WEBVTT

00:00:00.017 --> 00:00:04.957
Right. Are we on? We're not on. We're right. I guess he's not part of the interview.

00:00:05.097 --> 00:00:08.177
This is the Convergent Science Network podcast. Okay.

00:00:09.717 --> 00:00:13.237
Leading researchers in the domain of neuroscience, brain theory,

00:00:13.397 --> 00:00:18.537
and technology are interviewed by Paul Verschure and Tony Prescott. Seats, are we running?

00:00:20.317 --> 00:00:23.297
It's Paul Verschure with the Convergent Science Network podcast,

00:00:23.477 --> 00:00:27.517
together with my colleague Tony Prescott, and we're here in the room with David

00:00:27.517 --> 00:00:33.877
Reddish. And David, you presented us your research on what you call the cognitive rat.

00:00:34.517 --> 00:00:37.457
So what were you having in mind there with the cognitive rat?

00:00:37.957 --> 00:00:43.937
The idea is that if we can actually define cognitive functions carefully,

00:00:44.157 --> 00:00:51.357
then we can actually go in and we can look at whether animals also perform those cognitive functions.

00:00:52.697 --> 00:00:53.817
Particularly neurophysiologically.

00:00:53.977 --> 00:00:57.097
So if we look at the information processing that you need within cognition,

00:00:57.097 --> 00:01:00.117
cognition then you can actually find those and

00:01:00.117 --> 00:01:03.177
we found i think several examples of that do you

00:01:03.177 --> 00:01:07.137
mean that rats have knowledge in some tangible describable

00:01:07.137 --> 00:01:12.297
way yes rats definitely have knowledge about tasks that they have to do about

00:01:12.297 --> 00:01:17.697
the lives that they have can we can we define cognition here and a bit more

00:01:17.697 --> 00:01:24.977
specifically so what how would you define cognition um i the definition that i've been using,

00:01:24.997 --> 00:01:27.877
and I like to run with this and kind of see how far it will go,

00:01:27.977 --> 00:01:30.997
is the idea that there is covert knowledge,

00:01:31.197 --> 00:01:38.077
knowledge about things that are not immediately available to the animal, to you.

00:01:38.197 --> 00:01:43.057
So if you, for example, think about something, you will mentally time travel,

00:01:43.197 --> 00:01:45.617
you will imagine yourself into a future.

00:01:45.757 --> 00:01:50.157
So one of the things we've been able to show, for example, is that rats can

00:01:50.157 --> 00:01:53.717
mentally time travel. They can imagine other places and other times.

00:01:54.517 --> 00:01:59.617
For example. But now you emphasized in the introduction to your talk quite a

00:01:59.617 --> 00:02:04.397
bit on the one that it's historical antecedents, like in Tolman and then Hull.

00:02:04.937 --> 00:02:08.257
And also you spent quite some time on really explaining, let's say,

00:02:08.297 --> 00:02:14.197
the decoding methods that you have developed to actually make sense of the neural response.

00:02:14.337 --> 00:02:19.197
So what is this historical context in which you're doing this?

00:02:20.213 --> 00:02:23.333
Well, historically, of course, there's been a big question of whether or not

00:02:23.333 --> 00:02:25.633
animals can think, right?

00:02:25.733 --> 00:02:29.733
And whether they are more than just stimulus response creatures.

00:02:30.233 --> 00:02:35.153
And that, of course, was a big fight throughout much of psychology for many, many years.

00:02:35.813 --> 00:02:41.753
And um at this point i think of course lots of people not just us have kind

00:02:41.753 --> 00:02:46.393
of come to the conclusion that there are in fact cognitive information within

00:02:46.393 --> 00:02:51.793
that and we've come to that in large part because there are these information

00:02:51.793 --> 00:02:53.893
consequences that we can see.

00:02:55.093 --> 00:02:57.893
Historically you know i mean that's where i my understanding of the history

00:02:57.893 --> 00:03:05.593
comes okay but then And so the decoding methods you use to actually interpret

00:03:05.593 --> 00:03:07.773
these cellular responses that you find,

00:03:07.893 --> 00:03:10.513
and particularly you have looked at the hippocampus, but you also talked about

00:03:10.513 --> 00:03:12.253
other structures in the brain of the rat.

00:03:13.433 --> 00:03:17.633
Why do you put so much emphasis on the methods you use to decode these responses?

00:03:18.193 --> 00:03:23.253
Well, I think the key is to really understand. So to me, the key to the decoding

00:03:23.253 --> 00:03:26.093
story is that we are in some sense reading the mind.

00:03:26.313 --> 00:03:29.213
And I think that's the key here is that,

00:03:29.993 --> 00:03:33.933
If we really believe, as I think the evidence is now very, very solid,

00:03:34.033 --> 00:03:39.033
that the mind is the brain, and that in fact the brain is a physical instantiation

00:03:39.033 --> 00:03:42.513
which generates this psychological construct we call the mind,

00:03:42.713 --> 00:03:47.813
then we should be able to access the mind by looking at the physical properties of brain.

00:03:47.993 --> 00:03:53.093
And particularly in this case, that neurons communicate by sending these action

00:03:53.093 --> 00:03:56.753
potentials, these spikes, and we can listen to those spikes,

00:03:56.853 --> 00:04:00.833
and then we could ask, can we in fact find what those spikes represent.

00:04:01.213 --> 00:04:05.073
So the idea of the decoding is that we really have to, at some level,

00:04:05.713 --> 00:04:11.313
once we understand that information, we understand how it's encoded, then we can decode it.

00:04:11.593 --> 00:04:14.853
And at some level, the argument is we have to believe that decoding,

00:04:14.973 --> 00:04:18.813
even when it tells us that the rat is actually thinking about some other place.

00:04:19.113 --> 00:04:24.693
So you put quite a lot of emphasis on representations or knowledge in the rat

00:04:24.693 --> 00:04:28.293
mind of things which are not present at the current moment.

00:04:28.293 --> 00:04:34.653
Clearly because you want to distinguish between thinking about things and experiencing

00:04:34.653 --> 00:04:37.033
them in a direct way. Right.

00:04:38.993 --> 00:04:39.693
Is that...

00:04:40.729 --> 00:04:44.809
Is that where you were going out in terms of cognition? Do you link it to memory

00:04:44.809 --> 00:04:49.989
or do you go more towards reasoning, being able to reason and think about these

00:04:49.989 --> 00:04:52.109
things that are in your memory? Well, I think it's both.

00:04:52.189 --> 00:04:54.529
So one thing, I don't want to dismiss perception.

00:04:54.749 --> 00:04:57.929
Perception is an extremely complex question, and there's a lot of very,

00:04:58.029 --> 00:05:03.889
very interesting questions about how perception works, which ends up being much more complicated.

00:05:04.049 --> 00:05:08.889
Like, as I like to say, we don't perceive colors, we perceive objects, right?

00:05:09.069 --> 00:05:12.329
So how do you construct those objects? It's a whole interesting question there.

00:05:12.329 --> 00:05:17.709
But most of the cognitive question has been, I mean, everybody understands that

00:05:17.709 --> 00:05:20.549
animals can perceive objects. We know because they can manipulate them.

00:05:20.629 --> 00:05:28.189
The question is, does the animal actually use information beyond what's immediately available to it?

00:05:28.509 --> 00:05:32.569
And yes, the evidence is very solid now that it does. But to me,

00:05:32.609 --> 00:05:33.789
that's the cognitive question.

00:05:34.109 --> 00:05:37.969
So it's both a memory and a decision. In some sense, memory is decision,

00:05:38.149 --> 00:05:41.769
right? The only reason we remember things is to make better decisions in the future.

00:05:42.329 --> 00:05:44.089
I mean, otherwise, what's the evolutionary point of memory?

00:05:45.209 --> 00:05:48.309
Well, that's, of course, only when you want to use the notion of decision in

00:05:48.309 --> 00:05:51.789
a very broad sense. Yes. So I want to be very careful about the term decision.

00:05:52.109 --> 00:05:55.469
I'm going to define the decision as any time the animal takes an action.

00:05:55.549 --> 00:05:56.849
Then that's going to be a decision.

00:05:57.609 --> 00:06:02.549
So also this would be, let's say, a reflex-driven action? A reflex is a decision.

00:06:02.729 --> 00:06:06.549
And I emphasize that because the question then is not.

00:06:08.689 --> 00:06:13.289
How do you make a deliberative choice? That becomes a special kind of decision.

00:06:13.689 --> 00:06:19.789
But it becomes a what is the information processing happening when an animal does a reflex.

00:06:20.089 --> 00:06:22.849
A reflex actually is a decision because if you're measuring,

00:06:22.949 --> 00:06:27.569
for example, how hot your hand is relative to a stove, you put your hand on

00:06:27.569 --> 00:06:29.169
a cold stove, you don't have a reflex.

00:06:29.529 --> 00:06:33.289
Put your hand on a very hot stove, you have a reflex of pulling your hand away.

00:06:33.969 --> 00:06:38.589
At some point, there's a threshold. This reflex system has made a decision.

00:06:38.589 --> 00:06:43.629
Now, the question becomes, what is the mechanism by which the reflex has learned

00:06:43.629 --> 00:06:45.969
that decision? It's learned it over evolutionary time.

00:06:46.169 --> 00:06:49.789
And what is the mechanism by which the reflex has made that decision?

00:06:49.989 --> 00:06:54.729
And it's much simpler than deliberating between which job you're going to take, right?

00:06:54.849 --> 00:06:58.629
But they're both fundamentally, in the end, taking an action.

00:06:58.749 --> 00:07:02.509
But then how many levels of decisions do you distinguish in the rat?

00:07:03.568 --> 00:07:08.328
Well, I think that the way to think about it is not so much in terms of levels as processes.

00:07:08.588 --> 00:07:16.188
And as I see it, there are four information processing sequences that are very

00:07:16.188 --> 00:07:18.448
identifiably different in the mammalian brain.

00:07:18.748 --> 00:07:23.828
And those four systems tend to track. There's reflexes, there's deliberation,

00:07:23.908 --> 00:07:26.108
which is actually an imagination of the future.

00:07:26.608 --> 00:07:31.248
There's procedural learning, which is basically learning to do a procedure.

00:07:31.248 --> 00:07:38.028
I like to use the example of hitting a baseball or catching a football or something like that.

00:07:38.108 --> 00:07:41.748
I guess in Europe, it has to be not catching it, but hitting it with your foot.

00:07:41.848 --> 00:07:42.908
That's right, you got it.

00:07:43.848 --> 00:07:49.488
And then there's this fourth system, which ends up being, I call Pavlovian.

00:07:49.508 --> 00:07:54.808
I'm not sure that's the right word for it, but it's a species-specific behavior you learn to release.

00:07:55.028 --> 00:07:58.068
And I call it Pavlovian because this is what Pavlov's dogs were doing.

00:07:58.068 --> 00:08:01.308
Right they there was you salivate in

00:08:01.308 --> 00:08:04.308
response to food and they learn that when the bell comes there's food

00:08:04.308 --> 00:08:07.128
coming so they salivate so that's turns out those

00:08:07.128 --> 00:08:13.208
four systems end up being all of the fundamental information processing to take

00:08:13.208 --> 00:08:17.148
an action okay so you're saying then they're there and each of these systems

00:08:17.148 --> 00:08:21.028
will have then their own quality of decision making yes but now for their own

00:08:21.028 --> 00:08:26.608
neural structure sorry and their own neural structures you overlapping or uniquely Uniquely defined.

00:08:28.188 --> 00:08:31.448
Mostly separate. Okay. Not 100% separate, definitely.

00:08:32.268 --> 00:08:37.088
But they each have primary systems that are clearly quite different from each other.

00:08:37.268 --> 00:08:40.568
Right. But then to push you into a decision-making definition,

00:08:40.948 --> 00:08:44.388
also because you like to be clear about the definitions, right?

00:08:45.428 --> 00:08:50.788
If I take something like a scratch reflex, right? Or if I take like an eye blink reflex. Mm-hmm.

00:08:52.208 --> 00:08:58.108
A decision would imply that there is a sense of having options and that you select among options.

00:08:58.488 --> 00:09:02.848
But now, if you talk about the scratch reflex on the frog, which has been shown

00:09:02.848 --> 00:09:07.368
to be implemented in the spinal cord, once you generate that stimulus on the

00:09:07.368 --> 00:09:11.108
skin of the frog, there's just nothing else it can do but just scratch on that spot.

00:09:11.528 --> 00:09:13.688
So, there are no options involved.

00:09:14.148 --> 00:09:18.708
Well, but in fact, there are options involved because the amount of stimulus

00:09:18.708 --> 00:09:23.408
changes. So if you have very little stimulus, you basically don't actually touch

00:09:23.408 --> 00:09:25.028
the frog. The frog won't scratch.

00:09:25.588 --> 00:09:30.468
You poke it very sharp, the frog will scratch. There's some level in there where it will sweat.

00:09:30.708 --> 00:09:34.388
Now, I agree, reflexes, there's not a lot of these variability.

00:09:34.848 --> 00:09:37.728
It's not like deliberation where you're making from many, many choices.

00:09:37.728 --> 00:09:45.728
One of the examples I like to talk about is the famous scene in Lawrence of Arabia where T.E.

00:09:45.728 --> 00:09:49.728
Lawrence is showing off how cool he is and how tough he is by holding a match

00:09:49.728 --> 00:09:53.148
and letting it burn to his fingers and not executing the reflex.

00:09:53.768 --> 00:09:57.368
And so I like to give this example because what this actually is,

00:09:57.408 --> 00:10:01.708
in my view, is a conflict between two decision-making systems.

00:10:02.288 --> 00:10:06.308
A reflex system that wants to shake the match out before it burns to his fingers,

00:10:06.708 --> 00:10:11.548
and a deliberative system that wants to stop and say, no, don't do that because

00:10:11.548 --> 00:10:12.908
I want to show how cool I am.

00:10:13.528 --> 00:10:17.468
So you have this, I mean, a lot of people talk about it as kind of a top-down

00:10:17.468 --> 00:10:20.488
control, and I think it makes more sense to think of it as conflict.

00:10:21.628 --> 00:10:24.788
But still a conflict that must be resolved one way or the other.

00:10:24.908 --> 00:10:28.648
That's right. In fact, one of the very interesting open questions is how are

00:10:28.648 --> 00:10:29.488
those conflicts resolved?

00:10:29.888 --> 00:10:33.868
Right, exactly. And that's something that is not actually really well known right now.

00:10:34.401 --> 00:10:39.661
In the history of psychology, this idea that the rat is capable of cognition,

00:10:39.821 --> 00:10:41.721
of course, goes back a long way.

00:10:41.881 --> 00:10:45.661
And the work you referenced, you talked about Tolman, but there's better known

00:10:45.661 --> 00:10:50.921
experiments on, for instance, the cognitive map, the ability to take shortcuts, and so on.

00:10:51.121 --> 00:10:56.381
But you're wanting to go beyond that ability to say that the rat has more cognitive

00:10:56.381 --> 00:11:00.941
powers that are perhaps closer to things which you think of or have thought of as uniquely human.

00:11:00.941 --> 00:11:04.341
Actually, one of the really interesting things about Tolman's original cognitive

00:11:04.341 --> 00:11:06.961
map is that it's much more cognitive than map.

00:11:07.721 --> 00:11:11.321
It actually got translated and really thought of in terms of space,

00:11:11.501 --> 00:11:13.701
because of course, they were running rats on mazes.

00:11:13.861 --> 00:11:17.741
And of course, once they had the discovery of place cells in the 1970s,

00:11:17.741 --> 00:11:22.341
that kind of suggested, and John O'Keefe and Lynn Nadel's suggestion that the

00:11:22.341 --> 00:11:26.501
hippocampus was the seat of this cognitive map, really started to talk about

00:11:26.501 --> 00:11:27.541
it as a spatial paradigm.

00:11:27.541 --> 00:11:32.361
But the original Tolman concept was much more a structure of the world,

00:11:32.481 --> 00:11:38.501
and that the cognitive map was an understanding of the structure of the world

00:11:38.501 --> 00:11:41.901
with which you could essentially imagine yourself.

00:11:42.281 --> 00:11:46.361
And so it actually ended up being, as I like to say, much more cognitive than map.

00:11:46.641 --> 00:11:50.821
But what has tended to happen in the comparative literature as I know it,

00:11:50.861 --> 00:11:53.641
as people have said, okay, rats are great at spatial cognition,

00:11:53.961 --> 00:11:58.101
but that's where it stops. Would you like to go beyond that? Definitely.

00:11:58.601 --> 00:12:02.021
One of the things is that I think that the reason that rats are great at spatial

00:12:02.021 --> 00:12:07.621
cognition is it's much easier to construct a spatial task that rats understand.

00:12:08.201 --> 00:12:14.701
And I think it's very hard to construct non-spatial tasks that rats actually understand.

