WEBVTT

00:00:03.557 --> 00:00:10.517
This is the Convergent Science Network podcast. Leading researchers in the domain

00:00:10.517 --> 00:00:16.737
of neuroscience, brain theory and technology are interviewed by Paul Verschure and Tony Prescott.

00:00:19.497 --> 00:00:23.257
This is Paul Verschure with the Convergent Science Network podcast together

00:00:23.257 --> 00:00:24.857
with my colleague Tony Prescott.

00:00:25.657 --> 00:00:30.837
And we're here with Matthew Diamond who was speaking at our BCBT Summer School 2015 edition.

00:00:31.797 --> 00:00:36.297
Matthew, you were talking about the whisker system in rats,

00:00:36.897 --> 00:00:44.317
but in some sense you give your analysis of the whisker system an expansion, if you want,

00:00:44.437 --> 00:00:49.037
into a domain that is in some sense counterintuitive when you look at people

00:00:49.037 --> 00:00:52.117
who do work on rodents because we've started to really look at pretty complicated

00:00:52.117 --> 00:00:53.957
decision-making tasks.

00:00:53.957 --> 00:00:56.797
Tasks so so how did you actually get

00:00:56.797 --> 00:00:59.817
from the study of the whisker system and its intricacies

00:00:59.817 --> 00:01:04.537
into this domain of of decision making well

00:01:04.537 --> 00:01:09.857
we it it took a number of steps and i should say that one of uh one of my colleagues

00:01:09.857 --> 00:01:13.817
in this field always tells me that we do behaviors that are far too complicated

00:01:13.817 --> 00:01:18.837
why do we do such difficult why do we train rats in such difficult tasks we

00:01:18.837 --> 00:01:21.937
We would do better with a go-no-go whatever,

00:01:22.137 --> 00:01:24.377
and there's some truth to that.

00:01:25.357 --> 00:01:29.177
But we got there because we're in a cognitive neuroscience department,

00:01:29.317 --> 00:01:33.057
and we've always wanted to study cognition, that is, thinking,

00:01:33.177 --> 00:01:39.357
decision-making, perception, all these things together.

00:01:40.897 --> 00:01:46.737
And when I started in neuroscience, my PhD and as a postdoc,

00:01:46.737 --> 00:01:49.277
there were issues that we wanted to study.

00:01:49.377 --> 00:01:52.077
People knew a long time ago in the 1990s,

00:01:53.035 --> 00:01:56.175
People knew what were the interesting properties of the brain,

00:01:56.495 --> 00:01:59.655
some of the most fascinating properties of the brain, the intriguing things

00:01:59.655 --> 00:02:03.875
like decision-making and perception, but there just weren't ways to do it.

00:02:03.995 --> 00:02:09.855
And so all of us would say to each other and to ourselves, what I really want

00:02:09.855 --> 00:02:13.335
to understand is perception and decision-making, but I can't.

00:02:13.535 --> 00:02:18.275
I don't have the methods, they aren't available, so let's anesthetize the rat

00:02:18.275 --> 00:02:22.115
and give a controlled hold stimulus and measure the response to the stimulus,

00:02:22.215 --> 00:02:23.635
but it was always a compromise.

00:02:23.975 --> 00:02:29.235
And then in the last 10 years, there's been a lot of progress in instrumentation and,

00:02:29.895 --> 00:02:36.775
neurophysiological methods and understanding of how to train rats that has allowed

00:02:36.775 --> 00:02:41.415
us actually to begin to explore the things that we knew for a long time we wanted to do,

00:02:41.975 --> 00:02:43.715
but we just couldn't get at them before.

00:02:44.135 --> 00:02:49.895
It's interesting that the training of animals is actually, some of it is a reinvention

00:02:49.895 --> 00:02:54.515
of the wheel, because a long time ago in experimental psychology.

00:02:55.615 --> 00:03:00.435
People did very, very interesting things with animals, going back to 1920s,

00:03:00.435 --> 00:03:08.115
30s, 40s, and then neuroscience became much more fascinated with neurophysiology.

00:03:08.975 --> 00:03:14.235
Behavioral training was sort of put on the back burner and forgotten,

00:03:14.495 --> 00:03:20.655
and then in the last 10 years, There's been a rebirth of the attempts to study

00:03:20.655 --> 00:03:22.275
interesting behaviors in rats.

00:03:23.595 --> 00:03:28.095
Right. But now for the whisker system, you made this distinction between,

00:03:28.195 --> 00:03:31.995
let's say, active sensing and receptive sensing.

00:03:32.375 --> 00:03:35.255
So why do you think that's an important distinction to make?

00:03:35.255 --> 00:03:43.815
Well, I joined Tony years ago, five, ten years ago, in projects in which we

00:03:43.815 --> 00:03:46.395
focused on what we called active sensing.

00:03:46.755 --> 00:03:50.775
Ehud Aissar was also very important in this way of thinking. And.

00:03:51.496 --> 00:03:57.536
And because active, because the whisker system goes out to explore the world,

00:03:57.576 --> 00:03:58.976
it doesn't wait for things to happen.

00:03:59.036 --> 00:04:06.536
It makes things happen by a very deeply ingrained interaction between sensory

00:04:06.536 --> 00:04:11.896
systems and motor systems. and I bought into that and used that terminology.

00:04:12.476 --> 00:04:17.776
But when one is faced with active sensing, which usually means whisking,

00:04:17.816 --> 00:04:22.116
moving the head, moving the whiskers, and then we say, well,

00:04:22.196 --> 00:04:26.776
what if the animal's not moving the whiskers? What kind of sensing is that?

00:04:26.916 --> 00:04:33.456
And the natural terminology would be passive because it's contrary to active sensing.

00:04:33.836 --> 00:04:39.156
So if the animal is palpating a surface or moving around an arena and feeling

00:04:39.156 --> 00:04:43.496
the walls, and we say that's active sensing, then when it receives a vibration

00:04:43.496 --> 00:04:47.896
without movement, one would be left calling that passive sensing.

00:04:48.136 --> 00:04:52.596
And I didn't think that that was right. My students and postdocs and all of

00:04:52.596 --> 00:04:57.276
us in the laboratory observing the animals did not feel that they were in a passive state.

00:04:58.136 --> 00:05:03.156
We felt that they were actually actively controlling the whiskers in such a

00:05:03.156 --> 00:05:06.476
way as to collect a vibration from an external object.

00:05:06.856 --> 00:05:09.836
And so we refer to that now as receptive sensing.

00:05:10.416 --> 00:05:16.476
And when the animal creates the stimulus by its own movement,

00:05:16.556 --> 00:05:18.156
we call that generative sensing.

00:05:18.576 --> 00:05:20.636
And we believe that both are active.

00:05:21.596 --> 00:05:26.796
So there are two states of the system within the realm of active sensing.

00:05:27.116 --> 00:05:32.556
So there's still passive sensing as another possibility here where there's some

00:05:32.556 --> 00:05:34.276
unexpected stimulus on the whisker.

00:05:34.656 --> 00:05:39.536
Yes, but now from the perspective of the of the whisker system,

00:05:39.676 --> 00:05:43.896
how you use it, aren't you always operating in a mixed mode and isn't there

00:05:43.896 --> 00:05:48.716
to make that clean distinction isn't that only holding in the laboratory and

00:05:48.716 --> 00:05:52.236
it will be more difficult to maintain that under under real world conditions?

00:05:53.558 --> 00:05:57.998
Well, we just don't know enough about real world rats.

00:05:58.218 --> 00:06:00.878
We don't know enough about their natural history, what they do,

00:06:01.058 --> 00:06:03.718
how they do it, how they live, how they move.

00:06:04.898 --> 00:06:09.798
I hope that new research programs will begin on that.

00:06:09.938 --> 00:06:15.558
But as of now, we just don't know what they really do with their sensory systems.

00:06:16.478 --> 00:06:20.058
I think that there may be, as Tony said,

00:06:20.218 --> 00:06:24.678
something that is passive sensing, which means that the animal doesn't know

00:06:24.678 --> 00:06:31.578
what to expect and receives the stimulus that happens without being able to

00:06:31.578 --> 00:06:34.538
prepare the system for what's going to happen because they don't know.

00:06:35.058 --> 00:06:44.378
So for instance, an unexpected arrival of a predator or an unexpected stimulus could be

00:06:44.478 --> 00:06:50.578
processed by a passive sensory system simply because the rat is not able to

00:06:50.578 --> 00:06:53.978
actively set the state of processing.

00:06:55.378 --> 00:07:01.438
So then, let's assume we just sort of, we have the sensor system worked out.

00:07:01.558 --> 00:07:06.618
These rats are really great in using their whisker system for different kinds of discriminations.

00:07:07.858 --> 00:07:14.098
But now you're going to use this modality in actually a pretty intricate task, right?

00:07:14.098 --> 00:07:17.938
Where also you described a number of stages that you want to manipulate,

00:07:18.718 --> 00:07:23.058
which starts controlling, let's say, the attention of the animal to the upcoming stimulus.

00:07:23.598 --> 00:07:29.018
Then there will be a first stimulus, this is a vibration on the whiskers that

00:07:29.018 --> 00:07:31.018
must be encoded. Then there will be a pause.

00:07:31.398 --> 00:07:34.038
So now we are dealing with the working memory task.

00:07:35.038 --> 00:07:41.838
Second stimulus. Then there's a comparison, person, um, a subsequent delay,

00:07:42.098 --> 00:07:45.698
uh, before the action can be executed, then we get the go signal.

00:07:45.798 --> 00:07:48.398
The animal can, can choose left or right to get a reward.

00:07:48.778 --> 00:07:53.798
Yes. So if you just would look at this protocol and you would say,

00:07:53.838 --> 00:07:56.218
look, I'm going to train my rats to do this.

00:07:56.718 --> 00:07:59.958
Uh, I think most people would say, look, look, uh, Matthew, you're mad.

00:08:00.078 --> 00:08:00.958
This is just too complex.

00:08:01.558 --> 00:08:07.278
So, so how much time does it take to get the rat to be really sort of.

00:08:07.898 --> 00:08:11.838
Sufficiently capable of performing such a task.

00:08:13.291 --> 00:08:20.911
Well, there are different questions that you can ask the animal to do in the context of this task.

00:08:21.051 --> 00:08:28.551
But I'd say that for the first level questions, you can get answers with about two months of training.

00:08:29.451 --> 00:08:36.711
If you want to introduce new variations or new variables, it can take as long as three months.

00:08:37.451 --> 00:08:40.791
And it varies from rat to rat. Some are very clever and learn very quickly,

00:08:40.871 --> 00:08:46.311
and some are slower. and we tend to be patient with the slow rats and try to

00:08:46.311 --> 00:08:51.451
train them until the end rather than discarding them. So we see differences between rats.

00:08:52.011 --> 00:08:55.611
But I think it's worthwhile for a number of reasons.

00:08:56.091 --> 00:09:01.711
First of all, just the fact that the rat can do this is informative because,

00:09:01.931 --> 00:09:08.031
as I mentioned during the talk, this is, compared to the primate brain, it's a very small brain.

00:09:08.191 --> 00:09:12.891
It has many fewer modules in the cerebral cortex. So the question is,

00:09:12.951 --> 00:09:18.111
having available a smaller brain and a simpler brain in some way,

00:09:18.271 --> 00:09:21.131
are there things that the rat simply can't do?

00:09:21.571 --> 00:09:23.491
And what can it do?

00:09:24.011 --> 00:09:28.071
And so we're sorting this out. There are a number of laboratories that have

00:09:28.071 --> 00:09:34.031
progressively, in the last five to ten years, introduced into the world of rats

00:09:34.031 --> 00:09:36.151
a series of primate tasks.

00:09:36.151 --> 00:09:40.271
In fact, there was an article in Nature about this three or four years ago,

00:09:40.311 --> 00:09:42.371
I think they called them the rat pack.

00:09:43.331 --> 00:09:49.691
And there was the belief in the field of systems neuroscience,

00:09:50.031 --> 00:09:52.951
cognitive neuroscience, that rats simply couldn't do them.

00:09:53.591 --> 00:09:56.351
But it turns out that it takes a

00:09:56.351 --> 00:09:59.171
lot of training of the research team in order

00:09:59.171 --> 00:10:02.111
to learn how to train the rats so we we train ourselves

00:10:02.111 --> 00:10:04.871
and we learn we learn how to train the

00:10:04.871 --> 00:10:10.471
rats and it's a question of finding the right methods the right the right apparatus

00:10:10.471 --> 00:10:16.351
and once that's done the rats can do very primate like things even the visual

00:10:16.351 --> 00:10:23.331
system we all know that primate visual processing is extremely advanced,

00:10:23.651 --> 00:10:29.831
and we're not saying that rats can have visual perception up to the level of primates.

00:10:30.031 --> 00:10:34.971
But there are some properties of visual processing that five years ago people

00:10:34.971 --> 00:10:38.291
didn't believe that rats have, and it turns out they do have them.

00:10:38.391 --> 00:10:44.071
I'm referring to invariance according to position, according to size,

00:10:44.351 --> 00:10:45.931
according to viewing angle.

00:10:47.091 --> 00:10:53.731
My colleague at CISA, David A. Zocolan has found that rats do have these invariance

00:10:53.731 --> 00:10:57.571
capacities, and people thought they were a property of primates.

00:10:57.751 --> 00:11:01.031
So if you put the time and effort into...

00:11:02.047 --> 00:11:04.847
Finding the right kind of stimuli, the right kind of training apparatus,

00:11:05.207 --> 00:11:05.967
the right training regime.

00:11:06.687 --> 00:11:12.047
Rats can do primate-like things, and this alone tells us something.

00:11:12.387 --> 00:11:19.567
Is this really to use rats as a substitute for primates in brain research,

00:11:19.787 --> 00:11:22.527
or is there some other reason to do this?

00:11:22.827 --> 00:11:28.107
Well, I think that that's one of the outcomes, is to be able to do certain kinds

00:11:28.107 --> 00:11:34.947
of cognitive neuroscience, perceptual neuroscience, in rats and not have to rely on primates.

00:11:34.967 --> 00:11:38.987
That's certainly one benefit of this, but it's not the only reason for doing it.

00:11:39.987 --> 00:11:43.747
Even if there were no constraint on using primates, it's interesting to know

00:11:43.747 --> 00:11:48.007
what rats can do, because they do it with a very different organization of the brain.

00:11:48.467 --> 00:11:54.467
And the fact that a complex task, a primate-like task, can be done with a very

00:11:54.467 --> 00:11:58.407
different brain structure, tells us something about the way the brain works.

00:11:58.487 --> 00:12:02.807
It tells us that the brain, in particular the cerebral cortex,

00:12:03.047 --> 00:12:08.167
can find solutions to doing certain kinds of computations, but in a different way.

00:12:08.827 --> 00:12:14.127
And discovering how the rat brain can accomplish this actually tells us a lot

00:12:14.127 --> 00:12:16.007
about overall brain organization.

00:12:16.507 --> 00:12:21.387
So I'm a bit surprised you say that, because as in part of the argument you'd

00:12:21.387 --> 00:12:25.027
want to make is that though the rat brain is simpler than the primate brain,

00:12:25.027 --> 00:12:32.487
There are some fundamental mechanisms in tasks like decision-making which are probably similar,

00:12:32.747 --> 00:12:36.667
and the neural substrates may even have very interesting similarities.

00:12:37.827 --> 00:12:47.027
Yes, they may. The point is that these mechanisms can be installed in a circuit

00:12:47.027 --> 00:12:49.167
with fewer neurons and with fewer modules,

00:12:49.327 --> 00:12:53.807
and yet somehow the same computation can be accomplished.

00:12:53.987 --> 00:12:59.587
For instance, in what I talked about today, there's a very clear working memory component.

00:13:00.707 --> 00:13:06.067
And if you use the word working memory with systems neuroscientists,

00:13:06.147 --> 00:13:08.447
they immediately think of prefrontal cortex.

00:13:09.187 --> 00:13:12.347
Now, rats may have a prefrontal cortex.

00:13:12.567 --> 00:13:16.767
They probably do, but nobody's absolutely certain what region it is.

00:13:16.887 --> 00:13:22.267
It's very hard to find an analogy or homology between rat prefrontal cortex

00:13:22.267 --> 00:13:24.127
and primate prefrontal cortex.

00:13:24.927 --> 00:13:32.047
Some people believe that the rat homology to prefrontal cortex is a prelimbic

00:13:32.047 --> 00:13:35.927
area along the medial wall of the cortex, but nobody's really sure.

00:13:36.527 --> 00:13:38.387
And yet, with this...

00:13:39.506 --> 00:13:46.206
Possessing or maybe not possessing a prefrontal cortex rats can do uh can carry

00:13:46.206 --> 00:13:52.926
out working memory as well as primates nearly as well as primates and and and and so they're somehow,

00:13:53.766 --> 00:13:57.966
getting these computations done with a different brain structure and that's

00:13:57.966 --> 00:14:04.266
it's interesting to know how they do this but i think your data but maybe we

00:14:04.266 --> 00:14:07.986
should go through that first but i think you're the physiology that you have

00:14:07.986 --> 00:14:12.386
gathered on this task is actually also partially answering that question,

00:14:12.486 --> 00:14:15.426
whether it is that different or whether there are similarities.

00:14:16.526 --> 00:14:20.826
But maybe we can come back to that once we understood the task a little bit better.

00:14:20.906 --> 00:14:28.846
So what now happens is that the animal is exposed to two stimuli.

00:14:29.226 --> 00:14:35.406
These stimuli are a certain velocity of movement of this plate that you manipulate.

00:14:35.406 --> 00:14:40.106
Calculate and the variance of that velocity you control.

00:14:40.486 --> 00:14:46.426
So that means if you want the overall amplitude of the signal you're integrating

00:14:46.426 --> 00:14:49.066
is now controlled, it's different, right?

00:14:49.466 --> 00:14:54.326
But the second parameter you control of your stimulus is now the duration of the stimulus.

00:14:54.506 --> 00:15:00.406
Yes. And the question now is, okay, to what extent can the animal make this

00:15:00.406 --> 00:15:04.886
discrimination and say, oh, this was shorter or longer? Like stimulus two was

00:15:04.886 --> 00:15:06.866
shorter or longer than stimulus one.

00:15:07.006 --> 00:15:11.286
And this informs me whether I should go left or right to get my reward.

00:15:11.486 --> 00:15:15.406
And the reward, there's a liquid reward. Animals are water deprived or it's…

00:15:15.406 --> 00:15:17.446
Yes, it's actually fruit juice.

00:15:17.966 --> 00:15:23.006
Okay. So they're not food or water deprived. They have their water restricted.

00:15:23.966 --> 00:15:29.366
They drink a lot during the task because in a few hundred trials,

00:15:29.366 --> 00:15:35.986
they get almost as much liquid as a rat in its home cage would have in the course of a day.

00:15:36.526 --> 00:15:42.066
Then after the training session, they also have another hour to top up if they feel like it,

00:15:42.792 --> 00:15:46.492
And then they're restricted until the next session.

00:15:46.592 --> 00:15:54.312
So they arrive in the session thirsty, but certainly not in any sense in physiological stress.

00:15:54.832 --> 00:15:58.952
Right. So now we have this parametric control of our stimulus.

00:15:59.132 --> 00:16:03.552
And you made the point that this actually is a qualitatively different task

00:16:03.552 --> 00:16:07.152
than when we have to do some sort of categorical decision making.

00:16:07.972 --> 00:16:13.492
Let's say I show you Tony and then I show you a car and then you can go left or right.

00:16:13.772 --> 00:16:20.432
So what is the principal difference in your mind, certainly from the perspective

00:16:20.432 --> 00:16:27.252
of the red brain, between this categorical decision making and this more parametric decision making?

00:16:27.492 --> 00:16:34.892
Okay. In the parametric decision making, we actually presented two kinds of behavioral paradigms.