00:12:16.521 --> 00:12:21.301
Primates really like pushing levers for buttons, right? That's kind of the ultimate

00:12:21.301 --> 00:12:24.621
primate machine, right, is a soda machine.

00:12:24.901 --> 00:12:28.261
If you do something, you put some money and food comes out.

00:12:29.901 --> 00:12:35.501
Rats don't track that as well. But if you take these very cognitive things and

00:12:35.501 --> 00:12:39.341
translate them into a spatial place, not so much because of the space,

00:12:39.441 --> 00:12:44.761
but even just moving them to different locations, the rats can do very cognitive events.

00:12:45.081 --> 00:12:51.581
So we've, for example, done a task in which animal rats have to balance delay

00:12:51.581 --> 00:12:54.921
against food reward. How much food are you going to get?

00:12:55.381 --> 00:13:00.061
And what we did is we actually made the animals run to two different locations.

00:13:00.541 --> 00:13:05.221
And then we changed the delays in this complex contingency based on their choices.

00:13:05.641 --> 00:13:09.321
And they can do that. They understand that contingency and their behavior proves

00:13:09.321 --> 00:13:11.061
that they understand that contingency.

00:13:11.741 --> 00:13:15.101
And it was very hard to do until we moved them to a spatial location.

00:13:15.321 --> 00:13:21.421
It's not that they're doing the cognition in space, it's that the space just

00:13:21.421 --> 00:13:24.321
becomes easier to train them to do that task.

00:13:25.453 --> 00:13:28.773
But you might argue that spatial cognition is where the system evolved,

00:13:28.933 --> 00:13:33.733
and then it became more flexible, perhaps, in the evolutionary path leading

00:13:33.733 --> 00:13:38.013
to humans, so that we can now use these skills, but in a much more domain-independent way.

00:13:39.233 --> 00:13:44.513
Possibly. I'm not sure. The data as I see it doesn't seem to suggest that.

00:13:45.893 --> 00:13:49.733
But I'm trying to think if I can think of any good data offhand that's going

00:13:49.733 --> 00:13:51.273
to do that. Well, we know, for

00:13:51.273 --> 00:13:55.853
example, that rats see these same kind of cognitive effects across time.

00:13:56.333 --> 00:14:02.113
So there are these cells in the hippocampus, which you both know about,

00:14:02.273 --> 00:14:06.373
which are called place cells, which represent the location of an animal and

00:14:06.373 --> 00:14:12.253
can be used to both imagine positions in an environment and to navigate within an environment.

00:14:12.253 --> 00:14:17.333
But these same cells actually break up delays across time.

00:14:17.673 --> 00:14:24.333
And basically, it's almost as if the animal has a spatial map across the delay,

00:14:24.613 --> 00:14:29.033
which is a very cognitive event and is not spatial. It's actually temporal.

00:14:29.273 --> 00:14:38.533
So there we see a very similar use of the neural system, but in a temporal aspect

00:14:38.533 --> 00:14:40.093
instead of in a spatial aspect.

00:14:40.093 --> 00:14:45.213
But you could argue that you'd get that for free because movement in space implies

00:14:45.213 --> 00:14:48.853
that you have a temporal component, right? I agree. I agree.

00:14:49.033 --> 00:14:52.893
And I'm not – one of the dangers, of course, is how much if you say,

00:14:52.953 --> 00:14:55.553
well, it's for free, does that make it less cognitive?

00:14:56.333 --> 00:15:00.013
Ah, sure. Right. But I think that's an important point, right?

00:15:00.073 --> 00:15:05.613
That a lot of these things, these processes are evolved to work in such a way

00:15:05.613 --> 00:15:10.033
that they can solve these cognitive problems by building on other components.

00:15:10.073 --> 00:15:14.053
Right. But then it seems you would agree with Tony's contention that it might be co-opted.

00:15:14.795 --> 00:15:18.635
For other kinds of functions and cognitive operations. Well,

00:15:18.755 --> 00:15:21.015
so the big example, for example,

00:15:21.155 --> 00:15:24.835
that we've been looking at has been this thing called mental time travel,

00:15:24.935 --> 00:15:29.895
which is the ability to look to the future and to imagine yourself.

00:15:29.975 --> 00:15:32.895
We know that when humans do this, they actually construct a complete,

00:15:33.175 --> 00:15:36.235
essentially an episodic event of that future.

00:15:36.375 --> 00:15:39.555
What am I going to be wearing? Where am I going to be?

00:15:39.875 --> 00:15:42.835
Who am I going to be with? What's the world going to look like?

00:15:42.835 --> 00:15:49.275
And we know that the hippocampus is very involved in that future construction in humans.

00:15:49.935 --> 00:15:53.795
We know that the rats can construct spatial components in that future.

00:15:54.155 --> 00:15:59.935
What we don't know, and I emphasize the word don't know as compared to no, it's not true, right?

00:15:59.995 --> 00:16:03.695
We don't know whether rats, when they imagine that future, are simply imagining

00:16:03.695 --> 00:16:09.315
the location or imagining the entire episode of all of the flavors and other components.

00:16:10.175 --> 00:16:16.115
I suspect, given some of the data we've seen, that they are in fact constructing that complete future.

00:16:16.655 --> 00:16:21.295
But are they doing it in a fairly context-bound way in that of the options I

00:16:21.295 --> 00:16:25.635
have now, I could go that way and I can imagine mental time travel if I went

00:16:25.635 --> 00:16:29.335
this way, but I can't imagine what I would do tomorrow or next week.

00:16:29.675 --> 00:16:34.495
Right. So the short answer is, I'm not sure how we'd measure it, right?

00:16:34.615 --> 00:16:39.335
And that's, to me, the real problem is that until I know how to measure it,

00:16:39.415 --> 00:16:41.655
I don't want to say it doesn't exist.

00:16:43.675 --> 00:16:47.435
That's been, you know, for me over the last few years, this idea of,

00:16:47.495 --> 00:16:51.555
well, I didn't think rats could even look at the future until we figured out

00:16:51.555 --> 00:16:55.515
how to measure it, which is why I brought up this whole decoding stuff at the beginning,

00:16:55.615 --> 00:17:00.375
because it's that mathematical technology of decoding that allows us to measure

00:17:00.375 --> 00:17:04.115
these non-local representations of this future space.

00:17:04.555 --> 00:17:12.295
So the question is, how do we measure that episodic future event beyond space? And.

00:17:13.895 --> 00:17:18.855
I mean, yes, I'm pretty sure that rats are not writing Harry Potter, right?

00:17:18.955 --> 00:17:24.175
They're not, you know, actually constructing big fantasy novels that they can,

00:17:24.235 --> 00:17:27.355
you know, about all possibilities. That's pretty clear.

00:17:28.115 --> 00:17:31.735
But another, and so it's something you go back to the old question posed by

00:17:31.735 --> 00:17:35.295
Tolman about these cognitive maps, whether they're local or global, right?

00:17:35.335 --> 00:17:40.975
He was talking about whether they have sort of very delineated strips of information

00:17:40.975 --> 00:17:43.475
in there or whether it provides more global information.

00:17:43.895 --> 00:17:47.335
You're saying right now, given the data we have, we actually don't know whether

00:17:47.335 --> 00:17:51.435
these episodic memories that the hippocampus forms in the red are local or more

00:17:51.435 --> 00:17:53.215
integrated and contextual.

00:17:54.155 --> 00:17:56.235
But is that really fair to say?

00:17:57.595 --> 00:18:01.115
I would not have separated the local-global separation that way.

00:18:01.215 --> 00:18:06.595
I don't think that's what Tolman's local-global distinction was, as I understood it.

00:18:06.635 --> 00:18:10.715
I understood it being more, can you construct novel sequences,

00:18:11.055 --> 00:18:13.555
novel connections within your map?

00:18:13.555 --> 00:18:17.255
If you have a very thin strip of information and you're looking,

00:18:17.515 --> 00:18:22.795
I know what this street looks like, and I know what stores are on this street,

00:18:22.995 --> 00:18:25.855
and I have another street a block away that I know the stores,

00:18:26.055 --> 00:18:27.895
but I don't know how to cross those.

00:18:28.235 --> 00:18:31.555
Whereas if you have the global connection, you can do all those crosses,

00:18:31.635 --> 00:18:32.975
which is the shortcut story.

00:18:33.275 --> 00:18:37.215
Right. Right? One of Tolman's points was if you have a cognitive map, you can do shortcuts.

00:18:37.515 --> 00:18:43.515
That's right. it's very, very clear that the rats have a broad map on which

00:18:43.515 --> 00:18:50.735
they can do shortcuts and that they can actually connect up information in novel ways. That's solid.

00:18:51.895 --> 00:18:58.675
Whether the information is more than context within a specific location,

00:18:58.935 --> 00:19:04.615
that is, could they imagine a completely new non-spatial context connection,

00:19:04.935 --> 00:19:12.275
such as, you know, well, I can imagine elves and dwarves and Tolkien and all that stuff, right?

00:19:13.949 --> 00:19:18.449
I don't know whether rats can do that. But again, I don't know how I'd measure it.

00:19:18.529 --> 00:19:22.029
And if I don't know how to measure it, I don't know how to disprove it. Okay.

00:19:22.229 --> 00:19:26.189
But look, even though we might not be that clear about the boundaries of these

00:19:26.189 --> 00:19:31.329
episodic memories rats can form, what you have shown us a lot of data on today

00:19:31.329 --> 00:19:32.589
is this time travel component.

00:19:32.989 --> 00:19:37.869
Yes. Right? So what are the key observations there that you think are providing

00:19:37.869 --> 00:19:40.449
us insight on the ability of time travel in rats?

00:19:40.449 --> 00:19:44.929
Well, the key is Mental time travel, I should say.

00:19:44.929 --> 00:19:49.469
Right. The key is that the representations are self-consistent representations.

00:19:49.469 --> 00:19:55.049
That is, the neural signature is not noise. It's actually a specific neural

00:19:55.049 --> 00:19:59.469
signature of that other location. So the neurons at the current lo- that represent the current.

00:20:10.449 --> 00:20:13.429
Good representation of that other place right so that's one

00:20:13.429 --> 00:20:16.209
piece but practically that means you first

00:20:16.209 --> 00:20:19.389
run the rat you identify place cell responses

00:20:19.389 --> 00:20:22.369
you know okay here i have a set of cells that respond to location a

00:20:22.369 --> 00:20:26.369
i have another set of cells responding location b and what you now observe is

00:20:26.369 --> 00:20:30.809
that while in the future trial the animal is in location a the cells that correspond

00:20:30.809 --> 00:20:34.749
to location b start to fire correct right this is the signature that's the signature

00:20:34.749 --> 00:20:40.009
right so there's a few processes that connect that up to cognition and i think that it's important

00:20:40.289 --> 00:20:44.569
that when we talk about it, this is really coming primarily in our data in terms

00:20:44.569 --> 00:20:45.869
of this deliberation story.

00:20:46.389 --> 00:20:50.169
And so what you really need to do in order to get there is ask,

00:20:50.249 --> 00:20:55.129
what are the features that you should see in a deliberation event?

00:20:55.689 --> 00:20:59.849
And those features are that, for example, that the signal should be ahead of

00:20:59.849 --> 00:21:03.929
the animal, not behind the animal, because he's presumably deliberating about

00:21:03.929 --> 00:21:06.869
what he's going to do, for example.

00:21:07.049 --> 00:21:09.829
Well, that is an interesting assumption, right because.

00:21:10.952 --> 00:21:16.112
There you're just saying, look, the spatial trajectory the animal follows will

00:21:16.112 --> 00:21:18.292
also correspond with its future.

00:21:18.492 --> 00:21:21.372
Correct. But that's more like a constraint. You impose the task.

00:21:21.452 --> 00:21:26.272
But there's nothing that prevents, for example, the mental time travel from

00:21:26.272 --> 00:21:27.332
going backwards, right?

00:21:27.692 --> 00:21:32.532
Theoretically, the animal could just as easily be representing past events,

00:21:32.732 --> 00:21:34.672
events behind him, right?

00:21:34.752 --> 00:21:36.832
But practically, we see them being in front.

00:21:37.032 --> 00:21:40.632
Okay. The second piece, which I think is really critical, is their cereal.

00:21:40.952 --> 00:21:46.652
That is, it is representation of one side or the other, not both together.

00:21:47.512 --> 00:21:50.432
And so it's not just that it's spreading activation into the future.

00:21:50.592 --> 00:21:53.472
It's actually spreading down a very specific path.

00:21:53.872 --> 00:21:58.272
And I think that's another key factor that suggests, actually suggests what

00:21:58.272 --> 00:22:04.432
Tony's asking about, that it's much more of an actual episodic event, right?

00:22:04.512 --> 00:22:08.412
It's this, what if I go right? What happens?

00:22:08.752 --> 00:22:11.372
And then what if I go left? what happens.

00:22:12.692 --> 00:22:17.852
So but then you do impose certain constraints with the task.

00:22:18.072 --> 00:22:23.052
That means it's an alley that an animal runs through or an elevated platform.

00:22:23.672 --> 00:22:26.312
So it's not that the animal can move freely in an open space.

00:22:26.652 --> 00:22:30.212
Correct. Right. So do you see that as a limitation to the decoding?

00:22:31.672 --> 00:22:36.932
I see it as a practical step to being able to see the data we've seen so far.

00:22:37.132 --> 00:22:39.372
And I would very much like to do in open space.

00:22:39.592 --> 00:22:45.332
There are physical problems in terms of running on a physical maze.

00:22:45.572 --> 00:22:51.392
Because if you're going to say, have the animal come back, if you go to left,

00:22:51.652 --> 00:22:54.312
you go the same distance left and right, you can come back.

00:22:54.492 --> 00:22:58.952
But if you have a kind of 45 degree angle, the animal can't,

00:22:58.952 --> 00:23:00.572
the path back will be too long.

00:23:00.952 --> 00:23:07.392
So there's physical realities that we have to put the animal in VR to break.

00:23:08.172 --> 00:23:12.412
We haven't done that yet, but it's certainly on the table.

00:23:12.592 --> 00:23:16.392
Okay, very good. Well, we built technology for that, so we can talk about it.

00:23:16.512 --> 00:23:23.952
But now you also mentioned that you might see these forward sweeps, but at decision points.

00:23:24.592 --> 00:23:29.112
However, at the feeder sites, which in some sense is a termination point of

00:23:29.112 --> 00:23:32.292
a behavioral sequence, you get a very different kind of dynamic.

00:23:32.792 --> 00:23:35.732
Right, you have these sort of these high frequency events.

00:23:36.312 --> 00:23:39.632
So are they significant for this idea of mental time travel?

00:23:39.932 --> 00:23:44.452
Well, they are also good representations in the sense that they are consistent

00:23:44.452 --> 00:23:47.372
representations of other components of the task.

00:23:47.632 --> 00:23:52.472
So for example, again, the animals at location A and now the cells at A do not

00:23:52.472 --> 00:23:54.152
fire and the cells at B do fire.

00:23:54.412 --> 00:23:56.932
So again, we have this consistent jump.

00:23:57.827 --> 00:24:02.387
What's interesting about those data is that they're actually sequences on the

00:24:02.387 --> 00:24:05.247
track, and the sequences are both,

00:24:05.487 --> 00:24:10.487
they'll go in directions the animal has never traveled, they'll go across paths

00:24:10.487 --> 00:24:13.007
the animal has never actually traveled sequentially.

00:24:13.387 --> 00:24:18.407
And so it might be, I know how to go from, you know, the house to the library

00:24:18.407 --> 00:24:23.127
to the stadium, or I go from the house to the library and I know how to go from

00:24:23.127 --> 00:24:26.587
my house to the stadium, but now I actually know how to go from the library to the stadium.

00:24:26.587 --> 00:24:28.787
I just go library to house to stadium, right?

00:24:28.867 --> 00:24:31.847
That's a new path that I've never actually run.

00:24:31.967 --> 00:24:37.907
And we know that rats can, we will see representations of that during these

00:24:37.907 --> 00:24:38.787
high-frequency events.

00:24:38.787 --> 00:24:43.627
So it tells us that, you know, coming back to this cognitive map point,

00:24:43.847 --> 00:24:49.387
it suggests that animals have that more global representation of the cognitive

00:24:49.387 --> 00:24:53.487
map where they are able to physically,

00:24:53.947 --> 00:24:59.167
or sorry, mentally, not physically, but mentally actually connect up things

00:24:59.167 --> 00:25:01.367
that have never been physically connected. Right.

00:25:01.707 --> 00:25:06.307
But now, so at these termination points, these high-frequency events,

00:25:06.587 --> 00:25:08.587
you might see forward and backward sequences.

00:25:08.827 --> 00:25:12.067
Yes. Okay, so now the situation is changing. Correct.

00:25:13.467 --> 00:25:17.187
But on top of that, if you start to see what you call shortcuts,

00:25:17.587 --> 00:25:21.387
actually you could argue that it's neither forward nor backward.

00:25:21.667 --> 00:25:25.727
That's right. It's more like an implication of the information you have sampled.