00:16:34.892 --> 00:16:39.832
One is in which the rat or the human subjects, we also study humans,

00:16:40.052 --> 00:16:44.552
has to compare this intensity value that you talked about, this variance.

00:16:45.572 --> 00:16:50.572
And in that case, they should measure the physical characteristics of the stimulus,

00:16:50.712 --> 00:16:54.972
but optimally, they should not attend to the duration of the stimulus.

00:16:54.972 --> 00:17:02.052
Then, as a second task, we ask animals and human subjects to judge the duration

00:17:02.052 --> 00:17:07.972
of the stimulus while not attending to the intensity value.

00:17:09.612 --> 00:17:16.292
In both these cases, the comparison is, as you say, a parametric one,

00:17:16.432 --> 00:17:23.412
which means that that every stimulus is distributed along the same dimension

00:17:23.412 --> 00:17:26.732
and vary only in their location along that dimension.

00:17:26.972 --> 00:17:31.492
So we can draw an axis, which we could call variance of the stimulus,

00:17:31.572 --> 00:17:35.092
or intensity or amplitude, we could use different terms, but we can draw an

00:17:35.092 --> 00:17:40.292
axis, and then every stimulus simply has one position along this axis.

00:17:40.872 --> 00:17:47.332
And the object of the brain then is to find the relative position of two stimuli to compare them.

00:17:47.792 --> 00:17:51.372
This is parametric because there's one parameter that defines a stimulus.

00:17:51.572 --> 00:17:54.652
In the case of categorical stimulus comparison.

00:17:55.837 --> 00:18:02.357
Which ecologically is hugely important, but it's a different kind of strategy

00:18:02.357 --> 00:18:07.437
for the experimentalist because the stimuli that have to be compared,

00:18:07.577 --> 00:18:11.617
let's say, living thing in a visual system, in a visual task,

00:18:11.717 --> 00:18:12.597
you could ask the subject,

00:18:12.777 --> 00:18:15.137
do you see a living object or not see a living object?

00:18:15.337 --> 00:18:17.797
And you could show an image for a half second.

00:18:18.217 --> 00:18:20.957
That's an example of a categorical task.

00:18:21.597 --> 00:18:29.857
The stimuli that have to be processed may engage different sets of neurons.

00:18:30.277 --> 00:18:35.657
So the comparison could be between one stimulus that evokes activity in one

00:18:35.657 --> 00:18:41.677
set of neurons, a second stimulus that evokes activity in a second set of neurons,

00:18:41.857 --> 00:18:44.357
and eventually the brain of course

00:18:44.357 --> 00:18:47.517
has to converge the two populations in order to make the comparison.

00:18:47.717 --> 00:18:54.417
But at this stage of our understanding, we don't know in categorical tasks which

00:18:54.417 --> 00:18:59.137
we can't identify exactly the sets of neurons, and we don't know where they converge.

00:18:59.477 --> 00:19:04.117
And so we selected a task in which we could actually study the coding,

00:19:04.197 --> 00:19:11.537
the computations done by neurons in that the exact same set of neurons encodes both stimuli.

00:19:11.977 --> 00:19:15.897
Yeah, but the difference now is, of course, that the neurons that we're going

00:19:15.897 --> 00:19:20.417
to look at, are in some sense forming an analog representation of the stimulus itself.

00:19:20.797 --> 00:19:24.277
Yes. One of the categorical case that would not be the case.

00:19:24.597 --> 00:19:30.957
Well, a categorical representation still has to be built on a preceding parametrically controlled one.

00:19:31.397 --> 00:19:37.237
So is it possible that you are then looking at a lower, let's say,

00:19:37.277 --> 00:19:42.937
processing step in the hierarchy than what would play out when we have categorical decision making?

00:19:43.117 --> 00:19:47.517
Or you think that in this in this part of the rat brain where you're looking.

00:19:48.317 --> 00:19:54.117
That you would also find categorical representations if you knew what to look for? So we don't know.

00:19:55.577 --> 00:20:00.057
I think rats can make, for sure, they can make categorical decisions.

00:20:00.457 --> 00:20:08.417
For example, there's very systematic studies in the olfactory system where rats compare odors.

00:20:08.717 --> 00:20:15.017
These I would consider to be categorical because the odors are different from

00:20:15.017 --> 00:20:16.657
each other. They engage different receptors.

00:20:17.117 --> 00:20:20.797
They create activity in different populations of neurons.

00:20:20.917 --> 00:20:24.017
And then at some stage, perhaps in the olfactory cortex, we're not sure where,

00:20:24.097 --> 00:20:28.257
there's convergence and a decision is made.

00:20:30.660 --> 00:20:37.240
So rats can do this. We've explored in rats and we've had some success.

00:20:37.520 --> 00:20:43.380
We've also done a large set of studies in humans that I didn't present today

00:20:43.380 --> 00:20:46.620
in which the comparison is actually made between modalities.

00:20:46.860 --> 00:20:53.080
So we can have an acoustic stimulus that has some value along a dimension,

00:20:53.080 --> 00:20:58.920
an intensity dimension if you wish, and a tactile stimulus that has another

00:20:58.920 --> 00:21:01.120
value along a dimension,

00:21:01.460 --> 00:21:05.120
and then the subject has to compare the two stimuli in different modalities.

00:21:05.540 --> 00:21:10.300
So the first thing they have to do in order to do this task successfully is

00:21:10.300 --> 00:21:18.100
create some scale whereupon stimuli in different modalities can both be projected.

00:21:18.480 --> 00:21:23.060
So you have to, for instance, if you want to, you could call the scale from 1 to 10.

00:21:23.480 --> 00:21:27.840
So the subject has to learn to put an auditory stimulus along the scale of 1

00:21:27.840 --> 00:21:32.120
to 10, tactile stimulus along the scale of 1 to 10, and then make a comparison.

00:21:32.560 --> 00:21:36.420
Humans can do this. With a lot of work, we got some rats to do this,

00:21:36.460 --> 00:21:37.660
but it's very difficult for them.

00:21:37.940 --> 00:21:45.280
And the rats actually had good days and bad days, depending on how much effort

00:21:45.280 --> 00:21:46.740
they wanted to put into it.

00:21:46.840 --> 00:21:52.720
So it's very hard for rats. It's quite a challenge for humans, but humans can do it.

00:21:52.900 --> 00:22:00.000
So that actually requires, to do the task requires a convergence between the

00:22:00.000 --> 00:22:05.940
auditory system and the tactile system, and a human subject can do this. Right.

00:22:06.600 --> 00:22:10.100
So now, if we look at the stimuli you're using, so these parametric stimuli.

00:22:11.420 --> 00:22:16.280
What does, let's say, the discrimination threshold at which,

00:22:16.280 --> 00:22:19.540
and also the just noticeable difference at which rats operate.

00:22:19.780 --> 00:22:23.980
So how big a difference can they really distinguish between these different stimuli?

00:22:26.580 --> 00:22:36.020
Well, that's a great question. I don't recall that we've actually looked at our stimuli.

00:22:36.180 --> 00:22:40.840
We haven't quantified it exactly in terms of just noticeable difference.

00:22:41.360 --> 00:22:44.320
I think that we could do that computation.

00:22:44.520 --> 00:22:48.380
We haven't. But I think it would require examining

00:22:48.380 --> 00:22:51.980
the psychometric curve and then making

00:22:51.980 --> 00:22:55.060
a model for how far along the psychometric

00:22:55.060 --> 00:22:57.820
curve how different to stimuli would have to be

00:22:57.820 --> 00:23:06.520
to produce a different a measurably different choice in the animal and we have

00:23:06.520 --> 00:23:13.440
the data available we have not made exactly that population okay so now if we

00:23:13.440 --> 00:23:15.780
if we talk about a psychometric curve

00:23:15.920 --> 00:23:21.060
that you defined or extracted from the performance of both rats and humans.

00:23:22.760 --> 00:23:28.440
They're sort of similar, but certainly not identical as far as I got your data.

00:23:28.880 --> 00:23:34.860
So what are the important differences here between the psychometric curve that

00:23:34.860 --> 00:23:41.280
maps the properties of the stimulus to rat performance versus those you find for humans?

00:23:41.280 --> 00:23:46.820
Well, as you saw, the rats can perform the task.

00:23:47.440 --> 00:23:51.820
The average, if we take a set of rats and a set of humans,

00:23:52.000 --> 00:23:55.940
the average value for the human is better than the average value for the rats,

00:23:56.000 --> 00:24:03.040
measured in either as percent correct or as a more derivative measurement is

00:24:03.040 --> 00:24:04.480
the slope of the psychometric curve.

00:24:04.620 --> 00:24:07.980
And these are two standard measures of performance.

00:24:08.980 --> 00:24:16.360
On average, rats are inferior to humans, but there's individual variability

00:24:16.360 --> 00:24:20.860
in both species, in humans and rats.

00:24:21.060 --> 00:24:23.360
And this is one of the many, many interesting things.

00:24:24.726 --> 00:24:31.166
Funny things that we find in rat behavior and rat psychophysics. If you talk to a typical.

00:24:33.266 --> 00:24:40.566
Psychologist, a human psychologist, their impression would be that humans differ

00:24:40.566 --> 00:24:45.626
very much from each other, but a laboratory animal like a rat, they're all the same.

00:24:45.826 --> 00:24:52.606
It turns out that rats vary among the individual variability is as much or more than that in humans.

00:24:52.786 --> 00:24:57.906
So a good rat in our data set, a rat that performs really well,

00:24:57.986 --> 00:25:02.006
is actually better than some of the humans that perform less well.

00:25:02.346 --> 00:25:07.646
So on average the humans are better, but there's very significant overlap in

00:25:07.646 --> 00:25:09.526
the two clouds of performance.

00:25:10.586 --> 00:25:16.046
Wouldn't you expect the rats to have less variability because there's less general

00:25:16.046 --> 00:25:18.726
genetic variability in the population of lab rats?

00:25:19.986 --> 00:25:26.126
That would be our expectation. Our rats are sort of outbred so there is quite a bit of variability.

00:25:26.566 --> 00:25:32.386
And we have completely different behavioral procedures where we confirm a lot

00:25:32.386 --> 00:25:38.386
of variability between rats, but they're not genetically, they're not completely inbred.

00:25:38.966 --> 00:25:41.786
But now in some sense we always make these normative judgments,

00:25:41.906 --> 00:25:46.406
right? We see these psychometric curves, which means we're making assumptions about optimality.

00:25:46.826 --> 00:25:51.366
And it's not not unreasonable to assume that rat brains have been optimized

00:25:51.366 --> 00:25:55.746
for somewhat different, to satisfy somewhat different constraints than human brains.

00:25:55.786 --> 00:26:00.166
So in that sense, the biases and the priors in the rat brain might already be

00:26:00.166 --> 00:26:02.666
different from the human brain and maybe given those priors,

00:26:02.666 --> 00:26:04.046
the rat is behaving optimally.

00:26:04.366 --> 00:26:10.386
So what do you see as the most significant differences in the biases that the

00:26:10.386 --> 00:26:13.626
human brings to bear to the task and that the rat brings to bear on the task?

00:26:14.751 --> 00:26:21.631
Well, the rate of learning is, of course, radically different.

00:26:21.751 --> 00:26:28.111
In the case of human subjects, we have a training session in which they are

00:26:28.111 --> 00:26:30.711
presented with the stimuli in the task, and they make a choice,

00:26:30.831 --> 00:26:33.811
and the computer tells them if they are correct or incorrect.

00:26:34.051 --> 00:26:39.251
We don't give them a verbal instruction to tell them what to extract from the

00:26:39.251 --> 00:26:41.911
data. We don't tell them, this is what you're going to feel.

00:26:42.471 --> 00:26:46.351
This is a parameter of the stimulus that you have to extract and then you have

00:26:46.351 --> 00:26:48.631
to compare this parameter in the first to the second.

00:26:48.831 --> 00:26:52.411
We don't do that because we think it would create lots of biases and lots of

00:26:52.411 --> 00:26:56.571
top-down processes and too much thinking.

00:26:57.531 --> 00:27:03.431
So instead we simply let them feel the stimuli, they receive the stimuli,

00:27:03.731 --> 00:27:07.011
they make a choice and the computer says correct or incorrect.

00:27:07.471 --> 00:27:14.571
So So during the training session, the human subjects are in fact testing possible

00:27:14.571 --> 00:27:18.291
hypotheses about what the rule might be, but we don't tell them the rule.

00:27:18.631 --> 00:27:25.831
Then, trying various rules, we're not sure exactly what the mental rules that

00:27:25.831 --> 00:27:29.551
they test are, but trying various ones,

00:27:29.751 --> 00:27:37.111
they eventually arrive at a choice based on the thing that the computer actually

00:27:37.111 --> 00:27:39.711
employs as the rule, and they start getting it right.

00:27:40.231 --> 00:27:45.971
And so when we see them get about 80% right in blocks of 10,

00:27:46.131 --> 00:27:51.711
then we judge them as having figured out the task and the training is over and

00:27:51.711 --> 00:27:52.631
they can move to testing.

00:27:52.831 --> 00:27:56.191
So in humans, this can take 10, 20, 30 minutes.

00:27:56.831 --> 00:28:02.191
In rats, the training takes, as we said before, one or two months.

00:28:02.191 --> 00:28:12.031
They have to go through multiple stages of training, beginning with simply letting them explore the box.

00:28:12.511 --> 00:28:19.131
Then they have to learn that they have to go into a specific sector of the box

00:28:19.131 --> 00:28:21.031
and put their nose in a specific place.

00:28:21.371 --> 00:28:26.351
We can't tell them to do that the way we can tell human subjects to hold their finger here.

00:28:26.931 --> 00:28:31.311
We just have to train the rats to do that, and each of these takes a week or

00:28:31.311 --> 00:28:35.391
two. So when you get through all the stages, two months have passed.

00:28:35.471 --> 00:28:38.431
So that's a major difference in the training.

00:28:38.631 --> 00:28:41.791
In the performance, one difference you noticed is that.

00:28:42.848 --> 00:28:48.108
On very easy stimuli, stimuli that we know that the sensory system can process correctly.

00:28:48.608 --> 00:28:54.948
Humans can perform at 95% or even 100%. Rats, even for very easy stimuli,

00:28:55.328 --> 00:28:59.508
will make the wrong choice 10 or 15% of the time.

00:28:59.608 --> 00:29:02.348
And this is something that psychologists sometimes call lapse rate,

00:29:02.488 --> 00:29:07.728
which means errors that are not due to sensory processing, because we think

00:29:07.728 --> 00:29:10.428
that the sensory system really can handle those trials.

00:29:10.688 --> 00:29:15.148
So the lapse rate is something that distinguishes rats from humans,

00:29:15.288 --> 00:29:19.508
and there are many theories about why rats make lapses.

00:29:19.508 --> 00:29:23.948
One of the most interesting theories is that when they make this kind of error,

00:29:24.068 --> 00:29:29.088
what they're actually doing is making sure that the rule has not changed.

00:29:29.228 --> 00:29:33.428
That is, if they go to the opposite side to where they know they'll get the

00:29:33.428 --> 00:29:37.228
reward, that if they go the opposite side, they want to confirm that there There

00:29:37.228 --> 00:29:38.628
is, in fact, not a reward there.

00:29:39.008 --> 00:29:43.088
So it means rats might be more tuned to exploring the task space than to keep

00:29:43.088 --> 00:29:46.408
on exploring the task space because they might have evolved for environments

00:29:46.408 --> 00:29:49.648
where the contingencies are less stable as humans have.

00:29:49.948 --> 00:29:54.188
Absolutely. I think humans, if they discover that there's a rule,

00:29:54.328 --> 00:30:00.368
then they are willing to act according to this rule until they have evidence that it doesn't work.

00:30:00.488 --> 00:30:07.548
Whereas rats are much more prone to exploring the environment because perhaps

00:30:07.548 --> 00:30:11.548
they've evolved to believe that environments can change, that rules do change,

00:30:11.728 --> 00:30:14.508
and they want to make sure that something hasn't changed.

00:30:14.988 --> 00:30:19.848
Do you also believe that based on that rats would operate at a different level

00:30:19.848 --> 00:30:22.148
of, let's say, anxiety and certainty than humans?

00:30:24.348 --> 00:30:30.768
Possibly. I mean, they may be more anxious on the basis that they're working

00:30:30.768 --> 00:30:36.668
for their daily water. they get as much as they need.

00:30:36.768 --> 00:30:44.468
If they don't get enough liquid during the testing session, then they can drink

00:30:44.468 --> 00:30:45.608
for a couple of hours afterwards.

00:30:46.568 --> 00:30:51.188
They're not deprived of water in any way, but they know it's something that's

00:30:51.188 --> 00:30:53.428
important to them and they want to get it right.

00:30:54.388 --> 00:30:58.728
And surprisingly, humans also want to get it right, but for different reasons.

00:30:59.148 --> 00:31:04.588
They don't want to disappoint you, I think. Yes, and most of the subjects are

00:31:04.588 --> 00:31:07.428
students, and I think that they want to be better than the other students.

00:31:07.908 --> 00:31:09.308
I think there's some rivalry.

00:31:09.888 --> 00:31:15.088
Right. So we have these psychometric curves now for rats and humans.

00:31:15.288 --> 00:31:19.088
We know rats are slightly worse than humans, but not dramatically so.

00:31:19.968 --> 00:31:24.268
The overall psychometric curve is a bit flatter than in the human case.

00:31:24.308 --> 00:31:28.168
Also, the extremes are a bit more compressed as far as I could tell.

00:31:28.448 --> 00:31:32.908
Yes. So, that basically means we know now the rat can perform the task.

00:31:33.028 --> 00:31:34.888
Yes. So, now we want to know, okay...

00:31:35.907 --> 00:31:39.927
How do then these two properties of the stimulus that you're manipulating,

00:31:40.147 --> 00:31:44.767
which is intensity and duration, affect task performance, right?

00:31:44.847 --> 00:31:46.847
So what can you say about that relationship?

00:31:47.507 --> 00:31:53.567
Well, the parameter that we first explored was the perception of the value of

00:31:53.567 --> 00:31:55.467
intensity, the parameter of intensity.

00:31:55.467 --> 00:32:04.247
But we arranged the experiment in such a way that intensity is not a value that

00:32:04.247 --> 00:32:06.607
emerges instantaneously from the stimulus.

00:32:06.767 --> 00:32:12.347
It's a stochastic stimulus. It's a vibration, but a vibration which is created

00:32:12.347 --> 00:32:17.927
by sampling a normal distribution, so you could call it modulated noise in a way.

00:32:18.567 --> 00:32:26.267
So the effect of this is that at any given instant, there's not really a complete

00:32:26.267 --> 00:32:29.507
signal available to the sensory system in order to make a choice.

00:32:29.887 --> 00:32:37.447
Instead, the information necessary to make a choice, to make the judgment, accumulates over time.

00:32:37.627 --> 00:32:39.647
That's the nature of a stochastic stimulus.

00:32:40.427 --> 00:32:44.687
Stochastic stimulus is defined by probabilities,

00:32:45.067 --> 00:32:49.867
by statistical distribution, and with an increasing number of samples,

00:32:50.027 --> 00:32:55.807
the statistical structure becomes more evident to the brain,

00:32:55.987 --> 00:32:57.447
to whoever sampling the stimuli.

00:32:57.867 --> 00:33:04.167
So from this, our prediction is that if the brain is is actually accumulating evidence.

00:33:05.627 --> 00:33:11.227
And using the history of events during the stimulus, then the performance would

00:33:11.227 --> 00:33:13.247
be better with a longer duration.

00:33:13.547 --> 00:33:18.227
And this is what we found in both in rats and in humans.

00:33:18.747 --> 00:33:26.567
You said that there was some evidence that maybe rats wouldn't be good at accumulation in that way.