00:25:25.847 --> 00:25:29.487
That's right. Would you agree with that? Yes. So what aspect of cognition would that reflect?

00:25:30.367 --> 00:25:34.727
I think it's about imagination. I think it's actually daydreaming.

00:25:35.387 --> 00:25:38.867
And I think it's like, you know, when somebody's sitting in the talk and they're

00:25:38.867 --> 00:25:42.027
not paying attention and they start thinking about other things, right?

00:25:42.087 --> 00:25:44.847
I think that's very much what we're seeing, you know? You know,

00:25:44.847 --> 00:25:49.047
I mean, I can't say the animal's dreaming, right?

00:25:49.387 --> 00:25:53.967
But we know that when humans, for example, dream that, or when they imagine,

00:25:54.047 --> 00:25:57.027
when a human imagines things, like imagines a face, right?

00:25:57.067 --> 00:26:01.527
The same part of the brain that is active when they perceive a face becomes active.

00:26:02.047 --> 00:26:08.547
So should we be surprised that when a rat is imagining other places,

00:26:08.787 --> 00:26:12.787
the areas representing those other places become active, right?

00:26:12.787 --> 00:26:19.607
So I think you're seeing kind of imagination and thinking about how the world is structured.

00:26:20.367 --> 00:26:22.387
But is it also, would you call it insight?

00:26:23.187 --> 00:26:28.847
Yeah, I'd be happy to call it insight. But then, can you say something about the dynamics of this?

00:26:28.927 --> 00:26:32.747
Because in some sense, if I'm exploring my maze, here I am, I'm the rat exploring

00:26:32.747 --> 00:26:37.687
the maze, I'm driving my hippocampal cells that you are decoding.

00:26:38.769 --> 00:26:42.129
Sort of time lock to my action in space yes right

00:26:42.129 --> 00:26:45.829
but now in these sweeps i'm also

00:26:45.829 --> 00:26:51.129
time morphing if time warping yes because i'm not replaying that at the same

00:26:51.129 --> 00:26:55.769
real time as correct so so what's the relationship there between the dynamics

00:26:55.769 --> 00:27:01.009
and of this kind of well cognition happens faster than behavior okay and i mean

00:27:01.009 --> 00:27:04.589
i agree with you and so there's definitely we know that it's faster faster.

00:27:05.129 --> 00:27:07.509
It's about 40 times faster than behavior.

00:27:08.329 --> 00:27:13.329
Just that's the data. Why it's 40 times faster, I have no idea.

00:27:13.549 --> 00:27:17.749
What is the mechanism that makes it 40 times faster? Again, I have no idea.

00:27:17.929 --> 00:27:22.669
It's a very interesting open question. But if, for example, you can use this

00:27:22.669 --> 00:27:29.849
decoding mathematics to ask what speed of sequence is the best description of your data?

00:27:29.989 --> 00:27:33.089
Is it the speed of the animal is it seven times faster

00:27:33.089 --> 00:27:36.669
is it 15 is it 40 so we did this and

00:27:36.669 --> 00:27:41.389
at every moment we said which is the most active which is the best model right

00:27:41.389 --> 00:27:45.669
so most of the time it'd be about one to seven times because of this this thing

00:27:45.669 --> 00:27:50.049
called phase precession where the play cells do at these moments of sweeps it

00:27:50.049 --> 00:27:55.029
seems to be 15 times faster it looks like it's kind of a long a faster event,

00:27:55.689 --> 00:28:00.249
and then at these sharp waves at these replay events these events you're talking

00:28:00.249 --> 00:28:02.029
about at the feeders, they were 40 times faster.

00:28:02.149 --> 00:28:06.229
What was interesting to us is we could actually go in with the look at the speed

00:28:06.229 --> 00:28:10.209
of decoding and then ask, can we find the events?

00:28:10.349 --> 00:28:14.429
And we actually could construct and find the events from those speeds.

00:28:14.769 --> 00:28:16.829
So I think it does happen faster.

00:28:18.046 --> 00:28:21.466
Why it happens faster is a really interesting question. But is this constrained

00:28:21.466 --> 00:28:24.626
by the basic rhythmicity of the structure you look at?

00:28:24.846 --> 00:28:28.426
Like hippocampus, you know, you have a very definite theta cycle.

00:28:28.766 --> 00:28:34.026
You have a gamma range of oscillations within the theta.

00:28:34.366 --> 00:28:39.786
Isn't that already constraining, let's say, the rhythmicity of these replay events?

00:28:40.146 --> 00:28:46.306
I would assume so. Okay. I mean, again, now we're talking the fact that this

00:28:46.306 --> 00:28:47.586
is a physical brain, right?

00:28:47.586 --> 00:28:51.686
And the fact that it's a physical brain means that you have connections,

00:28:52.026 --> 00:28:56.026
your circuits are actually driving your system.

00:28:56.226 --> 00:29:01.006
And so the question is, what is the mechanism within those circuits that enables

00:29:01.006 --> 00:29:06.126
this rhythmicity and this speed and this learning and all of this?

00:29:06.786 --> 00:29:10.166
That's a fascinating question. There's lots and lots of labs working on that.

00:29:10.286 --> 00:29:15.346
I don't think the answer is known at this point. Okay.

00:29:15.526 --> 00:29:19.086
So when people have talked about these sequences in the past,

00:29:19.126 --> 00:29:23.666
and it's been known for some time, they've often done so in the context of memory consolidation.

00:29:23.806 --> 00:29:27.966
And they've argued that if these processes are interrupted, so you don't get

00:29:27.966 --> 00:29:29.646
enough sleep, then you don't consolidate memories.

00:29:30.006 --> 00:29:33.786
So is that the same thing? Or is that a different explanation?

00:29:33.786 --> 00:29:38.826
So it turns out there's been fascinating data over the last five years or so,

00:29:38.886 --> 00:29:43.306
which has suggested that these events, these called replay events,

00:29:43.566 --> 00:29:48.046
these sharp wave events, during waking states are different than during sleep states.

00:29:48.926 --> 00:29:52.266
Now, there's nothing we've been able to directly observe within them in terms

00:29:52.266 --> 00:29:55.986
of the frequencies or other parameters that we've been able to figure out that's different.

00:29:56.246 --> 00:29:59.386
But if you interrupt them during sleep, you interrupt consolidation.

00:29:59.806 --> 00:30:04.666
You interrupt the transfer of memory from the hippocampus, the semanticization

00:30:04.666 --> 00:30:07.186
of memory from hippocampus to other structures.

00:30:08.786 --> 00:30:13.906
However, Lauren Frank's lab, Zhadov is the first author of the paper,

00:30:14.046 --> 00:30:17.066
showed that if you disrupt these events during waking states,

00:30:17.246 --> 00:30:18.706
you actually disrupt decision-making.

00:30:19.694 --> 00:30:24.934
And you actually disrupt working memory within the task, not the consolidation afterwards.

00:30:25.734 --> 00:30:31.634
We actually looked at the content of these events during waking states and sleep states.

00:30:31.794 --> 00:30:34.734
And during the waking states, you see a lot of things that relate to insight.

00:30:35.314 --> 00:30:40.374
Shortcuts, forward, backward. It seems it's almost kind of covering the space

00:30:40.374 --> 00:30:42.914
as if it's just exploring the space mentally.

00:30:43.494 --> 00:30:49.214
But during sleep, they're almost always forward. They're almost always replaying the actual events.

00:30:49.694 --> 00:30:52.054
And of course, that's what you want. If you're consolidating memory,

00:30:52.454 --> 00:30:56.354
you don't really want to explore the space.

00:30:56.474 --> 00:30:58.674
You want to consolidate what actually happened.

00:30:59.234 --> 00:31:04.054
And so it suggests that what may well be happening, and I emphasize the word

00:31:04.054 --> 00:31:06.934
may because this is exactly where the field is right now,

00:31:07.114 --> 00:31:15.274
is that these events during waking states are more involved in some sort of insight, imagination,

00:31:15.814 --> 00:31:18.434
processing, trying to figure out what's going on.

00:31:18.434 --> 00:31:21.394
And during sleep, it's more of a consolidation event.

00:31:22.774 --> 00:31:26.054
Not dreaming. The rats aren't dreaming. I think the rat's dreaming.

00:31:26.774 --> 00:31:30.814
I don't have proof it is, but of course there is this new data from,

00:31:31.754 --> 00:31:35.494
I should know the authors, where they did actually an fMRI study.

00:31:35.714 --> 00:31:39.234
I don't know how they got the people to sleep in the fMRI machine, but they did.

00:31:39.474 --> 00:31:44.874
And they were able to show that after dreaming, these are humans,

00:31:44.974 --> 00:31:47.874
so they could wake them up and actually say, what were you dreaming about?

00:31:48.434 --> 00:31:53.394
That they were actually able to use similar decoding methods to show that if

00:31:53.394 --> 00:31:56.534
the dream contained people, the face areas were active.

00:31:56.694 --> 00:31:59.634
And if the dream contained landscapes, the landscape areas were active.

00:31:59.954 --> 00:32:06.374
And so just as lots of data from Nancy Kanwisher's lab and many other labs have

00:32:06.374 --> 00:32:10.854
shown that if you imagine a face, you're using the face part of cortex.

00:32:11.194 --> 00:32:15.294
This paper, again, I apologize, I should know the authors offhand,

00:32:15.454 --> 00:32:19.754
but it was in science about a year ago. This paper...

00:32:20.738 --> 00:32:24.418
Showed that during during dreams and we know their dreams because we asked the

00:32:24.418 --> 00:32:27.318
people or they asked the people and the people said They were dreaming these

00:32:27.318 --> 00:32:29.978
same structures were active now our rats dreaming.

00:32:30.158 --> 00:32:35.398
I don't know well, I guess you would look for correlated activity or certainly.

00:32:36.418 --> 00:32:39.158
Temporary related activity in the sensory areas.

00:32:39.298 --> 00:32:44.778
Yes, and those exist right that has been shown to exist during these Sleep events

00:32:44.778 --> 00:32:51.518
which during the sleep events you definitely get correlated activity in cortical areas.

00:32:52.898 --> 00:33:01.078
And you also get that actually after these events, the activities in the cortical

00:33:01.078 --> 00:33:05.638
areas actually tie together better, suggesting that there is some sort of transfer

00:33:05.638 --> 00:33:07.878
of information going into the cortex.

00:33:08.698 --> 00:33:13.878
So, but you would want to allow some things that humans can do that rats can't do.

00:33:14.058 --> 00:33:18.338
Well, rats don't talk to us. Okay, but beyond language, which most people will

00:33:18.338 --> 00:33:24.258
agree, but for instance, an ability maybe to frame what you're thinking about

00:33:24.258 --> 00:33:28.518
or imagining about in the future is maybe something that we're particularly good at.

00:33:28.578 --> 00:33:33.058
So I can imagine some arbitrary event, or I can ask you to imagine yourself

00:33:33.058 --> 00:33:36.198
in some arbitrary place, arbitrary time, and you can think that through.

00:33:36.698 --> 00:33:40.998
Is that something maybe that rat circuits would be less able to do,

00:33:41.038 --> 00:33:44.638
or you just want to leave that for future research? I would leave that for future

00:33:44.638 --> 00:33:48.658
research, but I think that we, I mean, I don't want to claim that rats are small humans.

00:33:48.958 --> 00:33:53.298
That's pretty clear, right? And beyond language, you know, when we imagine the

00:33:53.298 --> 00:33:55.338
future, we're imagining decades,

00:33:55.578 --> 00:33:59.918
we can imagine centuries ahead, we can imagine, you know, normal humans plan

00:33:59.918 --> 00:34:05.078
about 10 years ahead, typically, a typical human, you know, why is a human going to college?

00:34:05.238 --> 00:34:07.998
Well, they'll tell you because I want to go get a job in five years,

00:34:08.138 --> 00:34:13.878
right? So, typical humans are planning years ahead. Rats are not planning years ahead.

00:34:14.418 --> 00:34:20.798
So, I think it's more a question of scale that's changing.

00:34:21.098 --> 00:34:26.998
And I mean, we're not surprised to see that there are similarities in other organs, right?

00:34:27.098 --> 00:34:33.018
So, why should we not expect to see similarities in these cognitive structures,

00:34:33.238 --> 00:34:36.358
right? Well, it could also be just a difference of degree, right?

00:34:36.438 --> 00:34:37.298
Yes. It doesn't need to be the

00:34:37.298 --> 00:34:41.958
case that the fundamental processes are qualitatively different. Exactly.

00:34:42.298 --> 00:34:47.078
But then the question is, so we have an idea now on, let's say,

00:34:47.098 --> 00:34:52.198
the basic memory dynamics of hippocampus as you have assessed it experimentally.

00:34:52.738 --> 00:34:57.518
And we have these different kinds of replay events forward and backward.

00:34:58.558 --> 00:35:03.478
But to whom's benefit? So I'm here, I'm the rat, I'm running in a maze.

00:35:03.718 --> 00:35:09.198
I'm at the reward side. I have a forward or backward, uh, replay. Um, and,

00:35:10.980 --> 00:35:14.960
Who's going to process that? Am I throwing this out for my neighboring or for

00:35:14.960 --> 00:35:18.680
my other cells in the hippocampus? I'm assigning this to other brain areas?

00:35:19.620 --> 00:35:25.320
The short answer is we don't know. Okay. We do know that other structures are

00:35:25.320 --> 00:35:26.980
listening to those events.

00:35:26.980 --> 00:35:31.720
We know, for example, that the ventral striatum, the nucleus accumbens,

00:35:31.880 --> 00:35:37.500
which is a structure involved in evaluation and motivation,

00:35:38.020 --> 00:35:43.880
has reward information in it that those reward cells also replay.

00:35:43.880 --> 00:35:48.920
That's Karian Lansing's data from Cyril Pennartz's lab, where they saw that

00:35:48.920 --> 00:35:55.840
the sequence of hippocampal replays, I think during sleep states,

00:35:55.900 --> 00:35:56.660
if I'm remembering right,

00:35:56.800 --> 00:36:02.240
actually triggers immediately following the appropriate ventral striatal information,

00:36:02.620 --> 00:36:05.180
which actually suggests that it's not just...

00:36:05.180 --> 00:36:08.580
Because in fact, they had different flavors, and they saw the correct flavors

00:36:08.580 --> 00:36:12.880
being decoded, which again suggests actually it's more than just space,

00:36:12.960 --> 00:36:16.900
that it actually contains space and these other information as well,

00:36:17.000 --> 00:36:18.760
at least in ventral striatum and accumbens.

00:36:20.160 --> 00:36:27.580
It's known that cortical systems are, as I said, listening to these replay events.

00:36:27.860 --> 00:36:35.340
But it's quite possible that the replay events are actually also affecting hippocampus internally.

00:36:36.040 --> 00:36:41.440
That would not surprise me. But I don't think it's known at this point.

00:36:41.640 --> 00:36:46.280
So when we talk about the human literature on hippocampus, we usually use the word episodic memory.

00:36:46.540 --> 00:36:50.800
And essentially, you're saying rats have episodic memory and that space is one

00:36:50.800 --> 00:36:51.940
part of that. That's right.

00:36:52.180 --> 00:36:57.060
And one of the things that I think is very interesting is that I would say actually

00:36:57.060 --> 00:37:01.400
that it's mostly about episodic future than episodic past.

00:37:01.660 --> 00:37:06.540
There's some very beautiful kind of discussions coming out in the human literature

00:37:06.540 --> 00:37:10.640
about hippocampuses being necessary for imagining the future,

00:37:10.720 --> 00:37:13.200
which is an episodic, it's called episodic future thinking.

00:37:13.580 --> 00:37:16.580
And one of the things that's interesting is, of course, episodic memory,

00:37:16.860 --> 00:37:20.200
episodic past thinking, is notoriously fragile.

00:37:21.450 --> 00:37:25.590
It's very easy to manipulate that by framing questions in strange ways,

00:37:25.750 --> 00:37:30.290
or not even strange ways, just even normal questions, can change what you think you remember.

00:37:30.750 --> 00:37:34.810
And I think, and I think that a lot of the human literature now is coming to

00:37:34.810 --> 00:37:41.410
this, is that the reason that episodic memory is so fragile is it's not really a memory.

00:37:41.770 --> 00:37:44.390
It's actually an imagination of the past.

00:37:44.810 --> 00:37:48.850
And in the same way that you imagine the future, you're actually taking these

00:37:48.850 --> 00:37:52.010
pieces of memory and putting them back together.

00:37:52.490 --> 00:37:57.330
And because you're reconstructing, you're rebuilding that past using hippocampus.

00:37:57.330 --> 00:38:01.930
This is why you need hippocampus for episodic memory, is that you're actually

00:38:01.930 --> 00:38:03.670
doing this kind of episodic past thinking.

00:38:04.090 --> 00:38:08.330
Well, one of the theories of what the hippocampus might be doing is not so much

00:38:08.330 --> 00:38:13.570
about events or processes in time, but this idea that it's doing pattern completion.