00:33:26.787 --> 00:33:31.807
Yes, there have been other perceptual tasks that have been explored in rats,

00:33:32.727 --> 00:33:35.627
uh change detection kinds of tasks in

00:33:35.627 --> 00:33:41.747
which the there's been evidence that under some conditions rats do not use a

00:33:41.747 --> 00:33:47.367
long history of of tactile stimulation in order to make the choice but use only

00:33:47.367 --> 00:33:53.727
very recent history and apparently this depends on the uh the way that the animal,

00:33:54.807 --> 00:34:00.647
what it decides to optimize in the task so is it possible that that your results

00:34:00.647 --> 00:34:05.567
are consistent with those other findings because your task emphasizes you can't

00:34:05.567 --> 00:34:07.087
solve your task without evidence accumulation.

00:34:08.649 --> 00:34:16.049
Yes, it's true that the task was actually set up in such a way as to reward

00:34:16.049 --> 00:34:18.069
evidence accumulation.

00:34:18.189 --> 00:34:24.729
It was set up with a statistical structure such that performance is better if

00:34:24.729 --> 00:34:25.689
evidence is accumulated.

00:34:26.029 --> 00:34:32.129
But that doesn't necessarily impose on an organism to do it.

00:34:32.209 --> 00:34:36.969
An organism can do the task without evidence accumulation. stimulation,

00:34:37.029 --> 00:34:39.389
it would simply have not good performance.

00:34:40.209 --> 00:34:46.089
But then what you showed, which was very surprising, is that this integration

00:34:46.089 --> 00:34:49.729
process of the information, the evidence provided to the animal,

00:34:49.849 --> 00:34:54.989
is following a summation rule and not so much an integration process, like a multiplication.

00:34:55.709 --> 00:35:02.549
And this had to do with an interaction effect between intensity and the duration of the stimulus.

00:35:02.829 --> 00:35:05.049
Yes. So how did that exactly play out?

00:35:06.049 --> 00:35:11.449
Well, the task involves comparison of two stimuli, and the two parameters that

00:35:11.449 --> 00:35:18.369
we can control are the intensity value, which was the variance in the vibration,

00:35:18.609 --> 00:35:23.149
the higher variance is perceived as being more intense, and the duration of the stimulus.

00:35:23.349 --> 00:35:28.489
Of course, there are many variables in the experiment that we didn't talk about today.

00:35:28.569 --> 00:35:32.529
For instance, the delay between the first and second stimulus is extremely interesting

00:35:32.529 --> 00:35:37.349
because a longer delay puts a heavier load on working memory.

00:35:37.569 --> 00:35:42.569
Today, instead, we focused on the parameters of intensity and duration.

00:35:42.909 --> 00:35:51.329
And our first question was whether performance in humans and rats or in rats

00:35:51.329 --> 00:35:54.709
would vary according to the duration of the stimulus.

00:35:54.909 --> 00:36:00.129
And as we were saying, the statistical structure is such that an ideal observer,

00:36:00.449 --> 00:36:07.109
a machine doing the task with all available data, would in fact do better with a longer stimulus.

00:36:07.589 --> 00:36:14.469
And we found that that is the case under the conditions in which the first and

00:36:14.469 --> 00:36:16.249
second stimulus were of equal duration.

00:36:18.369 --> 00:36:26.909
We then varied the duration, but included trials trials in which the first and

00:36:26.909 --> 00:36:30.909
second stimulus were not of equal duration, which we call the unbalanced condition.

00:36:31.589 --> 00:36:38.749
And in the unbalanced condition, we found, to our surprise, that the subjects,

00:36:38.969 --> 00:36:43.969
both humans and rats, did not measure intensity as an average over time,

00:36:44.149 --> 00:36:46.789
but instead with some form of summation,

00:36:47.049 --> 00:36:49.989
as we saw, not a linear summation,

00:36:50.269 --> 00:36:56.009
but with some form of summation that caused a longer stimulus to be perceived as stronger.

00:36:56.929 --> 00:37:01.169
So it was not purely an averaging, but some form of summation.

00:37:03.176 --> 00:37:06.956
But do you see this now as, let's say, I have to estimate, you could say I can

00:37:06.956 --> 00:37:09.936
estimate these two parameters of duration and intensity.

00:37:11.476 --> 00:37:15.416
But now if I would integrate over these parameters separately,

00:37:15.776 --> 00:37:20.076
then duration might give me, let's say, a false sense of intensity.

00:37:21.416 --> 00:37:25.056
Because imagine I just have, let's say, a clock and I'm just integrating from

00:37:25.056 --> 00:37:28.756
this clock longer, so my signal gets higher. and that might then start to have

00:37:28.756 --> 00:37:33.596
a crosstalk, adding noise, if you want, to my intensity estimate.

00:37:34.636 --> 00:37:38.696
So do you see it in those terms, like a crosstalk between informational channels?

00:37:39.536 --> 00:37:45.536
Or do you see it more as an intrinsic bias that the system has? It's an intrinsic bias.

00:37:47.136 --> 00:37:50.856
It's not that longer stimulus adds noise. In fact, the longer stimulus reduces

00:37:50.856 --> 00:37:57.656
noise in the sense that subjects perform better if the two stimuli to be compared

00:37:57.656 --> 00:38:00.176
last 600 milliseconds each,

00:38:00.416 --> 00:38:06.096
as opposed to if the two stimuli last 100 milliseconds each.

00:38:06.316 --> 00:38:11.196
So the longer stimulus doesn't add noise, it actually adds precision to the

00:38:11.196 --> 00:38:15.216
judgment of the statistical properties.

00:38:15.676 --> 00:38:22.076
The problem that I think the the perceptual system has, which it is not able

00:38:22.076 --> 00:38:26.296
to overcome, is that it doesn't have an absolute clock.

00:38:27.256 --> 00:38:32.216
If the brain had an absolute clock, then it could measure the total activity

00:38:32.216 --> 00:38:35.916
occurring over stimulus and then divide by this absolute value.

00:38:36.876 --> 00:38:43.756
But it's not available in the brain. And for some reason it's it's not there

00:38:43.756 --> 00:38:49.316
or it's not used and and so um and so not knowing exactly how much,

00:38:49.896 --> 00:38:53.236
to divide the the the accumulated value

00:38:53.236 --> 00:38:56.136
by there's a bias such that the longer stimulus

00:38:56.136 --> 00:38:59.556
feels stronger and this this effect you

00:38:59.556 --> 00:39:04.696
see as a linear effect so if i double the time then this subjective sense of

00:39:04.696 --> 00:39:08.896
intensity is also doubled or it follows an exponential curve as you should say

00:39:08.896 --> 00:39:13.236
grammatically or could you imagine that at the level of neural substrate it

00:39:13.236 --> 00:39:16.676
might be let's say more discretized or it might have different sensitivities

00:39:16.676 --> 00:39:19.356
because it has to write for instance on oscillatory activity,

00:39:20.296 --> 00:39:28.256
yeah we don't know in exactly that level of detail we built a model that simulates behavior.

00:39:29.536 --> 00:39:35.036
Correctly and accurately when the summation was done not linearly but exponentially

00:39:35.036 --> 00:39:40.856
with an exponentially decreasing weighting function so that the onset of the

00:39:40.856 --> 00:39:45.376
stimulus contributes to the perceived value with the highest weight,

00:39:45.476 --> 00:39:53.376
and then that weight decreases exponentially with a time constant of about 150 to 200 milliseconds.

00:39:53.976 --> 00:40:00.836
And then we compared that to the expected results if the,

00:40:02.015 --> 00:40:07.375
if the weight of the stimulus increased over time so that the end of the stimulus

00:40:07.375 --> 00:40:13.135
had more weight, and we found that the results were consistent with an exponentially

00:40:13.135 --> 00:40:16.695
decreasing weighting function, not increasing.

00:40:17.715 --> 00:40:23.675
Right. So in some sense, you put now the process that would account for this

00:40:23.675 --> 00:40:26.315
bias at really the sensor processing end of the pipeline.

00:40:26.755 --> 00:40:30.895
Well, in between, we still have our working memory buffer. We have another integration

00:40:30.895 --> 00:40:32.915
stage that has to come to the decision.

00:40:34.195 --> 00:40:39.035
So why do you put the full burden of that bias at the perceptual end?

00:40:39.195 --> 00:40:46.155
It might also be, let's say, some sort of capacity issue in your working memory, for instance, right?

00:40:46.235 --> 00:40:50.795
So we don't know where this waiting function occurs,

00:40:51.055 --> 00:40:55.535
but we've seen that if we look at the activity of sensory cortex,

00:40:55.735 --> 00:40:59.575
primary sensory cortex, which in rats is for the whisker area,

00:40:59.755 --> 00:41:01.995
this is called a barrel cortex,

00:41:02.275 --> 00:41:06.795
the activity of neurons there does not reflect a weighting function.

00:41:08.675 --> 00:41:16.975
So if we try to decode the choice of the rat based on the activity in the sensory

00:41:16.975 --> 00:41:23.015
cortex, the choices will not reflect the duration of the stimuli.

00:41:24.079 --> 00:41:29.119
So the effect of the duration must occur after sensory cortex.

00:41:29.459 --> 00:41:35.579
Another region that we're looking at is premotor cortex, which is in front of sensory cortex.

00:41:36.739 --> 00:41:44.739
And there we find many neurons that have activity during the time between the

00:41:44.739 --> 00:41:49.799
first and second stimulus, where the activity has a very clear relation to the first stimulus.

00:41:49.799 --> 00:41:53.919
So at first glance, one interprets these as working memory neurons.

00:41:54.139 --> 00:41:57.279
Many laboratories have found neurons like these.

00:41:57.399 --> 00:42:04.019
The ROMA laboratory in Mexico City has found many neurons like this in exploration of primate cortex.

00:42:04.659 --> 00:42:11.879
And the interpretation is that these are neurons involved in working memory.