00:38:14.490 --> 00:38:21.330
So you are given some information through your sensory systems and from what

00:38:21.330 --> 00:38:25.210
your hippocampus encodes about maybe where it is in space, it's able to fill that out.

00:38:25.470 --> 00:38:30.130
So that kind of pattern completion aspect of what the hippocampus is doing is

00:38:30.130 --> 00:38:34.430
another thing that it's contributing beyond being able to forecast future events.

00:38:34.750 --> 00:38:39.130
But that's the whole point of that future event is you need to pattern complete

00:38:39.130 --> 00:38:44.650
it because what you need to do is you need to take the pieces and put them together

00:38:44.650 --> 00:38:48.110
and complete the pattern sometimes in novel ways,

00:38:48.190 --> 00:38:50.830
sometimes in not novel ways, Basically, you have to, as I said,

00:38:50.870 --> 00:38:55.430
the reason you have the memory is to construct that future so that you have

00:38:55.430 --> 00:39:03.290
to take those pieces and you use pattern completion processes to rebuild that system.

00:39:03.650 --> 00:39:06.830
When people talk about autobiographical memory, though, they often talk about

00:39:06.830 --> 00:39:10.510
involuntary autobiographical memory, which is where they're referring to the

00:39:10.510 --> 00:39:15.170
fact that something about the current context reminds me of a past event.

00:39:15.170 --> 00:39:20.250
And one explanation of that is it's about reinterpreting the current context

00:39:20.250 --> 00:39:23.250
in relation to things that have happened in the past.

00:39:23.490 --> 00:39:27.030
So I think maybe you push it too far to say it's about imagining the future

00:39:27.030 --> 00:39:29.130
because it's also about understanding the present.

00:39:29.470 --> 00:39:34.250
Yes, absolutely. And in fact, I would very much argue that one of the reasons

00:39:34.250 --> 00:39:39.870
we have episodic past thinking and episodic memory is so we can reinterpret that past.

00:39:40.090 --> 00:39:43.030
And we can say, oh, that's what that really meant.

00:39:43.030 --> 00:39:49.810
Yeah, but so in some sense, you're sort of making a contrast to the more traditional

00:39:49.810 --> 00:39:55.230
view of memory, where it's like a storehouse of the facts that you encounter, right?

00:39:55.230 --> 00:39:57.410
Right, exactly. And then basically what you're saying is, well,

00:39:57.510 --> 00:40:02.490
maybe these so-called facts are more malleable to change than we thought they would be.

00:40:02.530 --> 00:40:05.350
Yes. Like memory is continuously constructive. That's right.

00:40:06.930 --> 00:40:10.470
There are still, if you want, ingredients of that past in there.

00:40:10.630 --> 00:40:12.510
Yes. That's not completely arbitrary.

00:40:12.870 --> 00:40:17.090
Absolutely. Okay. So I don't think, so if you say, look, it's focused on the

00:40:17.090 --> 00:40:23.050
future, is that not, let's say, a bit of an overstatement in that sense?

00:40:23.250 --> 00:40:27.030
I think it does both future and past. But I actually think we have to be careful

00:40:27.030 --> 00:40:32.290
because, you know, we're talking as if the future is completely open and malleable

00:40:32.290 --> 00:40:34.090
and plastic and all that.

00:40:34.090 --> 00:40:38.750
But the fact is, we use those same ingredients to predict the future.

00:40:39.290 --> 00:40:43.150
You know, when I came to give the talk, this is the first time I've been in

00:40:43.150 --> 00:40:46.010
Barcelona, it's the first time I've been in this auditorium,

00:40:46.030 --> 00:40:52.750
but I've given talks to other people with similar audiences and similar backgrounds, right?

00:40:52.750 --> 00:40:58.910
So I said to myself, okay, I can take my ingredients from all these other pieces

00:40:58.910 --> 00:41:01.450
and I can use this to construct that future.

00:41:01.750 --> 00:41:07.150
And again, I'm using facts to do it. I think that's why you construct with those facts.

00:41:08.010 --> 00:41:12.290
You're absolutely right. I mean, the goal of memory is not to imagine what it

00:41:12.290 --> 00:41:13.730
could have been. And you could do that.

00:41:13.810 --> 00:41:18.310
That's a counterfactual question and one of the things that people do.

00:41:18.410 --> 00:41:24.290
In fact, we now know it's one of the things rats do. But it is also true that

00:41:24.290 --> 00:41:27.690
sometimes your goal is to try to get as accurate a memory as possible.

00:41:28.110 --> 00:41:30.510
The key is that that's actually hard.

00:41:31.410 --> 00:41:34.550
Which is the whole, I mean, that goes back to Elizabeth Loftus. Right.

00:41:34.850 --> 00:41:37.670
No, but isn't the cool consequence or an interesting consequence of that,

00:41:37.730 --> 00:41:42.830
that maybe if you just take a hippocampal centric view on memory. Right.

00:41:43.830 --> 00:41:47.810
The hippocampus just gets information from other extra hippocampal areas.

00:41:48.150 --> 00:41:51.410
And this might come from sensors. It might come from other memory systems.

00:41:51.790 --> 00:41:54.150
And it's just sitting there chunking out episodes.

00:41:54.710 --> 00:41:59.130
So in some sense, you could argue, well, for that memory system,

00:41:59.230 --> 00:42:04.370
if you look at CA3, which is really a core memory system of the hippocampus,

00:42:05.070 --> 00:42:08.610
whatever information I dump in there, whether it comes from a past experience

00:42:08.610 --> 00:42:10.830
or a current event, I actually don't care.

00:42:10.930 --> 00:42:16.710
I just compress this together in an episode irrespective of the origins of that piece of information.

00:42:17.010 --> 00:42:21.450
And that would mean that as much as I'm predicting my past, I'm also predicting my future.

00:42:21.550 --> 00:42:27.230
If you want, always remodeling my past, given this mixing of current states

00:42:27.230 --> 00:42:28.810
with past states. Would you agree with that?

00:42:29.110 --> 00:42:33.570
I think that's a very plausible explanation, and it's probably the most likely one.

00:42:33.750 --> 00:42:38.470
But at this point, I think it's also very important to say, we don't actually

00:42:38.470 --> 00:42:41.970
have the connection between these mechanisms.

00:42:43.070 --> 00:42:46.450
And the underlying circuitry. That is, we don't know what it is.

00:42:46.510 --> 00:42:48.410
And this is the point you're asking about the oscillations.

00:42:48.570 --> 00:42:53.290
We don't actually know what it is about those fundamental circuitry that is

00:42:53.290 --> 00:42:55.010
enabling these processes.

00:42:55.370 --> 00:42:57.070
We know these processes exist.

00:42:57.530 --> 00:43:01.170
We know that they're there in hippocampus. We know hippocampus is doing them.

00:43:01.330 --> 00:43:03.750
We know a lot about the circuitry of hippocampus.

00:43:03.930 --> 00:43:08.290
But at this point, that connection is still actually an open question.

00:43:08.290 --> 00:43:14.550
But the interesting thing is that these hippocampal areas are equally predicting

00:43:14.550 --> 00:43:19.450
the past in the sense that if prediction errors relative to past events would

00:43:19.450 --> 00:43:22.310
exceed a certain acceptable threshold, you just change them.

00:43:24.773 --> 00:43:27.973
Would you agree with that? Say that again? Well, for instance,

00:43:28.093 --> 00:43:34.873
if you see sweeps in hippocampus that reflect memory, now you are in a position

00:43:34.873 --> 00:43:37.533
to actually measure the accuracy of that memory.

00:43:37.693 --> 00:43:42.153
You can pose the question, well, was the rat exactly in that position or was it sort of there?

00:43:42.333 --> 00:43:47.493
Is it like already, if you are massaging this memory to the current state?

00:43:48.453 --> 00:43:51.333
So in the data you have on that, do you see really,

00:43:51.453 --> 00:43:55.473
let's say, accurate factual replay of

00:43:55.473 --> 00:43:58.873
past events that you really know x y is identical yeah

00:43:58.873 --> 00:44:01.513
or do you already see a massaging of that

00:44:01.513 --> 00:44:04.593
we already see the massaging there you go absolutely no the

00:44:04.593 --> 00:44:13.533
the replay events are definitely not completely veridical the question is which

00:44:13.533 --> 00:44:19.753
we don't know is how much of that noise matters right right how much of that

00:44:19.753 --> 00:44:21.553
noise is actually reflecting,

00:44:22.413 --> 00:44:25.873
a or i should say not noise how much of that difference is

00:44:25.873 --> 00:44:28.813
actually reflecting a change and how

00:44:28.813 --> 00:44:31.893
much of that difference is actually just noise right exactly

00:44:31.893 --> 00:44:39.093
and we know that it's not i mean biology is messy right but how much of that

00:44:39.093 --> 00:44:43.693
is on which side you know is actually a very interesting open question okay

00:44:43.693 --> 00:44:48.553
so in your data you showed us a couple of examples one of those was was what

00:44:48.553 --> 00:44:50.653
you call Vicarious Trial and Error,

00:44:50.693 --> 00:44:54.593
which is an idea going back to Tomlin, but also before that,

00:44:54.673 --> 00:44:56.933
I think it was Mertzinger and Gentry, 1931.

00:44:57.213 --> 00:45:02.833
So a very nice example of how when the rat gets to the top of a junction in

00:45:02.833 --> 00:45:07.733
the maze, it might glance left and you can see a play out of what might happen

00:45:07.733 --> 00:45:09.173
if it was to go down that left tunnel.

00:45:10.297 --> 00:45:16.577
But in a way, that's evidence to me that, yes, he can plan ahead and he can

00:45:16.577 --> 00:45:20.857
use past experience in order to fill out what could happen in the immediate future.

00:45:21.337 --> 00:45:25.017
It's a little bit like the involuntary autobiographical memory ideas that you're

00:45:25.017 --> 00:45:27.737
reminded of stuff which is relevant to your current choice.

00:45:28.177 --> 00:45:32.657
But then you're also talking about situations where he was at a feeding point

00:45:32.657 --> 00:45:36.537
and then he would just replay being in other positions in the maze.

00:45:36.537 --> 00:45:40.977
So that perhaps is stronger evidence that the rat can imagine himself in other

00:45:40.977 --> 00:45:43.117
positions in the world. I mean, would you go that far?

00:45:43.357 --> 00:45:49.297
Well, I would actually say we don't know how involuntary that forward sequence is. Right.

00:45:49.497 --> 00:45:55.677
Right. So, in fact, to be honest, my gut is it's much more voluntary than involuntary.

00:45:57.037 --> 00:46:03.777
Why are you saying that? To be honest, mostly because when humans are doing

00:46:03.777 --> 00:46:11.577
this kind of deliberation moment, it's extremely voluntary, and it's very much very cognitive.

00:46:11.577 --> 00:46:14.717
In humans, reaches conscious, tends to reach consciousness.

00:46:14.937 --> 00:46:19.357
It tends to be, you compare it, for example, to a more automated behavior,

00:46:19.577 --> 00:46:25.357
where in a human, you've switched from this kind of very flexible,

00:46:25.497 --> 00:46:29.717
slow, deliberative mode to a very habitual kind of automatic mode.

00:46:29.857 --> 00:46:33.917
In fact, it tends to be less reaching immediate consciousness.

00:46:34.277 --> 00:46:38.097
And it often is kind of a, well, you ask, for example, a sports player,

00:46:38.277 --> 00:46:40.317
why did you do that? They say it felt right.

00:46:41.940 --> 00:46:45.160
And in fact, you ask a sports player, we'll talk about being in the zone when

00:46:45.160 --> 00:46:46.260
things just kind of click.

00:46:46.700 --> 00:46:50.700
And I think what they're doing is recognizing that their habitual system,

00:46:50.800 --> 00:46:53.120
their procedural system is kind of working correctly.

00:46:53.120 --> 00:47:04.060
So I don't know how much of the rat forward sequence is voluntary versus involuntary.

00:47:04.300 --> 00:47:09.780
And I don't know how much of the replay happening is voluntary versus involuntary.

00:47:10.560 --> 00:47:14.520
It's possible, and there are a lot of computational models that have suggested,

00:47:14.840 --> 00:47:20.660
that the events happening at the The feeder sites is just what happens if you

00:47:20.660 --> 00:47:24.660
have the right connection structure and you put noise into the system and it

00:47:24.660 --> 00:47:25.840
kind of percolates over.

00:47:26.820 --> 00:47:32.720
Okay, so the events happening at the corners in the maze are more cognitive.

00:47:33.160 --> 00:47:39.340
And what you're saying is that perhaps the evidence at the feeder sites is not perhaps active.

00:47:39.840 --> 00:47:44.420
I want to think about if I was in that part of the maze, it's just the noise

00:47:44.420 --> 00:47:46.780
in the Bacampus could give rise to these sorts of patterns.

00:47:46.780 --> 00:47:51.140
So that's the current theory, but let me give you a very interesting data point,

00:47:51.240 --> 00:47:56.200
which I still don't know what to make of it, which is that we had an experiment.

00:47:58.040 --> 00:48:02.100
And on this experiment, the animal would sometimes go only to one side of the

00:48:02.100 --> 00:48:06.400
maze for many, many trials, and then would go to the other side of the maze,

00:48:06.520 --> 00:48:08.460
and vice versa. And it would do.

00:48:08.520 --> 00:48:13.420
But the point is, we could ask, how recently had it been on the current side

00:48:13.420 --> 00:48:15.300
of the maze or on the opposite side of the maze?

00:48:15.860 --> 00:48:21.460
And what we found is that it was more likely to imagine in these kind of,

00:48:21.480 --> 00:48:25.500
during these events happening at the feeder sites, this thing we're talking

00:48:25.500 --> 00:48:27.420
imagination insight kind of things,

00:48:27.600 --> 00:48:32.600
was much more likely to be on the opposite side of the maze when the animal

00:48:32.600 --> 00:48:34.720
had not been there recently.

00:48:36.540 --> 00:48:41.680
And that's strange that doesn't fit the computational models of how we had understand

00:48:41.680 --> 00:48:48.260
the circuit to be and suggest that somehow whatever role this is playing it actually,

00:48:48.920 --> 00:48:54.640
depended on a lack of activity a lack of recent experience now the animal knew

00:48:54.640 --> 00:49:00.340
the maze very well it had been on this maze many many days but on this day it

00:49:00.340 --> 00:49:02.060
was say mostly on the left side,

00:49:02.520 --> 00:49:07.740
we tend to see more right-side replace, which is strange.

00:49:08.500 --> 00:49:12.420
And one possibility is that it's doing it to try to keep everything balanced,

00:49:13.140 --> 00:49:15.540
and trying to say, well, you can't forget what's over there.

00:49:15.640 --> 00:49:17.040
I may need to know it at some time.

00:49:19.057 --> 00:49:22.317
But again, you run into this mechanism problem. What's the mechanism of it?

00:49:22.477 --> 00:49:29.417
It doesn't sound like such a complicated phenomenon to interpret from a network

00:49:29.417 --> 00:49:34.677
perspective if you would allow dynamics of, let's say, short-term plasticity,

00:49:34.717 --> 00:49:37.137
where you say, look, I have known trajectories.

00:49:37.437 --> 00:49:38.997
I'm suppressing their responses.

00:49:39.657 --> 00:49:46.117
I have, let's say, lateral interactions among different locations in this environment I visited.

00:49:46.117 --> 00:49:49.217
Visited they can now become active because

00:49:49.217 --> 00:49:51.917
blah blah we can tell each other

00:49:51.917 --> 00:49:55.057
stories like that's right right yes so then given that

00:49:55.057 --> 00:49:58.177
i can if you want not trivialized but

00:49:58.177 --> 00:50:02.737
sort of interpreted in mechanical terms why don't do you find this such an important

00:50:02.737 --> 00:50:10.317
data point because the short-term plasticity story that we would want to tell

00:50:10.317 --> 00:50:16.057
that we would have assumed to tell is based that recent Recent activity would

00:50:16.057 --> 00:50:17.557
make a cell more likely to fire.

00:50:18.037 --> 00:50:21.937
And in order to explain this specific data point, you have to suggest that the

00:50:21.937 --> 00:50:23.657
cell is less likely to participate.

00:50:24.237 --> 00:50:28.917
No, but I can just use an habituation component. The question then is,

00:50:29.017 --> 00:50:34.497
why on other tasks is the more recent stuff the stuff that gets replayed?

00:50:35.257 --> 00:50:40.437
So the problem here is that I can explain this experiment by creating these

00:50:40.437 --> 00:50:44.957
kind of habituation, adding in habituation parameters and stuff like that.

00:50:45.257 --> 00:50:49.697
But there are other tasks where actually the most recent stuff seems to be replayed.

00:50:50.497 --> 00:50:56.497
So we now have to say, now our parameters for our mechanistic explanation depend

00:50:56.497 --> 00:51:00.537
on task in strange and complex ways. Right, okay.