00:42:12.139 --> 00:42:17.179
Now, when we look at how in rats the working memory neurons encode the first

00:42:17.179 --> 00:42:23.799
stimulus, we find that the firing rate has a stronger,

00:42:23.959 --> 00:42:31.039
better statistical correlation with the first stimulus if we consider the first

00:42:31.039 --> 00:42:38.399
stimulus as the perceived value of that stimulus rather than the actual physical

00:42:38.399 --> 00:42:39.399
property of the stimulus.

00:42:39.559 --> 00:42:44.919
In other words, we find that the activity of a neuron that varies according

00:42:44.919 --> 00:42:46.679
to first stimulus intensity.

00:42:47.899 --> 00:42:55.919
Will fire more if the first stimulus lasts longer, even if the intensity during

00:42:55.919 --> 00:42:57.719
that stimulus is constant.

00:42:57.939 --> 00:43:02.619
So as the stimulus continues over time.

00:43:03.969 --> 00:43:08.469
We know from a lot of behavioral work that as the stimulus continues over time,

00:43:08.669 --> 00:43:10.589
the perceived intensity increases.

00:43:11.689 --> 00:43:13.389
This came out from the psychometric curves.

00:43:14.529 --> 00:43:19.989
At the same time, we find that neurons in premotor cortex, if the firing is

00:43:19.989 --> 00:43:25.229
positively correlated with the stimulus intensity, then those same neurons fire

00:43:25.229 --> 00:43:26.989
more when the stimulus lasts longer.

00:43:27.209 --> 00:43:34.289
So the firing rate actually reflects the stimulus duration. so the firing rate

00:43:34.289 --> 00:43:39.109
in premotor cortex reflects the perceived value of the stimulus rather than

00:43:39.109 --> 00:43:40.589
the average value of the stimulus.

00:43:41.229 --> 00:43:50.349
Do you do a control where the intensity is not varying randomly within a given

00:43:50.349 --> 00:43:54.109
range but is fixed and so you get a continuous stimulus,

00:43:55.549 --> 00:43:57.909
at that right target frequency?

00:44:01.269 --> 00:44:08.029
So two stimuli are being compared and… So rather than getting a noisy stimulus

00:44:08.029 --> 00:44:09.689
with an average intensity,

00:44:09.989 --> 00:44:14.909
you would just have the stimulus continuously at the target value.

00:44:15.069 --> 00:44:16.509
That would be an easier task.

00:44:16.909 --> 00:44:18.849
Something like a sinusoid, for instance.

00:44:19.309 --> 00:44:24.989
No, we have not used sinusoidal stimuli in this stage of the experiment.

00:44:25.269 --> 00:44:33.829
We did in some very preliminary pilot studies and found that rats had a hard

00:44:33.829 --> 00:44:39.909
time maintaining attention for pulse trains and for sinusoids.

00:44:39.969 --> 00:44:46.289
That is, it was more difficult to convince the rat to remain in the nose poke,

00:44:46.469 --> 00:44:49.969
to remain immobile bowel with its whiskers in contact with the plate.

00:44:50.669 --> 00:44:53.829
For the first stimulus, the interstimulus delay, and the second stimulus,

00:44:53.849 --> 00:44:58.169
to wait for the go cue when the stimuli were sinusoidal.

00:44:58.269 --> 00:45:01.549
They tended to abort the trial, that is to leave early.

00:45:01.949 --> 00:45:03.949
So to put it in layman's terms.

00:45:06.984 --> 00:45:11.944
Periodic or regular or constant stimuli are simply not interesting.

00:45:12.284 --> 00:45:14.584
And for some reason, we're not sure why.

00:45:15.064 --> 00:45:21.244
And the noisy stimuli are interesting. The rat is very willing to wait for the entire stimulus.

00:45:21.384 --> 00:45:26.444
We can make the stimulus, if we want, even one second long, two seconds long, and they wait.

00:45:27.304 --> 00:45:31.904
That's interesting. So one thing that occurred to me is that you have a task

00:45:31.904 --> 00:45:36.644
in which the stimulus is varying along two parameters, intensity and duration.

00:45:36.984 --> 00:45:40.904
Um and but success in

00:45:40.904 --> 00:45:43.944
the task requires that you just attend to intensity

00:45:43.944 --> 00:45:46.804
and yes you could you ignore duration yes it's uh

00:45:46.804 --> 00:45:49.884
because it's is put there as a kind of confound

00:45:49.884 --> 00:45:52.824
yes um and your model

00:45:52.824 --> 00:45:59.224
is assuming that they are more or less successful in attending to the intensity

00:45:59.224 --> 00:46:04.384
but presumably i mean neither the rat nor the human knows that this rule is

00:46:04.384 --> 00:46:08.064
just about intensity So they might be working on all sorts of theories about

00:46:08.064 --> 00:46:10.284
mixtures of intensity and duration.

00:46:11.344 --> 00:46:18.684
And if you give either duration or intensity on their own, then they can solve those problems too.

00:46:18.944 --> 00:46:22.424
So I'm just wondering, in the model space that you've explored,

00:46:22.664 --> 00:46:29.024
have you explored all the potential variants of rules that the animals could

00:46:29.024 --> 00:46:32.924
be using to try and maximize performance on this task?

00:46:32.924 --> 00:46:44.844
Are there rules where you might do reasonably well by taking the duration into account in some way?

00:46:45.757 --> 00:46:54.017
Well, we put a reward rule into the computer, and that's applied to the rat

00:46:54.017 --> 00:46:55.777
or to the human during the experiment.

00:46:56.077 --> 00:47:01.957
And the reward rule, if we do an intensity experiment, the reward rule is intensity.

00:47:02.437 --> 00:47:08.257
That's what has to be compared. Or we can, in different routes or different human subjects,

00:47:08.497 --> 00:47:14.957
we can change the rule such that the subject is rewarded according to duration

00:47:14.957 --> 00:47:20.357
and to perform perfectly should ignore intensity.

00:47:20.357 --> 00:47:25.697
Now, if we consider the first case when the rule is intensity,

00:47:25.977 --> 00:47:32.137
there are many, many trials in which the two stimuli have the same duration

00:47:32.137 --> 00:47:34.777
and vary only by intensity.

00:47:35.457 --> 00:47:42.317
So an example would be a 400 millisecond stimulus followed by 400, or 600, 600, 200, 200.

00:47:42.757 --> 00:47:49.577
In all those cases, no choice could be made based on duration.

00:47:49.577 --> 00:47:54.677
In fact, the rats and the humans perform very well according to the intensity rule.

00:47:55.117 --> 00:48:01.497
So we have no reason to think that when they're actually doing the intensity

00:48:01.497 --> 00:48:06.897
task, that they think that they might be doing a time task.

00:48:07.157 --> 00:48:10.477
They know that when they're doing intensity, they're doing intensity.

00:48:11.497 --> 00:48:16.337
The problem that they have is that when they try to accomplish the intensity

00:48:16.337 --> 00:48:19.117
task, ask the sensory system.

00:48:20.371 --> 00:48:25.531
Is confounded by the duration. Well, that's your reading of what they're doing.

00:48:25.791 --> 00:48:30.531
I mean, is your assumption there that intensity is more salient to them than duration?

00:48:31.031 --> 00:48:36.591
No, because we can train rats and we can train humans to do duration.

00:48:36.971 --> 00:48:45.811
The question is the reward rule. In fact, there's actually no requirement that

00:48:45.811 --> 00:48:47.151
we use a different stimulus set.

00:48:47.151 --> 00:48:55.011
If we have two parameters to be varied, the intensity and the time duration,

00:48:55.351 --> 00:49:02.031
we can have stimulus pairs that differ in one or the other, and the subject

00:49:02.031 --> 00:49:05.771
is simply rewarded for making a choice based on one parameter and not the other,

00:49:05.871 --> 00:49:08.031
or the second parameter and not the first.

00:49:08.831 --> 00:49:13.271
So we don't even have to change the stimuli. We just change the rule and we

00:49:13.271 --> 00:49:18.531
find that a subject trained according to one rule makes its choice based on that rule,

00:49:18.711 --> 00:49:23.911
and a subject trained on the other rule makes its choice based on that rule.

00:49:25.191 --> 00:49:28.831
You still have some kind of interaction effect because with the intensity the

00:49:28.831 --> 00:49:31.591
longer stimuli are easier to classify.

00:49:32.771 --> 00:49:39.511
So this makes your task rather different from a lot of more standard judgment tasks of that kind.

00:49:40.391 --> 00:49:46.191
Right. So for the longer stimulus, there is more signal available for making

00:49:46.191 --> 00:49:47.611
the discrimination, and we found

00:49:47.611 --> 00:49:51.591
that in fact the discriminations are made better both by humans and rats.

00:49:51.791 --> 00:49:57.111
And this confirms work done by the ROMA laboratory with vibrations applied to

00:49:57.111 --> 00:50:05.451
the finger of monkeys, keys, and also in visual system, detecting the direction of noisy moving dots,

00:50:05.651 --> 00:50:11.731
dots which move in a coherent or incoherent way, performance is better for longer stimuli.

00:50:11.991 --> 00:50:16.911
So we've simply confirmed that for stimuli in which.

00:50:17.950 --> 00:50:22.170
More information, more signal is available in a longer time,

00:50:22.290 --> 00:50:23.790
even to the ideal observer.

00:50:24.150 --> 00:50:29.710
There's simply more information available that in fact rats and humans perform better.

00:50:29.910 --> 00:50:32.770
So that's a confirmation of many, many different studies.

00:50:33.290 --> 00:50:42.950
And the confound between time and intensity was something that emerged from our experiments.

00:50:43.150 --> 00:50:49.070
Then we went back in literature and looked for this and found that there have

00:50:49.070 --> 00:50:51.270
been reports consistent with this.

00:50:51.870 --> 00:50:57.870
And are you able to look at the sort of learning history and infer from that

00:50:57.870 --> 00:51:04.230
anything about perhaps the strategy, certainly in humans, you could imagine

00:51:04.230 --> 00:51:06.790
you have a strategy of testing various hypotheses.

00:51:07.210 --> 00:51:13.390
But even in the rat, you might imagine that there was some pattern of of changes

00:51:13.390 --> 00:51:18.210
in the trials where at one stage they have no idea and perhaps another stage

00:51:18.210 --> 00:51:20.890
they're testing that it's about intensity,

00:51:21.190 --> 00:51:23.770
these kinds of things. Is there anything to see from that?

00:51:24.110 --> 00:51:30.530
Yes. During training, rats, of course, have very low performance as they're

00:51:30.530 --> 00:51:33.350
learning, but they make a choice.

00:51:33.570 --> 00:51:39.270
And they might very well have some rule in mind,

00:51:39.310 --> 00:51:43.950
some algorithm, algorithm that they're by which they're acting and it's it's

00:51:43.950 --> 00:51:48.550
not working or it may be by chance works 60 or 65 percent of the time and they

00:51:48.550 --> 00:51:53.790
think that that's fine so so we have to simply.

00:51:55.110 --> 00:52:03.650
Simply train them day after day and if they if they had in mind a rule that

00:52:03.650 --> 00:52:06.970
was not the the one that the computer is set up to use,

00:52:07.130 --> 00:52:11.710
then sooner or later they'll discover that it's not reliably giving them the reward.

00:52:13.107 --> 00:52:18.747
So you made a model to explain the performance or the mapping of the stimulus

00:52:18.747 --> 00:52:19.827
properties to the performance.

00:52:20.207 --> 00:52:24.387
And in there, you see that you have this very fixed time constant at which you

00:52:24.387 --> 00:52:26.567
are ramping the sensory evidence that comes in.

00:52:26.647 --> 00:52:29.447
You just mentioned 150 milliseconds or something like that.

00:52:30.207 --> 00:52:35.127
Do you see this really as an invariant that is sort of wired into the human and the red brain?

00:52:35.347 --> 00:52:39.547
Or this is also in turn dependent on task properties?

00:52:40.967 --> 00:52:46.467
That's a really interesting question. And we've been able to,

00:52:46.547 --> 00:52:50.907
from the simulation, the model, we've been able to extract this time constant

00:52:50.907 --> 00:52:53.627
tau for a number of rats, a number of humans.

00:52:53.967 --> 00:53:00.927
And to our surprise, and I think probably it would be surprising to many colleagues,

00:53:01.187 --> 00:53:05.307
the tau, the time constant, was actually slightly longer in rats than in humans.

00:53:06.067 --> 00:53:12.827
About in the order of 150 to 175 milliseconds for rats and about 100 to 150

00:53:12.827 --> 00:53:14.207
milliseconds for humans.

00:53:15.087 --> 00:53:21.047
So while many of us would have predicted that humans accumulate evidence over

00:53:21.047 --> 00:53:25.407
a longer history, it turns out that rats accumulate evidence over a longer history.

00:53:25.587 --> 00:53:27.467
That was a big surprise to us.

00:53:30.467 --> 00:53:37.447
Under the same conditions. Now, the question is whether this tau is something intrinsic to the brain.

00:53:37.867 --> 00:53:42.367
Might it be the same for different kinds of stimuli, maybe even for different modalities?

00:53:42.907 --> 00:53:55.867
We don't know, but our guess is that tau does depend on the stimulus conditions.

00:53:57.527 --> 00:54:05.327
From a mathematical point of view if the stimulus is stochastic it has some,

00:54:06.567 --> 00:54:11.787
frequency properties what we can think of as a correlation time that is how

00:54:11.787 --> 00:54:16.387
much time has to pass for the stimulus to be uncorrelated with what it was in

00:54:16.387 --> 00:54:22.367
the past how much time has to pass before the stimulus becomes completely unpredictable,

00:54:23.427 --> 00:54:28.467
completely independent And in the conditions we've used, with a filter of about

00:54:28.467 --> 00:54:32.867
150 milliseconds, the correlation time is in the order of 10 milliseconds,

00:54:33.287 --> 00:54:35.747
plus or minus a few milliseconds.

00:54:36.987 --> 00:54:41.787
So that means that in 150 milliseconds,

00:54:42.287 --> 00:54:49.767
the rat could sample, the sensory system could sample approximately 10 to 20,

00:54:49.827 --> 00:54:52.847
roughly 10 to 20 independent samples.

00:54:54.094 --> 00:54:59.274
If the correlation time were longer, that is, if it took the stimulus longer

00:54:59.274 --> 00:55:04.434
to achieve a value unrelated to the previous value, for example,

00:55:04.454 --> 00:55:07.254
making the filter a lower pass,

00:55:08.014 --> 00:55:13.534
then it would take longer in order to accumulate the same number of independent samples.

00:55:14.094 --> 00:55:18.434
Therefore, to achieve the same performance, the sensory system would have to

00:55:18.434 --> 00:55:24.574
actually sample for longer. So one would expect that tau may adapt to that by

00:55:24.574 --> 00:55:27.954
increasing the integration time.

00:55:28.074 --> 00:55:30.994
This is a bit mathematically complex, but I hope it comes through.

00:55:31.294 --> 00:55:35.354
But then, so the point would be that either I have some sort of attenuation

00:55:35.354 --> 00:55:37.534
factor of the evidence I accumulate,

00:55:37.854 --> 00:55:44.454
or I might be sitting in some oscillatory dynamic that is basically dictating

00:55:44.454 --> 00:55:49.134
to me that, okay, the early samples will have a higher impact on the information

00:55:49.134 --> 00:55:50.394
to grade than later samples.

00:55:50.674 --> 00:55:55.374
So how do you see it more as a continuous attenuation factor or do you see it

00:55:55.374 --> 00:55:59.334
more as sort of an oscillatory sampling process.

00:56:00.694 --> 00:56:03.694
So until we have evidence to

00:56:03.694 --> 00:56:11.054
the contrary our guess is that this is a continuous a continuous a weighting

00:56:11.054 --> 00:56:15.174
function that changes continuously over time we haven't really looked for and

00:56:15.174 --> 00:56:20.914
therefore haven't seen any signs that it might be oscillatory maybe at a gamma

00:56:20.914 --> 00:56:22.254
frequency or something like that

00:56:22.474 --> 00:56:25.314
could have some role, but we simply haven't looked at that.

00:56:25.314 --> 00:56:29.194
But then it is a weighing function that in turn might be task-dependent.

00:56:29.194 --> 00:56:32.554
So that means for different tasks and different, let's say, stages of learning,

00:56:32.814 --> 00:56:36.434
the weighing function will change. Yeah, no, for sure it does.

00:56:36.674 --> 00:56:42.094
I mean, let me try to pull out an example that might make sense intuitively,

00:56:42.714 --> 00:56:46.574
although I'm not working through this mathematically,

00:56:46.574 --> 00:56:49.514
but if if i ask you is

00:56:49.514 --> 00:56:52.974
the climate of the earth changing you would

00:56:52.974 --> 00:56:58.094
collect samples you'd go back in history and use ice core or whatever might

00:56:58.094 --> 00:57:03.814
be available and collect samples uh over a long time to find out if if uh if

00:57:03.814 --> 00:57:10.414
the climate of the earth is changing and so so so the the the.

00:57:11.964 --> 00:57:15.504
Period of time over which you'd have to collect samples in order to give us

00:57:15.504 --> 00:57:18.844
an answer would be many hundreds or thousands of years.

00:57:19.184 --> 00:57:26.024
If I ask you, is something that changes in the order of seconds,

00:57:26.324 --> 00:57:29.084
you would need many fewer samples.

00:57:29.464 --> 00:57:43.144
So clearly, in any kind of task, in order to estimate the properties of some time series of data.

00:57:44.124 --> 00:57:48.584
You will need a different number of samples in order to make the estimate.

00:57:49.464 --> 00:57:56.264
So from the model, this now follows logically, you also predicted that if you

00:57:56.264 --> 00:57:58.384
would compare now a primacy or a recency effect,

00:57:58.564 --> 00:58:01.944
like more evidence in the beginning or in the end, And that actually for both

00:58:01.944 --> 00:58:05.564
the rats and the humans, you would see primacy. And that's also what you observe.

00:58:05.784 --> 00:58:08.944
Yes. So this was actually a really nice prediction that came out of the model.

00:58:09.364 --> 00:58:11.924
But then the next step is now that you have the model in your hands,

00:58:12.064 --> 00:58:14.604
okay, what are the neurons really doing? Okay.

00:58:14.944 --> 00:58:19.004
So with that, you went to a frontal area in the rat brain.

00:58:19.544 --> 00:58:23.744
And you start to look at the neural responses in this task. ask.

00:58:23.904 --> 00:58:29.964
So what were the features of the neural responses that stood out for you in

00:58:29.964 --> 00:58:31.944
these populations of neurons that you measured from?

00:58:32.444 --> 00:58:37.524
So the data that I presented today came from two regions of the cerebral cortex.

00:58:37.804 --> 00:58:43.224
The primary sensory cortex, which as we said in rats, it's called barrel cortex,

00:58:43.404 --> 00:58:45.424
the cortex that receives input from the whiskers.

00:58:45.584 --> 00:58:50.004
And then we looked at a more frontal region, which is usually called premotor

00:58:50.004 --> 00:58:56.084
cortex or whisker motor cortex or as it's also being called by Carlos Brody

00:58:56.084 --> 00:58:58.364
frontal orienting field FOF.

00:58:58.964 --> 00:59:04.404
So different people have different names for it and so I showed some evidence

00:59:04.404 --> 00:59:05.764
from from that area as well.

00:59:05.944 --> 00:59:08.464
In the sensory cortex we found that.

00:59:10.650 --> 00:59:17.490
Around half the neurons have a very clear relationship in their firing to the stimulus properties,

00:59:17.830 --> 00:59:25.590
but what the neurons report in their firing is only the most recent events of

00:59:25.590 --> 00:59:27.550
the stimulus, the last 10 or 20 milliseconds. seconds.

00:59:27.650 --> 00:59:29.930
So these are what we call local coding neurons.

00:59:30.150 --> 00:59:35.810
They encode what happened immediately, just a few milliseconds ago.