00:51:00.797 --> 00:51:03.077
And so, yes, that's possible.

00:51:03.357 --> 00:51:05.397
But why?

00:51:05.817 --> 00:51:09.857
It would not be a satisfactory story because it would be ad hoc given one task only.

00:51:09.877 --> 00:51:15.017
Exactly. But then my point is, why would you bring in an interpretation that

00:51:15.017 --> 00:51:16.417
uses the notion of voluntary?

00:51:18.477 --> 00:51:24.177
I was doing that more to point out that it wasn't just simply this noise model.

00:51:24.297 --> 00:51:25.397
Okay, okay. I understand.

00:51:25.717 --> 00:51:31.557
I don't know how voluntary these things are, because I'm not sure what the word voluntary means.

00:51:31.677 --> 00:51:34.897
No, right. That's what I want to ask you. What would be the signatures of voluntary

00:51:34.897 --> 00:51:38.677
operations in the preparation you look at?

00:51:39.600 --> 00:51:45.140
What I would say is that when we look at human sequences and we identify some

00:51:45.140 --> 00:51:49.260
sequences of voluntary and not voluntary in humans, which is its own debate,

00:51:49.500 --> 00:51:51.840
but once we've made that categorization,

00:51:52.120 --> 00:51:56.700
then we can say, what's the information processing in the brain structures in

00:51:56.700 --> 00:52:00.020
humans under these, quote, unquote, voluntary conditions?

00:52:00.740 --> 00:52:06.660
And then we can say, okay, does the rat show the same information processing

00:52:06.660 --> 00:52:07.960
in the same brain structures?

00:52:07.960 --> 00:52:12.800
And so then the question is, well, you could either argue humans are not voluntary,

00:52:12.820 --> 00:52:18.280
or you can argue that rats are voluntary, but you're no longer allowed to argue

00:52:18.280 --> 00:52:19.920
humans are voluntary and rats are not.

00:52:20.040 --> 00:52:24.900
Right. So I think in the episodic memory literature, people are using that in a very specific way.

00:52:25.220 --> 00:52:29.660
So involuntary memory is when things just come into your mind,

00:52:29.760 --> 00:52:33.420
and you didn't intend that to come into your mind.

00:52:33.460 --> 00:52:37.160
And it may be an unpleasant memory of something that's associated with what's

00:52:37.160 --> 00:52:40.860
happening. Whereas voluntary autobiographical memories, when you specifically

00:52:40.860 --> 00:52:45.480
try to remember what you were doing last Friday or Christmas Eve or whatever,

00:52:45.620 --> 00:52:47.540
so you're working to retrieve that memory.

00:52:48.260 --> 00:52:53.240
And that would be very interesting, obviously, if we had any evidence at all

00:52:53.240 --> 00:52:55.360
that a rat could do something like that.

00:52:55.360 --> 00:53:00.700
Well, one of the data points that we're still kind of finalizing the data on,

00:53:00.740 --> 00:53:04.240
so I don't want to make too much of it yet. We're still looking at this.

00:53:04.360 --> 00:53:07.300
We have an abstract at SFN, for example, talking about this.

00:53:07.880 --> 00:53:09.300
So it's going to be out soon.

00:53:10.120 --> 00:53:17.480
Is that it looks like hippocampus is being pulled during these moments by other brain structures.

00:53:18.420 --> 00:53:23.360
So it looks like, in fact, prefrontal cortex in particular is basically saying

00:53:23.360 --> 00:53:26.860
to the hippocampus, go find stuff.

00:53:27.180 --> 00:53:32.440
So in a sense, it's possible that that could get closer to your voluntary story,

00:53:32.600 --> 00:53:37.720
that some other system is actually saying to hippocampus, I need to remember this now.

00:53:39.540 --> 00:53:42.420
Go figure this out for me you need some way

00:53:42.420 --> 00:53:46.000
of ignoring the the sensory context and

00:53:46.000 --> 00:53:51.100
saying this is the context i want you to consider uh construct a a scenario

00:53:51.100 --> 00:53:56.180
starting from there but is there a way to actually i'm trying to think if there's

00:53:56.180 --> 00:54:00.040
a way so what we'd want to do is have a way of essentially shutting off the

00:54:00.040 --> 00:54:03.160
sensory context and still seeing that at this moment.

00:54:05.995 --> 00:54:12.495
And that's hard to do because it's really hard to shut off sensory context in a controlled way.

00:54:12.715 --> 00:54:18.315
It does enter the hippocampus over very specific pathways that are anatomically

00:54:18.315 --> 00:54:19.995
well-defined and not mixed.

00:54:20.235 --> 00:54:23.855
Yes. So this might suggest that this allows you to, let's say,

00:54:23.895 --> 00:54:25.595
segment these input streams.

00:54:26.635 --> 00:54:29.215
Yes. Right? Biasing towards, let's

00:54:29.215 --> 00:54:33.775
say, a memory component or a sensory component or an action component.

00:54:34.095 --> 00:54:37.035
Yes. Would you buy that? Yes. Okay. Absolutely.

00:54:37.695 --> 00:54:42.735
It's a very hard experiment to do, though. But yes. We're hoping that you will do it soon.

00:54:43.395 --> 00:54:48.415
But now, we looked at this. So Tony mentioned this vacation.

00:54:51.135 --> 00:54:56.175
Error is trial and error. Vacation is trial and error. And I was getting confused here.

00:54:57.055 --> 00:55:02.215
But it's not, you just not only looked at that behavior for historical reasons.

00:55:02.215 --> 00:55:06.535
You have tied it down in a very specific way to the memory dynamics and the

00:55:06.535 --> 00:55:08.015
task performance of the animal.

00:55:08.435 --> 00:55:12.635
So that means it's only in a very specific moment in the task that rats actually

00:55:12.635 --> 00:55:14.975
display this behavior. It's not that they do it all the time.

00:55:15.375 --> 00:55:19.395
So what is exactly the structure that you found there?

00:55:19.395 --> 00:55:25.755
So actually, Tolman actually saw this as well, that vicarious trial and error

00:55:25.755 --> 00:55:28.375
tends to happen during early learning stages,

00:55:28.635 --> 00:55:36.235
during stages when the animal knows the environment that he needs to work on

00:55:36.235 --> 00:55:41.055
or that it needs to work on, because both males and females do vicarious trial and error.

00:55:43.115 --> 00:55:47.935
That it knows the environment that it needs to work on, but it doesn't quite

00:55:47.935 --> 00:55:49.535
know what to do on that environment.

00:55:49.735 --> 00:55:53.595
And once it actually gets enough experience that it no longer,

00:55:53.715 --> 00:55:58.075
essentially no longer needs to search, and it says, I know at this moment I'm

00:55:58.075 --> 00:56:02.715
going to go left, then the vicarious trial and error goes away.

00:56:03.775 --> 00:56:09.035
And we can put it back. We can reinitiate it by actually doing a reversal reversal,

00:56:09.055 --> 00:56:13.735
where we force the animal and say, well, what you've been doing all this time doesn't work anymore.

00:56:14.715 --> 00:56:18.215
I think of actually, you know, my classic example is driving to work.

00:56:18.575 --> 00:56:22.915
The first time you're driving to work, you're planning, you're paying attention,

00:56:23.215 --> 00:56:26.095
you're thinking about where you're going to go, you're doing all this deliberative stuff.

00:56:26.495 --> 00:56:30.135
But if you do it every day for weeks, months, or years, suddenly you're doing

00:56:30.135 --> 00:56:32.395
it, you're doing it in your sleep, you're doing it, you know,

00:56:32.395 --> 00:56:35.035
you're driving your friend to your office instead of the airport because you

00:56:35.035 --> 00:56:35.975
got that into a good conversation.

00:56:36.335 --> 00:56:40.415
You know, luckily in Minnesota, the airport's close enough that that didn't

00:56:40.415 --> 00:56:41.295
mean he missed his flight.

00:56:42.735 --> 00:56:47.135
He was from New York and was really worried where that would have been a disaster.

00:56:47.295 --> 00:56:49.275
Right, exactly. Um, I don't know,

00:56:49.773 --> 00:56:55.073
But actually, so for example, in my drive to work, they actually closed one

00:56:55.073 --> 00:57:00.433
of the roads, and suddenly I had to recognize at this one intersection,

00:57:00.733 --> 00:57:03.253
which was actually significantly before the road got closed,

00:57:03.493 --> 00:57:06.673
that I needed to turn a different way.

00:57:07.153 --> 00:57:13.113
And it was actually very interesting kind of seeing my own internal system about

00:57:13.113 --> 00:57:17.053
identifying at this moment, I have to, I found myself doing a lot of vicarious

00:57:17.053 --> 00:57:19.853
trial and error, luckily in my head and not with the car.

00:57:20.453 --> 00:57:24.413
But, you know, actually this kind of reversal does trigger that.

00:57:24.613 --> 00:57:27.093
You were hoping for the chocolate to fall from the sky as well.

00:57:27.333 --> 00:57:29.713
It would be nice if it was European chocolate, not American chocolate.

00:57:29.713 --> 00:57:36.293
Yeah. But look, so the point is you said, okay, deliberation requires both search and evaluation.

00:57:36.673 --> 00:57:41.353
Yes. And you see this vicarious trial and error as a signature of search.

00:57:41.653 --> 00:57:43.233
Yes. That's correct. Yes. Okay.

00:57:43.553 --> 00:57:49.133
So are you suggesting with that that this is really an active strategy to obtain

00:57:49.133 --> 00:57:52.213
information from the environment? No. Okay.

00:57:53.013 --> 00:57:57.913
That's an open question. I don't think so, and I don't think so in large part

00:57:57.913 --> 00:58:03.213
because we have also seen vicarious trial and error in situations where the

00:58:03.213 --> 00:58:05.593
reward site is actually out of sight of the animal,

00:58:05.693 --> 00:58:08.093
so down tunnels that the animal has to go.

00:58:09.813 --> 00:58:17.013
And in our environments, the environmental signatures are very, very different.

00:58:17.213 --> 00:58:23.493
It's not a discrimination cue. So it's not a lot of difficulty for the the animal

00:58:23.493 --> 00:58:26.173
to know what's on the left side or what's on the right side.

00:58:26.433 --> 00:58:32.493
So I think this is actually more a reflection of an internal process in which

00:58:32.493 --> 00:58:36.493
the animal is trying to figure out what the consequences of going down that path are.

00:58:36.813 --> 00:58:39.633
So it's not so much that they're trying to trigger the sensory,

00:58:39.753 --> 00:58:45.733
although it might be that they're trying to physically trigger this involuntary memory sequence,

00:58:46.013 --> 00:58:50.793
or it may just be that actually the voluntary memory sequence,

00:58:51.073 --> 00:58:54.353
if I can use that distinction the voluntary memory

00:58:54.353 --> 00:58:57.433
sequence is actually triggering an initial motion

00:58:57.433 --> 00:59:00.333
right that is kind of saying well okay let's start

00:59:00.333 --> 00:59:03.153
to go wait no i don't want to go left right and i think

00:59:03.153 --> 00:59:06.053
that may well be what a lot of this vicarious trial and error is

00:59:06.053 --> 00:59:10.533
okay but so so you're saying it's not necessarily an orienting response to obtain

00:59:10.533 --> 00:59:15.553
information it's more like an action that you use as a recall cue well it's

00:59:15.553 --> 00:59:23.233
either an action as a recall cue or an epiphenomenon of the of the The internal orienting response,

00:59:23.353 --> 00:59:26.353
which is also driving kind of the beginning of the action. Right.

00:59:26.473 --> 00:59:27.793
And I don't know which one it is.

00:59:28.393 --> 00:59:31.433
Or it doesn't need to be an exclusive choice. It might be both,

00:59:31.613 --> 00:59:33.733
right? That's right. That's right. Mm-hmm.

00:59:36.081 --> 00:59:44.581
But then, can we really call that deliberation as such in these experiments, right?

00:59:44.621 --> 00:59:49.761
So also in your case, so the road to the airport or to your work is closed,

00:59:49.881 --> 00:59:53.141
and now suddenly you have to sort of find a new route.

00:59:53.641 --> 00:59:58.761
So you could say, well, look, I'm just obtaining sensory information. I'm consulting memory.

00:59:59.221 --> 01:00:04.061
I'm performing actions. It could also be like you search through a list of alternative

01:00:04.061 --> 01:00:09.921
options as opposed to really, really actively modeling the consequences of future actions.

01:00:10.301 --> 01:00:16.301
Well, we know that they are in fact actively modeling the consequences of those

01:00:16.301 --> 01:00:20.121
future actions because we know that there are direct representations through those options.

01:00:20.121 --> 01:00:25.721
What I think the real key is that we also know that there's reward extra information,

01:00:26.221 --> 01:00:29.601
particularly in the ventral striatum, particularly in the areas of ventral striatum

01:00:29.601 --> 01:00:34.521
that hippocampus projects to, that reflect those reward information.

01:00:34.521 --> 01:00:39.701
So not only is there a representation of those future options,

01:00:40.001 --> 01:00:44.861
but there's also a representation of the value or the reward that they're going

01:00:44.861 --> 01:00:50.081
to get at the ends of those options, which suggests that it actually is a search

01:00:50.081 --> 01:00:52.461
and evaluate process. Right. Okay.

01:00:53.221 --> 01:00:59.561
So now, we learn a lot now about the specific memory dynamics of the hippocampus,

01:00:59.561 --> 01:01:03.761
but as you already indicated, hippocampus doesn't operate in isolation, right?

01:01:03.821 --> 01:01:06.341
It is part of a circuit of a loop, essentially.

01:01:06.481 --> 01:01:10.141
It's a loop with many other parts of the brain. And at some point,

01:01:10.201 --> 01:01:15.061
you showed us at least how, in your view, this loop is interconnected with some

01:01:15.061 --> 01:01:18.101
other key areas in cortex and subcortical areas.

01:01:18.241 --> 01:01:23.421
So how do you see that overall loop play out? What do the different stages in this loop do?

01:01:24.121 --> 01:01:28.801
Well, the key, I think, that we have, I mean, there's issues of the loop from

01:01:28.801 --> 01:01:31.661
the circuitry perspective of how does information get in from,

01:01:31.721 --> 01:01:35.081
for example, entorhinal cortex and the different parts of entorhinal cortex.

01:01:35.081 --> 01:01:38.921
And we could certainly talk about that.

01:01:39.161 --> 01:01:44.941
But I think at this point from this deliberation moment, what we know is that

01:01:44.941 --> 01:01:49.141
hippocampus contains the information searching through that future.

01:01:50.761 --> 01:01:53.761
That it reflects that information.

01:01:54.001 --> 01:01:58.221
And the evidence, at least both from humans when they have damaged hippocampi

01:01:58.221 --> 01:02:04.481
and from other animal experiments where people have manipulated hippocampus

01:02:04.481 --> 01:02:08.201
suggest that hippocampus is critical to that future construction.

01:02:08.741 --> 01:02:14.721
We know that the, we have some evidence that there's prefrontal information coming in.

01:02:15.561 --> 01:02:18.981
We have some information that the ventral striatum is doing evaluation,

01:02:19.441 --> 01:02:25.261
but at this point how they all interact on a moment by moment basis is actually

01:02:25.261 --> 01:02:27.461
something that we don't know right now.

01:02:27.561 --> 01:02:30.501
And that's actually something we're very excited to start going to look for.

01:02:30.661 --> 01:02:32.921
But now in that circuit that you delineate,

01:02:34.056 --> 01:02:38.916
Which of these four different regions you mentioned would have the biggest impact

01:02:38.916 --> 01:02:40.256
when it would be lesioned?

01:02:40.536 --> 01:02:43.416
Would it be the hippocampus? So it would be ventral striatum?

01:02:43.436 --> 01:02:44.356
Would it be prefrontal cortex?

01:02:44.716 --> 01:02:49.916
They would all have different impacts. Of course. But which one would be the most critical?

01:02:50.296 --> 01:02:52.816
Well, critical for what? Solving the task.

01:02:54.256 --> 01:02:57.356
I think they're all involved in solving the task. They're all involved differently.

01:02:57.656 --> 01:03:01.576
So, for example, I think that without ventral striatum, although,

01:03:01.636 --> 01:03:05.356
to be honest, we have not done this experiment, I would predict that without

01:03:05.356 --> 01:03:09.976
eventual striatum, you'd have a lot of trouble differentiating values of rewards, right?

01:03:10.056 --> 01:03:14.516
And we know from other people have found that it's critical for recognizing evaluation.

01:03:17.096 --> 01:03:21.896
We know that, again, from other people, that orbitofrontal cortex,

01:03:22.116 --> 01:03:25.936
which is another structure involved in a lot of this, is very critical when

01:03:25.936 --> 01:03:32.716
you have different flavors, when you have to integrate across multiple reward information, right?

01:03:32.716 --> 01:03:38.896
That it's very important when you're doing logic from A implies B,

01:03:39.156 --> 01:03:43.456
B implies C, C implies I get reward, so I need to actually go to A.

01:03:43.996 --> 01:03:49.016
These kinds of logical chains, you need orbitofrontal cortex to be able to make

01:03:49.016 --> 01:03:50.156
these kind of logical chains.

01:03:53.416 --> 01:03:56.856
The how these structures actually

01:03:56.856 --> 01:04:00.216
are i mean they're all critical for the task and

01:04:00.216 --> 01:04:04.196
the key is that they're going to each play a different computational role and

01:04:04.196 --> 01:04:09.376
so the way i look at it is it's more a question of how does the computation

01:04:09.376 --> 01:04:14.676
of the animal change when you've taken that out of the circuit yeah but you

01:04:14.676 --> 01:04:18.456
also indicated yourself during your talk that for many of these tests,

01:04:18.616 --> 01:04:22.276
you can just remove the hippocampus and you can still solve it.

01:04:22.436 --> 01:04:25.536
You might be a bit less efficient, it might take you a bit longer to acquire,

01:04:26.796 --> 01:04:29.456
but the way we test at least the rat brain,

01:04:30.627 --> 01:04:34.527
In many of these tasks, you can actually manage your hippocampus. Yes.

01:04:34.607 --> 01:04:39.087
So I think the key here is that, and this is why we have multiple decision systems, right?

01:04:39.207 --> 01:04:43.047
Because in fact, there are many ways to solve many of these problems.

01:04:44.407 --> 01:04:53.087
One of my favorite examples is Varga Khatam's data on children who have damaged hippocampi.

01:04:53.087 --> 01:04:58.007
I believe they have congenital issues where they have no hippocampus.

01:04:58.007 --> 01:05:02.987
They actually do okay in school, but it turns out that they do okay in school

01:05:02.987 --> 01:05:09.187
by working completely semantically, and they work in school by a very different process.

01:05:09.267 --> 01:05:12.827
They actually are able to pass their classes, and they do well and all that

01:05:12.827 --> 01:05:19.667
stuff, but they have a completely different process by which they solve their behaviors, right?

01:05:19.707 --> 01:05:23.127
And I think that's why we have these many, many systems, right?

01:05:23.187 --> 01:05:27.627
Because evolutionarily, right, we didn't usually have a doctor we could go fix, right?

01:05:28.007 --> 01:05:34.427
So, when a person solves or an animal solves a task without a hippocampus,

01:05:34.427 --> 01:05:35.287
they do it very differently.

01:05:35.627 --> 01:05:38.687
And in fact, you can provide, if you get the right probe trial,

01:05:38.887 --> 01:05:43.907
you can actually construct probe trials whereby they behave differently, right?

01:05:43.907 --> 01:05:48.107
One of my favorite examples is the Tolman-Hull debate on the T-Maze,

01:05:48.307 --> 01:05:53.407
where they trained animals to go from the south arm to the west arm of a T-Maze.

01:05:53.867 --> 01:05:58.247
And they both did it identically. And if you look at the path from the south

01:05:58.247 --> 01:06:00.447
arm to the west arm, they look the same.

01:06:00.847 --> 01:06:03.907
But if you put the animal on the north arm, so I should say,

01:06:04.067 --> 01:06:07.387
Tolman argued that the animals were going to the place.

01:06:08.347 --> 01:06:12.427
Tolman said, I know where I am. I'm at the south arm. I know where I want to

01:06:12.427 --> 01:06:16.267
be. I want to be at the west arm. How do I get there? This is Tolman's explanation.

01:06:17.107 --> 01:06:20.027
Hull says, no, no, no, it's all stimulus response. The animal says,

01:06:20.127 --> 01:06:21.147
I'm on the maze. I turn left.

01:06:22.347 --> 01:06:24.887
Now put the animal on the north arm of a plus. Paul Jay,

01:06:25.873 --> 01:06:30.813
So Tolman's animals will say, I'm on the north arm, I want to be on the west

01:06:30.813 --> 01:06:34.313
arm, I want to make a different action, turn right and go to the same place.

01:06:35.013 --> 01:06:39.793
Hull's animals will say, I'm on the maze, turn left. They end up at a different

01:06:39.793 --> 01:06:42.413
location. It's not that one of these is right and one of these is wrong.

01:06:42.953 --> 01:06:46.113
It's that this is two ways of solving the task computationally.

01:06:46.533 --> 01:06:51.913
And in fact, turns out that what happens is that early on, the animals look

01:06:51.913 --> 01:06:53.773
like Tolman. and late, the animals

01:06:53.773 --> 01:06:56.813
look like Hull and they actually switch from one system to the other.

01:06:56.933 --> 01:06:59.553
But the Tormund-like system can be training the other one.

01:06:59.833 --> 01:07:04.553
There's models by, for instance, Richard Sutton where he has reinforcement learning

01:07:04.553 --> 01:07:08.853
going on but he also builds a forward model and very much in the same way as

01:07:08.853 --> 01:07:12.853
your rats, he plays sequences through the forward model and uses that to train

01:07:12.853 --> 01:07:14.353
the reinforcement learning system.

01:07:14.773 --> 01:07:22.253
Yes. So, and do you see that as happening then in the rat brain? So, I suspect yes.

01:07:23.413 --> 01:07:28.193
We know, for example, we've done models, we've built models where this kind

01:07:28.193 --> 01:07:33.753
of consolidated replay will train up a downstream structure that can learn a

01:07:33.753 --> 01:07:36.753
more habit-based kind of action chain story.

01:07:37.953 --> 01:07:42.353
One of the experiments I've always wanted to run, I've never actually run it,

01:07:42.413 --> 01:07:45.233
mostly because I just haven't done it yet.

01:07:45.533 --> 01:07:50.933
And if somebody else wants to do it, it's fine, is to run one of these T-Maze

01:07:50.933 --> 01:07:54.013
plus maze experiments where the animals switch from one to the other,

01:07:54.133 --> 01:07:58.993
but actually to train them for only a limited time and then put them back in

01:07:58.993 --> 01:08:04.273
their home cage and leave them for several weeks and then test them. Do they switch?

01:08:04.973 --> 01:08:08.433
Right? Is it actually the experience that creates the switch?

01:08:08.733 --> 01:08:11.013
Is it the mental part that creates the switch?

01:08:11.893 --> 01:08:14.813
Is it time? We don't know that answer.

01:08:15.913 --> 01:08:21.213
But we know that when somebody is learning from, actually, the example we just

01:08:21.213 --> 01:08:25.353
came up with, I was just talking to an interview from Sports Illustrated,

01:08:25.433 --> 01:08:26.773
which is an American magazine.

01:08:26.993 --> 01:08:31.153
They're asking me about American football players who have to learn these big,

01:08:31.173 --> 01:08:35.753
large playbooks, which of course are all Xs and Os and declarative memory.

01:08:36.673 --> 01:08:40.373
And yet they have to actually execute it in procedural memory on the field.

01:08:40.933 --> 01:08:44.873
And so there's this very interesting transition that these people have to do,

01:08:44.933 --> 01:08:51.453
right, by imagining and simulating and practicing and learning and do that transition.

01:08:52.673 --> 01:08:58.533
It definitely happens. What the exact processes are is a fascinating question.

01:08:58.773 --> 01:09:00.233
But David, I would like to...

01:09:01.289 --> 01:09:05.389
Come back to your earlier remark where you said, well, yes, it's true,

01:09:05.489 --> 01:09:08.369
you can lesion hippocampus, animals can still perform the task,

01:09:08.489 --> 01:09:09.889
but maybe that just shows redundancy.

01:09:10.589 --> 01:09:14.529
But isn't that a little bit too easy? Because you could argue,

01:09:14.609 --> 01:09:18.169
look, hippocampus does stand out on anatomical grounds.

01:09:18.309 --> 01:09:21.669
It's a very unique kind of organization that you will actually not really find

01:09:21.669 --> 01:09:22.849
in that way anywhere else in the brain.

01:09:22.849 --> 01:09:29.209
As you said yourself the mind is the brain so that should imply that that anatomical

01:09:29.209 --> 01:09:32.909
structure that structure is telling us something about a unique function yes

01:09:32.909 --> 01:09:38.029
and I want to the second part is that it's a two pronged attack okay so ahead

01:09:38.029 --> 01:09:40.149
the second the second prong here is that.

01:09:42.069 --> 01:09:47.069
Maybe this implies that the tests we are using to understand the function of

01:09:47.069 --> 01:09:50.309
that structure are just not sensitive enough yes maybe the tests we're using

01:09:50.309 --> 01:09:57.689
are just two let's say course to really titrate out the specific contribution of hippocampus.

01:09:57.729 --> 01:10:02.409
And as a result, we are forced into this redundancy interpretation because our

01:10:02.409 --> 01:10:03.769
test is just not sensitive enough.

01:10:04.049 --> 01:10:09.109
I want to be careful with the word redundancy. I'm sorry if my use of the word

01:10:09.109 --> 01:10:12.489
redundancy implied that they were all equivalent. I don't mean that at all.

01:10:13.369 --> 01:10:20.709
The computation by which the other systems actually perform is very different.

01:10:20.869 --> 01:10:25.469
And so an animal, for example, I mean, giving the Tolman Hull example that I

01:10:25.469 --> 01:10:30.289
showed, an animal with a hippocampal lesion, in fact, looks Hullian very quickly,

01:10:30.389 --> 01:10:35.129
much more quickly than an animal actually with dorsal striatal damage,

01:10:35.289 --> 01:10:39.989
who kind of stays Tolmanian for a longer time, right? Right.

01:10:40.529 --> 01:10:45.809
I, the task, the key is that these tasks are not clinical tasks and we should

01:10:45.809 --> 01:10:50.389
not be saying that the Morris water maze is a hippocampal task. It's not.

01:10:51.468 --> 01:10:56.128
The hippocampus is necessary to solve the Morris water maze in a very specific way.

01:10:56.648 --> 01:11:01.668
But yet, and so if you train the Morris water maze in a different way, right?

01:11:01.708 --> 01:11:05.548
For example, you train the animal for months and months, or you train the animal

01:11:05.548 --> 01:11:08.688
using this very large platform shrinking down to a small platform,

01:11:08.868 --> 01:11:14.488
you can train an animal to do the task without a hippocampus, right?

01:11:14.548 --> 01:11:19.088
So the key is to actually think of what is the computation the hippocampus is

01:11:19.088 --> 01:11:23.648
performing? how does that computation then get used in this task?

01:11:24.108 --> 01:11:29.068
And can we construct a probe trial which will differentiate that computation

01:11:29.068 --> 01:11:30.388
from other computations?

01:11:30.808 --> 01:11:37.268
And I actually think, in fact, when I say redundancy, what I mean is many of

01:11:37.268 --> 01:11:38.988
the things we have to do in life

01:11:38.988 --> 01:11:45.368
are able to or depend on are able to be solved in these multiple ways.

01:11:45.628 --> 01:11:49.868
But I don't mean to imply that they're being solved in the same way at all. Okay.

01:11:50.968 --> 01:11:55.408
So the last set of experiments you presented to us, which I thought were really

01:11:55.408 --> 01:11:58.948
very exciting, was dealing with the notion of regret in rats,

01:11:59.168 --> 01:12:01.668
which seems very counterintuitive. Yes.

01:12:02.428 --> 01:12:06.288
So what does regret really mean in the case of a rodent?

01:12:06.528 --> 01:12:10.688
So let me be very careful with what we actually found in this data,

01:12:10.768 --> 01:12:14.488
because, of course, it's easy to kind of go way overboard with the term.

01:12:14.488 --> 01:12:20.428
What we found is in situations in which we would expect it to induce regret

01:12:20.428 --> 01:12:27.168
in humans, the animals behave differently, and their neurophysiology is different.

01:12:27.548 --> 01:12:33.008
That behavior proves that they understand their own agency, that they understand

01:12:33.008 --> 01:12:35.768
that they made a mistake, and they recognize that mistake.

01:12:37.008 --> 01:12:41.288
And so we were able to differentiate that. And we were able to show that during

01:12:41.288 --> 01:12:44.428
those moments, there are representations of the previous event,

01:12:44.628 --> 01:12:47.668
that is the previous moment when they made the previous decision.

01:12:48.767 --> 01:12:53.867
And whether the animal feels regret at that moment is kind of the thing that

01:12:53.867 --> 01:12:55.987
makes all the popular media happy, right?

01:12:56.067 --> 01:13:00.107
But the truth is what we actually found is there's different information processing

01:13:00.107 --> 01:13:05.907
happening in conditions where the animal recognizes it made a mistake by its own agency.

01:13:06.247 --> 01:13:10.007
And a case where the animal experiences a similar set of cues or an equivalent

01:13:10.007 --> 01:13:16.447
set of cues, but makes a mistake not by, the mistake is not of its own agency.

01:13:16.927 --> 01:13:22.567
And that we then saw that the information processing tracks that counterfactual,

01:13:22.707 --> 01:13:25.507
which is critical to human regret.

01:13:25.967 --> 01:13:31.007
But maybe we should try to understand the task a little bit better, right?

01:13:31.047 --> 01:13:38.247
So we have a circular arena or a little tunnel through which the animal runs. It's actually open.

01:13:38.847 --> 01:13:44.307
Okay, it's open field? Well, it's a circular arena, but it's a racetrack. Okay, yeah, right.

01:13:44.407 --> 01:13:47.147
It's a racetrack. It's a racetrack. So there's cues everywhere is kind of the key.

01:13:47.247 --> 01:13:50.707
So at four equally spaced points along this circular track,

01:13:50.867 --> 01:13:55.767
you then have little zones, which you call restaurants if you want,

01:13:55.867 --> 01:14:00.947
where the animal can sample a certain type of food reward or a certain flavor.

01:14:01.067 --> 01:14:02.287
Do they smell it or do they eat it?

01:14:02.467 --> 01:14:05.007
They eat it. They actually eat it. And they're different flavors,

01:14:05.087 --> 01:14:09.327
and each flavor remains at a constant location throughout the entire training.

01:14:09.427 --> 01:14:14.147
So the animal knows that in the southwest corner, there is going to be chocolate.

01:14:14.367 --> 01:14:19.587
Exactly. And now you saw across the animals you tested, they have individual preferences.

01:14:19.967 --> 01:14:23.707
Right. So there's an important step before we get to the individual preferences,

01:14:23.867 --> 01:14:27.767
which is that every time the animal encounters one of these zones,

01:14:27.967 --> 01:14:32.587
enters a restaurant, we like to say, there's a tone,

01:14:32.867 --> 01:14:37.487
a pit, where the pitch of the tone tells the animal how long he's going to have to wait for food.

01:14:37.607 --> 01:14:41.707
And that allows the animal to make a decision to either stay or go. Right.

01:14:41.827 --> 01:14:46.367
Right? And so then, because animals have thresholds for each of these different

01:14:46.367 --> 01:14:50.967
flavors, we can identify their revealed preferences. So the animals will stay.

01:14:52.017 --> 01:14:56.257
If it's less than, if the offer, right, the cost of this restaurant,

01:14:56.417 --> 01:15:00.677
of this food pellet is going to be small enough, then the animal will stick around.

01:15:00.997 --> 01:15:04.177
And if it's going to be too long a delay, the animal skips it.

01:15:04.297 --> 01:15:08.817
Right. And so that threshold allows us to measure how willing the animal is

01:15:08.817 --> 01:15:10.977
to spend its time for that food.

01:15:11.117 --> 01:15:16.197
Yeah. And what we found is that different animals had different thresholds for

01:15:16.197 --> 01:15:19.037
different flavors, but that they were consistent with an animal.

01:15:19.437 --> 01:15:21.997
Right. So that one animal would wait a long time for chocolate,

01:15:22.177 --> 01:15:27.557
which tells us that that animal is willing to spend more time on chocolate than on other things.

01:15:27.777 --> 01:15:32.017
Right. And also what's interesting there, in the behavioral signature,

01:15:32.237 --> 01:15:35.277
they either decide to wait or they move on.

01:15:35.397 --> 01:15:38.237
It's not that they start waiting and then interrupt the waiting. Correct.

01:15:38.737 --> 01:15:43.877
So now what's the longest waiting time these rats are supposed to,

01:15:43.957 --> 01:15:49.577
are willing to suffer? So for two of the rats, we only went up to 30 seconds

01:15:49.577 --> 01:15:51.117
because that gave us enough range.

01:15:51.337 --> 01:15:56.737
But for two of the rats, actually, the rats were willing to wait every time

01:15:56.737 --> 01:15:59.477
at 30 seconds. And so we went up to 45 seconds.

01:15:59.677 --> 01:16:02.057
And that was long enough. But

01:16:02.057 --> 01:16:06.057
we've seen rats wait some of the trials even 45 seconds for some cases.

01:16:06.777 --> 01:16:13.697
But then you also impose a timeout. So a total of 60 minutes to just consume

01:16:13.697 --> 01:16:17.597
whatever they can get. That's right. And then the point is that you're manipulating

01:16:17.597 --> 01:16:19.877
now these waiting times because you know the threshold value.

01:16:20.237 --> 01:16:24.497
That's right. So in that way, you can create, if you want, disappointment or regret.

01:16:24.817 --> 01:16:29.657
That's right. The difference being disappointment is when you have unexpectedly,

01:16:29.877 --> 01:16:34.757
let's say, changed the property they might find that is different from what

01:16:34.757 --> 01:16:35.437
they thought it would get.

01:16:36.197 --> 01:16:41.897
Right. Well, it's actually a random distribution from the 1 to 30 seconds.

01:16:41.897 --> 01:16:46.017
So it's kind of on the tail end of that distribution, that they just kind of,

01:16:46.017 --> 01:16:50.077
they know it's somewhere in this thing, and they kind of get the bad end of the deal.

01:16:50.737 --> 01:16:55.437
So a good deal will be a short waiting time for a reward you like,

01:16:55.557 --> 01:16:59.277
and a bad deal is a long waiting time for a reward you don't like.

01:16:59.737 --> 01:17:04.457
That's right. Yeah. So, but now, so now we understand the task,

01:17:04.717 --> 01:17:06.037
and then the question is,

01:17:07.276 --> 01:17:11.896
Is it fair to interpret this in terms of agency? Because in some sense,

01:17:12.036 --> 01:17:14.936
agency implies that there is

01:17:14.936 --> 01:17:19.776
a knowledge of the causal relationship of the agent with the environment.

01:17:19.916 --> 01:17:22.376
So you can say, yeah, I did it. It was my choice. That's right.

01:17:25.076 --> 01:17:29.736
So what you observe is that when animals hit this point where they get,

01:17:29.916 --> 01:17:32.316
let's say, they have to wait longer than expected.

01:17:33.116 --> 01:17:39.336
Right. By their own choice. You see certain signatures that you interpret as

01:17:39.336 --> 01:17:41.676
indicating regret. So what are these specific signatures?

01:17:41.936 --> 01:17:45.136
Well, the key is actually, you have to actually look, in order to get the regret

01:17:45.136 --> 01:17:48.256
on this task, you have to actually look at a pair of samples.

01:17:48.776 --> 01:17:53.276
And they have to actually have skipped a good deal to reach a bad deal.

01:17:53.736 --> 01:17:57.876
And it's the skipping a good deal that, so we know by the fact that they either

01:17:57.876 --> 01:18:02.256
stay or go, that they're making a decision, right? And the fact that they skip

01:18:02.256 --> 01:18:05.556
a good deal means they've made a decision that is against their preferences.

01:18:06.376 --> 01:18:10.436
And now they encounter a bad deal. And because they're time limited,

01:18:10.676 --> 01:18:12.436
it means they've messed up.

01:18:12.676 --> 01:18:17.816
They've made a mistake and they've erred where they should have taken the good

01:18:17.816 --> 01:18:20.216
deal. And they're not allowed to go backwards.

01:18:20.316 --> 01:18:24.556
Once they've left a deal, the deal is rescinded. No more, you know,

01:18:24.636 --> 01:18:27.656
it's only good while you're in the restaurant. Once you leave,

01:18:27.996 --> 01:18:30.916
you have to, you know, basically you have to go all the way around and try again,

01:18:31.016 --> 01:18:32.636
and it could be a completely different deal.

01:18:34.016 --> 01:18:37.856
So what we see is that at the moment when they have this mistake,

01:18:38.036 --> 01:18:40.156
where they've made a mistake and now they hit a bad deal,

01:18:40.376 --> 01:18:45.476
they'll stop, they look backwards, and at that moment, the orbitofrontal cortex

01:18:45.476 --> 01:18:51.216
and the ventral striatum represent the moment of entering the previous restaurant.

01:18:51.216 --> 01:18:54.676
And in the same way that we talked at the beginning of the podcast about this

01:18:54.676 --> 01:18:59.296
decoding, we're doing the same decoding operation, and this time not for the

01:18:59.296 --> 01:19:02.896
location of the animal, but for entering each of these different zones.

01:19:03.196 --> 01:19:08.796
And so we can say this is a good representation of the previous zone. Mm-hmm.

01:19:09.878 --> 01:19:13.418
Which is a mental time travel to that moment. Yeah, so in some sense,

01:19:13.458 --> 01:19:18.118
what the mental time travel entails is that you recall, let's say,

01:19:18.138 --> 01:19:20.718
the value, essentially, of that other zone.

01:19:20.898 --> 01:19:26.058
That's right. So it's more like you call up a reference, like,

01:19:26.078 --> 01:19:31.178
okay, but if I would have stuck it out on the other side, this is what I would have gotten.

01:19:31.478 --> 01:19:35.238
Well, so what's interesting is that we didn't actually see a very strong,

01:19:35.298 --> 01:19:38.238
although there's very strong representations of reward in these structures,

01:19:38.238 --> 01:19:42.478
That is, at the moment of reward, the cells fire, you know, a subset of cells

01:19:42.478 --> 01:19:44.578
will fire massively for each different flavor,

01:19:44.778 --> 01:19:48.058
telling us very differentiable what each reward is.

01:19:48.238 --> 01:19:53.638
At this moment of, quote, regret, unquote, the animal did not represent the

01:19:53.638 --> 01:19:56.518
reward it should have gotten. That is, it didn't represent the reward.

01:19:57.138 --> 01:20:00.798
It actually represented the entry point into the other restaurant.

01:20:01.118 --> 01:20:03.858
Yeah, but wait, if you decode that from the ventros triatum,

01:20:03.958 --> 01:20:07.618
as you said earlier, that must reflect some sense of valuation.

01:20:08.238 --> 01:20:10.698
Or not? That's a hypothesis.

01:20:11.778 --> 01:20:14.818
But given the literature, it would be consistent. Given the literature,

01:20:14.818 --> 01:20:17.058
it's quite likely that it's some sense of valuation.

01:20:17.338 --> 01:20:20.078
Actually, I suspect the orbital frontal cortex is also some sense of valuation.

01:20:20.078 --> 01:20:21.158
Exactly right. Yes, absolutely.

01:20:21.498 --> 01:20:25.018
But we don't know that. What we know is that at that moment,

01:20:25.138 --> 01:20:28.558
there's a representation of that moment in the task. Right.

01:20:29.258 --> 01:20:34.898
Whether there's actually a valuation judgment associated with it would be very, very interesting.

01:20:34.898 --> 01:20:38.338
Interesting one of the things i'd really like to look at is

01:20:38.338 --> 01:20:41.518
whether there is some relationship between

01:20:41.518 --> 01:20:45.118
the self-firing patterns and the the value of

01:20:45.118 --> 01:20:48.678
each of the rewards but it's very noisy and it's very hard to decode with the

01:20:48.678 --> 01:20:52.678
limited data we have right exactly my hope is that if as we gather more data

01:20:52.678 --> 01:20:56.818
we'll be able to actually determine whether in fact there's value representation

01:20:56.818 --> 01:21:01.518
at that moment if it's just the moment that's being represented and what right

01:21:01.518 --> 01:21:05.078
and of course one One of the interesting questions is, what's Hippocampus doing at that moment? Yeah.

01:21:06.070 --> 01:21:10.670
But now you could also argue that your task is like a rat gambling task and

01:21:10.670 --> 01:21:12.630
that you misinterpret it in some sense, right?

01:21:12.690 --> 01:21:16.790
Because I could also say, well, if I'm the rat, I'm here standing in front of

01:21:16.790 --> 01:21:20.430
this, the zone with the banana flavor, which I really detest.

01:21:20.450 --> 01:21:21.510
I want to get to the chocolate.

01:21:21.990 --> 01:21:26.210
So I don't care what he offers me. For me, it's not a bad deal because I want

01:21:26.210 --> 01:21:28.910
to get to the chocolate. And then at the chocolate, you offer me a bad deal.

01:21:29.210 --> 01:21:33.230
So I'm not regretting anything I did with the banana because I'm not interested in banana.

01:21:33.230 --> 01:21:38.650
But you would still in relative terms just interpret the local event because

01:21:38.650 --> 01:21:40.710
you say well you decided not to

01:21:40.710 --> 01:21:44.510
wait there and relative to the delay i was giving you it was a good deal,

01:21:45.010 --> 01:21:48.190
and you still didn't take it and that the chocolate gave you a bad deal but

01:21:48.190 --> 01:21:51.830
the rat is saying look i'm not interested in banana i don't care about your

01:21:51.830 --> 01:21:56.350
deal i want to just get the chocolate well but we don't see the rat behavior

01:21:56.350 --> 01:22:02.370
look like that that is the rats actually take all four deals when When they're good deals, right?

01:22:02.530 --> 01:22:07.150
It's not, their thresholds, the differences are five seconds out of the 30.

01:22:08.010 --> 01:22:12.630
The other thing is that the next offer after, let's say it's chocolate to banana

01:22:12.630 --> 01:22:18.150
to cherry to plain, after the banana, it's gonna, he's got two more offers before

01:22:18.150 --> 01:22:19.310
he's coming back to the chocolate.

01:22:19.710 --> 01:22:24.250
And in fact, we know from these representations that when he's left the banana

01:22:24.250 --> 01:22:27.810
and he's done with that, he's thinking about the next one, which is the cherry,

01:22:27.990 --> 01:22:31.950
not the one that he previously came. So normally, it's going to be representing

01:22:31.950 --> 01:22:33.830
what's next in my sequence.

01:22:35.170 --> 01:22:40.250
So I don't think... The point is that only in these very specific conditions

01:22:40.250 --> 01:22:43.370
do you see this representation of the previous choice.

01:22:43.730 --> 01:22:47.530
And one of the things that I think is important about this task is because there

01:22:47.530 --> 01:22:51.330
are four options, it's not just, well, I'm not thinking about...

01:22:51.330 --> 01:22:52.570
I don't want to be where I am now.

01:22:52.830 --> 01:22:56.910
It's actually thinking about every specific case of that previous option.

01:22:57.510 --> 01:23:00.550
Whereas again, normally the end will be thinking about the next option.

01:23:02.073 --> 01:23:06.773
So, you've presented evidence that rats can feel disappointment,

01:23:07.013 --> 01:23:12.373
maybe regret, that they can mentally time travel, potentially even quite far

01:23:12.373 --> 01:23:15.293
in the future, and imagine the rewards or punishments they might get.

01:23:15.633 --> 01:23:19.113
So as a scientist, you must obviously think, well, what are the implications

01:23:19.113 --> 01:23:21.693
here for how we use animals in our research?

01:23:25.173 --> 01:23:28.453
Yeah, well, I think that the key, I don't think it changes anything,

01:23:28.553 --> 01:23:33.593
because I think the question still has to be, how do we make sure that our animals

01:23:33.593 --> 01:23:34.893
are treated as well as possible?

01:23:35.193 --> 01:23:39.653
That we want to make sure that every experiment we do is fully justified.

01:23:40.073 --> 01:23:45.493
That we want to make sure that unless we're studying a stress condition,

01:23:45.733 --> 01:23:47.633
we don't want to be stressing our animals.

01:23:47.873 --> 01:23:50.253
I mean, I know I think differently under stress.

01:23:51.013 --> 01:23:54.513
So if I want to understand how normal behavior is happening,

01:23:54.653 --> 01:23:57.453
I want to understand how a non-stressed animal.

01:23:57.593 --> 01:24:01.553
So I think it's always a question of you know

01:24:01.553 --> 01:24:05.133
really thinking about this question and always asking yourself is

01:24:05.133 --> 01:24:07.953
this experiment important is this

01:24:07.953 --> 01:24:11.293
actually something that has to happen and I think that's a valid question and

01:24:11.293 --> 01:24:14.953
I think that we every scientist I know asks that question as they're doing experiments

01:24:14.953 --> 01:24:20.013
I guess so that one thing that we might have done in the past is think well

01:24:20.013 --> 01:24:26.053
it's a rat it's not going to worry about what's happening to it tomorrow but you know we

01:24:26.213 --> 01:24:30.173
now have this extra consideration perhaps that these animals are able to think

01:24:30.173 --> 01:24:33.713
back on things that have happened in the past or maybe look forward to events

01:24:33.713 --> 01:24:35.973
in the future and therefore we have to be more careful.

01:24:37.597 --> 01:24:42.557
Yes, but to be honest, I think that intuitively, we've always known this.

01:24:42.737 --> 01:24:47.697
You ask a pet owner, that pet owner has always believed that their animals are

01:24:47.697 --> 01:24:51.377
emotional, intuitive, you know, connecting.

01:24:51.517 --> 01:24:57.897
And I think the whole animal experimental question has always included that. And I think it has to. do.

01:24:58.757 --> 01:25:05.937
I think that it's easy to say, well, yes, there's been a historical science,

01:25:06.037 --> 01:25:10.537
but we go back to the Harlow experiments, those horrible Harlow experiments.

01:25:11.077 --> 01:25:19.217
He was arguing the importance of this from an emotional perspective, that this is critical.

01:25:19.597 --> 01:25:21.717
I mean, there's wonderful data.

01:25:22.617 --> 01:25:26.277
Deborah Blum in her book Love at Goon Park talks extensively about this.

01:25:26.337 --> 01:25:29.417
I think it's one of the the best books on the whole animal issue,

01:25:30.137 --> 01:25:34.237
talks about how every infant who has survived the NICU unit,

01:25:34.357 --> 01:25:40.157
the neonatal intensive care unit, owes its life to Harry Harlow and those nightmare experiments.

01:25:41.157 --> 01:25:46.657
Because the NICU units changed, and the survival rate went from basically nothing

01:25:46.657 --> 01:25:49.277
to tremendous, to very successful.

01:25:49.457 --> 01:25:53.617
Because people started actually, they started actually touching the infants,

01:25:53.837 --> 01:25:57.157
right? Before that, it was assumed that you couldn't touch an infant because

01:25:57.157 --> 01:25:58.557
it'd be a sterile problem.

01:25:58.697 --> 01:26:00.937
But then human infants need contact.

01:26:01.737 --> 01:26:04.977
So we changed. We changed the NICU units.

01:26:05.117 --> 01:26:09.457
And those NICU units changed because of Harry Harlow and those experiments that

01:26:09.457 --> 01:26:11.117
are incredibly hard to justify.

01:26:12.517 --> 01:26:17.917
So this is a very difficult and complex issue, and is one that I think needs

01:26:17.917 --> 01:26:24.017
to be thought of in terms of the importance of this, and how do we connect that up?

01:26:24.357 --> 01:26:27.917
And to be honest, I'm not sure that this data changes.

01:26:28.917 --> 01:26:33.057
It convinces scientists that, you know, the animals really do this,

01:26:33.117 --> 01:26:35.877
and I think it tells us a lot about how these processes work,

01:26:36.077 --> 01:26:37.797
which is, to me, the really important thing.

01:26:38.377 --> 01:26:43.137
That we understand a lot more, I think now, about how these processes work,

01:26:43.257 --> 01:26:46.257
which means, of course, we can start to ask what happens when they go wrong, right?

01:26:46.457 --> 01:26:51.757
And in fact, a lot of human psychiatry, for example, is fundamentally dependent

01:26:51.757 --> 01:26:55.237
on breakdowns in computation, right?

01:26:55.297 --> 01:26:59.337
The way to really think about psychiatry, I think, is to think of it as failure

01:26:59.337 --> 01:27:00.857
modes of an engineering system.

01:27:01.197 --> 01:27:04.117
But we can't know what the failure mode is until we know what the engineering

01:27:04.117 --> 01:27:08.417
system is. So where do you see the impact of this work in, for instance,

01:27:08.517 --> 01:27:10.017
treating human mental illness?

01:27:10.397 --> 01:27:14.897
Well, I think that, for example, if we can identify the fundamental structures,

01:27:15.197 --> 01:27:18.917
that is the fundamental computational components that are driving things like

01:27:18.917 --> 01:27:22.497
psychiatry, then we will actually be able to understand treatments better.

01:27:22.577 --> 01:27:25.297
We'll understand what the actual symptoms are.

01:27:25.417 --> 01:27:27.997
One of the things I like to say, we've been doing a lot of work on addiction,

01:27:28.157 --> 01:27:31.557
actually, in trying to understand what is human addiction.

01:27:31.837 --> 01:27:36.977
And I like to say addiction is a symptom, not a disease. that there's actually lots of diseases,

01:27:37.057 --> 01:27:42.017
lots of dysfunctions in the decision-making and other systems that we can identify

01:27:42.017 --> 01:27:47.137
from first principles now that we know how these computations work.

01:27:47.197 --> 01:27:49.157
We know better how these computations work.

01:27:49.417 --> 01:27:52.897
And that then leads us into being able to start to say, well,

01:27:52.937 --> 01:27:55.457
okay, that's not a cocaine addict.

01:27:55.737 --> 01:27:59.357
That's somebody who has an evaluation problem in the deliberation system.

01:27:59.957 --> 01:28:02.777
Now, of course, we have to figure out how do you fix an evaluation problem in

01:28:02.777 --> 01:28:06.917
your deliberation system. But at least we know what it is, right?

01:28:07.317 --> 01:28:11.137
And there are many, there are cases, there's discussion happening.

01:28:11.277 --> 01:28:15.397
I mean, right now, this is exactly where the field is, with this new term called

01:28:15.397 --> 01:28:16.797
computational psychiatry.

01:28:16.917 --> 01:28:19.997
I have to say, I'm not fond of that term, but it is the term.

01:28:20.097 --> 01:28:23.977
And the idea is to take this new computational understanding of decision making

01:28:23.977 --> 01:28:27.957
systems, and really this engineering view of the brain, right?

01:28:28.017 --> 01:28:34.177
The brain as a system that's It's doing things explicitly and has a physical

01:28:34.177 --> 01:28:36.377
process, running a physical computation,

01:28:36.657 --> 01:28:39.877
and identifying where the failure modes, where the fault lines are,

01:28:39.957 --> 01:28:44.317
and then connecting that up. And there's a conversation happening.

01:28:44.737 --> 01:28:49.777
My hope is that this can change treatment and even definitions of what some

01:28:49.777 --> 01:28:50.937
of these dysfunctions are.

01:28:52.199 --> 01:28:59.399
The short answer is it looks like the answer is yes, but we don't know it yet.

01:28:59.479 --> 01:29:01.059
I suspect we're five or ten years away.

01:29:01.619 --> 01:29:06.919
And one of the hopes, I guess, for the more distant future is that one of the

01:29:06.919 --> 01:29:10.879
treatments we might be able to have is to replace damaged circuits with artificial

01:29:10.879 --> 01:29:12.899
circuits, what people call neuroprostheses.

01:29:13.219 --> 01:29:17.619
And people are talking about hippocampus as a potential target for neuroprothetics.

01:29:17.859 --> 01:29:20.639
I mean, what do you think about that? Is that a realistic possibility?

01:29:20.639 --> 01:29:23.899
Oh, I certainly think neuroprosthetics are a realistic possibility.

01:29:25.419 --> 01:29:28.599
I think actually the first things that are going to come, though,

01:29:28.699 --> 01:29:31.619
are going to be better understandings of learning systems.

01:29:32.059 --> 01:29:36.819
And that as we understand how these systems learn and modify themselves, right?

01:29:36.899 --> 01:29:39.599
I mean, anytime you interact with somebody, you are changing the brain,

01:29:39.679 --> 01:29:43.279
right? The fact that somebody remembers anything means the brain has changed.

01:29:43.599 --> 01:29:48.099
So I think there's ways to do it that are less physically invasive,

01:29:48.339 --> 01:29:50.239
and I think those are going to happen first.

01:29:51.379 --> 01:29:56.539
I think that we're going to see a lot kind of small... I think the first things

01:29:56.539 --> 01:29:58.819
that are going to happen is small modifications of treatments,

01:29:59.039 --> 01:30:05.139
where we say, oh, well, actually, this treatment depends on working memory computations.

01:30:05.379 --> 01:30:11.579
So if we just added a working memory training component, that will change how this treatment works.

01:30:11.579 --> 01:30:16.279
I think that's going to be the first steps, because neuroprostheses require

01:30:16.279 --> 01:30:20.739
not only the understanding of the computational process, but a second engineering

01:30:20.739 --> 01:30:25.459
piece of how do you actually interface with the brain, which is non-trivial.

01:30:25.619 --> 01:30:29.999
And lots of people are working on that, but I think there's a complicated second

01:30:29.999 --> 01:30:32.179
step that's going to have to happen there. Mm-hmm.

01:30:32.939 --> 01:30:37.939
So now, in the beginning, when you sort of had to, in a nutshell,

01:30:37.999 --> 01:30:41.859
define what you're trying to achieve, you said that you're trying to decode

01:30:41.859 --> 01:30:42.979
the mind from brain states.

01:30:43.339 --> 01:30:46.299
Yes. Right? But if we now look at the data….

01:30:47.184 --> 01:30:50.884
And let's say a bit more detached cynical perspective. I can say,

01:30:50.904 --> 01:30:54.464
oh, great, but you've been decoding neuron states from brain states.

01:30:54.924 --> 01:31:00.504
Where is the mind, right? How do we get that link to mind?

01:31:00.724 --> 01:31:06.184
And I see that you're trying to get there also looking at these high-level constructs

01:31:06.184 --> 01:31:09.264
like regret, for instance, right? Or agency, right?

01:31:09.464 --> 01:31:13.604
But this is also if you want creating a risk because now if people,

01:31:13.764 --> 01:31:18.904
if you would not be able to really nail that, Because these contracts are just very complicated.

01:31:19.764 --> 01:31:23.424
You're left in this position that all you've been doing then is decoding neuron

01:31:23.424 --> 01:31:24.444
states from brain states.

01:31:24.704 --> 01:31:28.524
Right. So how do we really cross that bridge? How are we going to do that?

01:31:28.624 --> 01:31:32.044
How confident are you that we're actually close in reaching that goal?

01:31:32.424 --> 01:31:35.104
I'm actually pretty confident that we are reaching that goal.

01:31:35.184 --> 01:31:35.784
You're absolutely right.

01:31:35.884 --> 01:31:40.304
There's danger in using these terms, particularly when we have about half of

01:31:40.304 --> 01:31:43.024
the term, which is I think what's happening in regret, for example.

01:31:43.024 --> 01:31:46.684
We don't have evidence that the animals feel the emotion of regret.

01:31:47.784 --> 01:31:51.344
We talked a little at the talk whether we could actually figure out how to do that.

01:31:52.024 --> 01:31:57.504
But for deliberation, I'm much more confident, actually, that we have the pieces of that. Yeah.

01:31:58.500 --> 01:32:03.320
I think the key here is that this new viewpoint, and it really is,

01:32:03.440 --> 01:32:08.620
to be honest, new in the last 20 or 30 years, that the way to understand the

01:32:08.620 --> 01:32:11.100
brain is as a computational device,

01:32:11.380 --> 01:32:15.660
and not just in some sort of digital computation, but in the mathematics of

01:32:15.660 --> 01:32:21.740
analog components, that it's actually performing some sort of fundamental computational process.

01:32:22.500 --> 01:32:27.320
And that's that what we call mind is, in fact, also a computational process.

01:32:27.320 --> 01:32:32.780
And that doesn't diminish, in my view, the who you are of a person.

01:32:32.840 --> 01:32:35.680
I'm very happy to be commander data. Sure, I have no problem with that. Right?

01:32:37.060 --> 01:32:42.720
And so, but if we have that computational process, then we should be able to access it.

01:32:42.820 --> 01:32:47.560
And so to me, the key is that those processes make very specific predictions

01:32:47.560 --> 01:32:50.000
about what the neuron state should look like.

01:32:50.320 --> 01:32:54.600
And so we can go in and look for those neuron states, checking those predictions.

01:32:54.600 --> 01:32:59.060
And I think I would not want to say that we just can go in and look, right?

01:32:59.140 --> 01:33:02.900
I think the key here is a full interaction of theory and experiment.

01:33:03.160 --> 01:33:04.520
We need a theoretical neuroscience.

01:33:05.040 --> 01:33:08.180
That would be very good, yes. I agree. But then what is interesting,

01:33:08.380 --> 01:33:14.020
though, is that now you do – actually, you mentioned the word computation a lot. Yes.

01:33:14.400 --> 01:33:18.080
So apparently you see that as a bridging level of description.

01:33:18.720 --> 01:33:22.260
Are you having in mind a specific set of computational operations?

01:33:22.260 --> 01:33:25.600
Or do you use it in a loose way, like some form of transformation?

01:33:26.300 --> 01:33:30.080
I mean in a loose way, in some form of transformation. I do not mean specific.

01:33:30.260 --> 01:33:37.100
What I mean is that mathematics and computation about information is the correct

01:33:37.100 --> 01:33:40.400
way to describe this process.

01:33:40.660 --> 01:33:43.700
And certainly we can also describe the process pharmacologically,

01:33:43.860 --> 01:33:46.360
we can describe it chemically, we can describe it physically.

01:33:46.800 --> 01:33:51.320
But I think that the way to understand what the brain is doing.

01:33:52.289 --> 01:33:55.549
Is through a transformation of information.

01:33:55.769 --> 01:33:59.289
So the idea is to take, in some sense, a computer science view,

01:33:59.429 --> 01:34:02.509
which is that you have representations, you have algorithms.

01:34:02.809 --> 01:34:05.989
The algorithms are not necessarily digital algorithms. The representations are

01:34:05.989 --> 01:34:07.349
not necessarily digital representations.

01:34:07.969 --> 01:34:11.929
But that you have, the question is, what are the representations?

01:34:11.949 --> 01:34:15.749
How are those representations transformed from structure to structure?

01:34:16.009 --> 01:34:21.169
How are those representations encoded? And that that language,

01:34:21.229 --> 01:34:27.749
that mathematical language, is the bridge to connect kind of psychological states with neural states.

01:34:28.229 --> 01:34:33.349
So then before we get to the finish line, in some sense, you could also look

01:34:33.349 --> 01:34:38.469
at your description of the brain as making it actually fairly simple.

01:34:38.589 --> 01:34:42.429
Because we have these different modules that perform certain operations.

01:34:43.069 --> 01:34:48.389
I have a memory in my hippocampus. I have valuation in my obitofrontal cortex, ventral striatum.

01:34:48.929 --> 01:34:51.669
I have some rule-based integration, a prefrontal cortex.

01:34:52.249 --> 01:34:56.069
And as long as I just decode what these different subsystems do,

01:34:56.149 --> 01:34:58.889
I can just glue them together and I understand how the brain works.

01:35:00.129 --> 01:35:03.889
I don't think the glue is so simple. Okay, tell me. Is that where the secret lies?

01:35:04.549 --> 01:35:08.749
No, I think there's important questions in all these components.

01:35:09.009 --> 01:35:11.989
Both, I mean, to be honest, I want to be careful about saying,

01:35:12.069 --> 01:35:13.369
you know, that we've solved the brain.

01:35:13.449 --> 01:35:15.809
We certainly haven't solved the brain. There's a lot of work to do,

01:35:15.849 --> 01:35:22.989
right? To say that we've identified that there exist future representations

01:35:22.989 --> 01:35:27.389
in hippocampus doesn't tell us how those future representations are generated. We talked about that.

01:35:28.249 --> 01:35:30.809
So there's a whole question of how does this computation happen?

01:35:31.189 --> 01:35:35.149
One of the things that's very exciting in terms of the gluing is there's some

01:35:35.149 --> 01:35:39.709
very exciting data coming on suggesting that there's dynamic gluing.

01:35:39.709 --> 01:35:42.589
That structures will talk to other

01:35:42.589 --> 01:35:45.429
structures by matching oscillations and

01:35:45.429 --> 01:35:48.529
by sometimes they'll connect up and sometimes they won't

01:35:48.529 --> 01:35:52.129
uh one of my favorite stories is the

01:35:52.129 --> 01:35:57.889
neuromodulator stories in the invertebrate literature in which neuromodulators

01:35:57.889 --> 01:36:02.169
basically completely rewire the network cells that were oscillating suddenly

01:36:02.169 --> 01:36:05.429
are not oscillating cells that were inhibitory are now excitatory it's almost

01:36:05.429 --> 01:36:08.809
like you have multiple networks hippocampus works the same way,

01:36:09.449 --> 01:36:14.189
In the presence of acetylcholine, the entorhinal cortex is driving most of the inputs.

01:36:14.409 --> 01:36:21.429
The recurrent inputs in CA3 are weak but have stronger LTP.

01:36:21.689 --> 01:36:26.089
In the absence of acetylcholine, right, hippocampus is doing more internal generation.

01:36:26.369 --> 01:36:30.069
The entorhinal inputs are weaker. The recurrent connections are stronger.

01:36:30.089 --> 01:36:33.149
And hippocampus is driving more to the deep entorhinal outputs, right?

01:36:33.269 --> 01:36:36.849
This is work, among other people, by Mike Hasselmo in the 1990s.

01:36:38.309 --> 01:36:44.989
And so you've got this where these computational states are really complicated, right? Right.

01:36:45.149 --> 01:36:49.609
So I don't want to trivialize the module story, right?

01:36:49.689 --> 01:36:53.369
But I still think it's a computation question, right? So the advantage of the

01:36:53.369 --> 01:36:58.249
acetylcholine from the Mike Haslamo story is that it prevents interference during storage.

01:36:58.449 --> 01:37:02.229
It's fundamentally a computational explanation for this process.

01:37:02.829 --> 01:37:09.529
Okay. So tell me, David, look, you're really leading the pack in a lot of this work, I have to say.

01:37:09.769 --> 01:37:15.849
Thank you. And this sort of system-level understanding of cognitive properties of rats.

01:37:16.569 --> 01:37:20.709
So if we'd like to follow in that in your tradition, what would be the radish

01:37:20.709 --> 01:37:25.589
law of brain science that we should adhere to? Yikes.

01:37:29.269 --> 01:37:32.709
That's a hard one. Bring in everything?

01:37:33.789 --> 01:37:41.309
I think to me, the key is to be able to bring in theory, to bring in the computation,

01:37:41.669 --> 01:37:45.569
to bring in the experiments, to have all of it talk to each other,

01:37:45.629 --> 01:37:50.669
and to try to actually build this conjoint interaction.

01:37:50.669 --> 01:37:59.269
And to say, you know, how does the, you know, how does this,

01:37:59.349 --> 01:38:01.949
I mean, you really want a full loop is to me the key.

01:38:02.069 --> 01:38:05.469
You want the theory making a prediction that you then test with the experiment,

01:38:05.589 --> 01:38:10.229
that you connect up with the modeling, that, you know, then changes your theory.

01:38:10.369 --> 01:38:16.149
And this whole cycle and thinking of it as a system is to me the key,

01:38:16.249 --> 01:38:18.049
though I'd hate to call it the Reddish law.

01:38:19.529 --> 01:38:22.509
Okay good then look tony here likes traveling

01:38:22.509 --> 01:38:25.209
um and i don't think he has been

01:38:25.209 --> 01:38:27.929
to minnesota yet so four years from now i'm gonna ship him

01:38:27.929 --> 01:38:31.189
to minnesota uh cool low cost

01:38:31.189 --> 01:38:34.809
but he's gonna get there one way or the other um and

01:38:34.809 --> 01:38:37.849
he's gonna he's gonna have a piece of paper in his hand that said i came here

01:38:37.849 --> 01:38:42.969
to test the hypothesis so what's the one hypothesis that your prediction that

01:38:42.969 --> 01:38:46.729
that you want to make today that Tony will come and check out four years from

01:38:46.729 --> 01:38:51.989
now to see whether you really tested it and what kind of outcome you found.

01:38:54.269 --> 01:38:58.749
Specific prediction specific prediction um.

01:39:02.660 --> 01:39:07.300
I'm not sure I could give one quickly like that, but what I could tell you is

01:39:07.300 --> 01:39:13.000
what the key question is that we'd like to be asking, which is, how do you actually.

01:39:15.020 --> 01:39:19.260
Integrate and decide when you have a conflict between decision systems?

01:39:19.500 --> 01:39:24.160
To me, that's the ignorance question.

01:39:24.360 --> 01:39:29.040
I really love the Stuart Feierstein ignorance point, that what science is about

01:39:29.040 --> 01:39:31.120
is questions, right? And finding the question.

01:39:31.560 --> 01:39:35.780
And to me, particularly questions that you didn't know were questions before

01:39:35.780 --> 01:39:37.300
you started working, right?

01:39:37.400 --> 01:39:40.720
To me, that's the question that I didn't know was a question, right?

01:39:41.080 --> 01:39:44.860
Until I actually had the point, you know, until I, it wasn't me,

01:39:44.920 --> 01:39:48.620
until, you know, the field had gotten to the point where we had these multiple

01:39:48.620 --> 01:39:52.580
decision systems, asking how you interact between decision systems is a meaningless question.

01:39:53.120 --> 01:39:58.100
To me, that's the the question i would love to know if tony comes and says that's

01:39:58.100 --> 01:40:03.940
the question have you answered it i would be ecstatic if i if i had an answer

01:40:03.940 --> 01:40:07.520
to that all right great david radish thank you very much for this conversation

01:40:07.520 --> 01:40:09.280
thank you for having me thank you,

01:40:10.320 --> 01:40:18.760
are we up yeah that was intense the csn podcast was produced by the convergent

01:40:18.760 --> 01:40:22.280
science network of Biometrics and Biohybrid Systems,

01:40:22.580 --> 01:40:27.500
a project funded by the European Sevens Research Framework Program.

01:40:29.060 --> 01:40:34.380
For more interviews, recorded lectures, or upcoming conferences in the field

01:40:34.380 --> 01:40:40.420
of biometrics and biohybrid systems, go to csnnetwork.com.

01:40:40.720 --> 01:40:49.360
Music.