00:59:37.790 --> 00:59:43.810
What happened 100 milliseconds ago or 200 milliseconds ago has very little impact

00:59:43.810 --> 00:59:46.690
on the likelihood of a spike at any given time.

00:59:46.810 --> 00:59:56.390
So at time t equals zero, Euro, the stimulus value at T minus 250 has no impact,

00:59:56.570 --> 01:00:01.390
but what happened at T minus 10 or 15 milliseconds has a large impact.

01:00:01.650 --> 01:00:07.930
So the neurons report the most instantaneous, most recent value of the vibration.

01:00:08.390 --> 01:00:16.310
So this is an essential element for the brain to be able to reconstruct the

01:00:16.350 --> 01:00:18.590
stochastic stimulus, this noisy vibration.

01:00:18.990 --> 01:00:22.790
The noisy vibration is made up of a sequence of local events,

01:00:23.010 --> 01:00:26.890
but any given local event does not define the stimulus.

01:00:27.130 --> 01:00:32.410
And so there has to be integration done after the sensory cortex in order for

01:00:32.410 --> 01:00:39.650
the brain to appreciate the overall statistical structure of the neuron, of the stimulus, sorry.

01:00:39.790 --> 01:00:42.390
So the sensory cortex neurons are.

01:00:43.544 --> 01:00:47.544
Are local coders, and the integration has to occur elsewhere.

01:00:47.964 --> 01:00:54.124
In the frontal region that we refer to as premotor cortex, the activity of neurons

01:00:54.124 --> 01:00:59.524
did in fact reflect the overall statistical structure of the stimulus,

01:00:59.764 --> 01:01:02.764
but not the local history.

01:01:03.064 --> 01:01:07.484
So for instance, from looking at the firing of a neuron in prefrontal cortex,

01:01:07.744 --> 01:01:13.984
we cannot say what happened in the whisker vibration 10 milliseconds ago or

01:01:13.984 --> 01:01:19.824
15 milliseconds ago, but we can say what has happened over the course of the

01:01:19.824 --> 01:01:21.364
last 100 or 200 milliseconds.

01:01:21.744 --> 01:01:26.364
So those neurons already reflect integration, but we're not sure if they're

01:01:26.364 --> 01:01:33.344
doing the integration or are receiving a neuronal signal that's already processed.

01:01:33.724 --> 01:01:37.844
Right, but now, would you see these primary sensory neurons as wavelets,

01:01:38.524 --> 01:01:39.744
which is like neural wavelets?

01:01:40.444 --> 01:01:43.284
That are now encoding a time series.

01:01:43.544 --> 01:01:49.624
Yeah, that's one way to look at it. In fact, our colleague Rodrigo Quiroga is

01:01:49.624 --> 01:01:56.344
looking at spike trains through wavelet analysis, and it's actually quite a promising approach.

01:01:56.744 --> 01:02:03.044
Right. So now we have the prefrontal or the premotor area neurons looking at

01:02:03.044 --> 01:02:05.164
the signal at a longer time window. Yes.

01:02:05.644 --> 01:02:10.244
But can you correlate the neural response directly with the parametric control

01:02:10.244 --> 01:02:16.144
of your stimulus, like duration and the width of the vibration distribution?

01:02:16.664 --> 01:02:20.564
Yeah, the neurons in the premotor cortex,

01:02:20.884 --> 01:02:26.984
most of them, those that do have firing that's related to the stimulus,

01:02:27.204 --> 01:02:32.844
are better correlated with what we would call the perceived value of the stimulus

01:02:32.844 --> 01:02:35.664
than the physical value of the stimulus.

01:02:35.744 --> 01:02:40.904
That is, each stimulus is characterized by the intensity, which we call sigma, and by the duration.

01:02:42.275 --> 01:02:45.475
The neurons in premotor cortex,

01:02:45.835 --> 01:02:52.815
when they encode the memory of the stimulus or they encode the choice made by the animal,

01:02:53.055 --> 01:02:54.975
their activity is better related

01:02:54.975 --> 01:03:02.375
to the value of sigma after we take into account the effect of time.

01:03:05.935 --> 01:03:11.935
And that's one side of the PFC or the frontal neurons. And then there are others

01:03:11.935 --> 01:03:17.455
that encode the working memory element, or is that what you're talking about, the working memory?

01:03:17.615 --> 01:03:24.915
So we've seen in premotor cortex, as Renufo Romo, who is present today,

01:03:25.075 --> 01:03:30.055
he noted that he's seen the same thing in primate premotor cortex,

01:03:30.295 --> 01:03:34.835
that there's really a mixed soup of neurons there.

01:03:34.835 --> 01:03:38.075
It's incorrect in our data

01:03:38.075 --> 01:03:41.495
and his data it's incorrect to believe that one module

01:03:41.495 --> 01:03:44.695
one region of cortex is working in a

01:03:44.695 --> 01:03:48.595
homogeneous way one sees in the same trainings

01:03:48.595 --> 01:03:54.355
in the same test session even recorded adjacent electrodes in other words neurons

01:03:54.355 --> 01:04:00.935
sitting side by side do very different things in premotor cortex we see a mixture

01:04:00.935 --> 01:04:05.455
of neurons that include those that encode the stimulus as the stimulus occurs,

01:04:05.775 --> 01:04:12.335
but not the local history of the neuron, but rather the time-integrated value of the neuron.

01:04:12.675 --> 01:04:16.915
During the time-integrated value of the stimulus, of course.

01:04:18.155 --> 01:04:20.455
During the stimulus presentation.

01:04:20.595 --> 01:04:26.075
So those are online as the stimulus occurs, those neurons encode that stimulus.

01:04:26.635 --> 01:04:33.875
Other neurons encode the the preceding stimulus during the delay interval between the two stimuli.

01:04:34.015 --> 01:04:38.995
So those neurons seem to participate in working memory, and then still other neurons.

01:04:40.358 --> 01:04:44.338
That encode during the second stimulus the value of that stimulus,

01:04:44.498 --> 01:04:49.678
the parameter of that stimulus, and others the choice that the rat has to make

01:04:49.678 --> 01:04:53.198
based on the comparison between the first and the second stimulus.

01:04:53.598 --> 01:05:00.278
And to make matters even more complicated, some neurons encode combinations

01:05:00.278 --> 01:05:02.278
of these different features.

01:05:02.398 --> 01:05:06.718
That is, it's not unusual for a neuron to encode the stimulus as as it occurs,

01:05:06.978 --> 01:05:11.318
and the memory of the stimulus, whereas another neuron may encode the memory

01:05:11.318 --> 01:05:13.438
but not the stimulus as it occurs.

01:05:13.958 --> 01:05:19.338
So there's really no single label that you can give to the full population of neurons.

01:05:19.638 --> 01:05:24.258
So you can see why there might need to be different neuron types in order to

01:05:24.258 --> 01:05:28.258
solve this problem, but if there's no topography, as you suggest,

01:05:28.518 --> 01:05:33.238
then we have a real challenge for reading out what is the right answer here.

01:05:33.238 --> 01:05:38.538
So it's not just a case of looking at average firing in say the comparison neurons

01:05:38.538 --> 01:05:43.518
because they're intermixed with the working memory neurons and so on so what

01:05:43.518 --> 01:05:48.578
do you think is have you got any idea of how that readout might happen?

01:05:48.958 --> 01:05:56.538
Well I think that there's two issues I would say two ways of looking at this

01:05:56.538 --> 01:06:02.598
problem one is the computation that's being done physiologically by the cortex

01:06:02.598 --> 01:06:04.478
cortex, and the other is the readout.

01:06:04.698 --> 01:06:11.098
So about the computation, I would say we're close to the starting point.

01:06:11.218 --> 01:06:15.698
We just don't know what it is. We don't know, for instance, how a memory is

01:06:15.698 --> 01:06:17.938
stored for one or two seconds.

01:06:18.158 --> 01:06:23.238
We just don't know. We don't know whether firing rate is, which is what we've

01:06:23.238 --> 01:06:26.938
looked at and most laboratories look at, we don't know if firing rate is actually

01:06:26.938 --> 01:06:30.478
the memory or if there's something beyond firing rate.

01:06:30.538 --> 01:06:34.718
There could be population codes, there could be trajectories through multi-dimensional

01:06:34.718 --> 01:06:39.698
population space, there could be latent synaptic weights during the delay interval.

01:06:39.838 --> 01:06:44.998
There are many ways, many theories for how memory is stored,

01:06:45.298 --> 01:06:49.118
and we really can't confirm or exclude any of these.

01:06:49.798 --> 01:06:55.398
So about this we really don't know much, and we don't know the transformation

01:06:55.398 --> 01:07:00.198
from the feeling of an ongoing stimulus to a memory of the the feeling.

01:07:00.518 --> 01:07:06.618
This is completely unknown to us. The second point that you raise, decoding.

01:07:07.458 --> 01:07:15.138
Well, we can quite easily make networks that take our activity and decode it,

01:07:15.238 --> 01:07:18.418
and that's not difficult to do.

01:07:19.398 --> 01:07:22.578
From the same population of neurons, we can,

01:07:23.785 --> 01:07:28.845
We can look at the population from, say, two different angles,

01:07:29.565 --> 01:07:34.065
and by one angle we see a memory, and from another angle we see a choice.

01:07:34.605 --> 01:07:39.085
This has been shown, for instance, in a paper recently.

01:07:43.245 --> 01:07:48.245
Valerio Monte, I think, if I'm not mistaken, is the first author of the paper,

01:07:48.245 --> 01:07:52.845
from neurons in primate prefrontal cortex,

01:07:53.125 --> 01:07:58.585
where there were two features of a stimulus, I believe color and motion.

01:07:59.885 --> 01:08:03.865
And the monkeys were trained to do either a color task or a motion task.

01:08:04.265 --> 01:08:12.205
And he found that the same neurons encoded both features and could be decoded

01:08:12.205 --> 01:08:16.925
from the the activity of those neurons, either the color or the motion could be decoded.

01:08:17.045 --> 01:08:25.205
So it's not that difficult using sort of offspring of principal component kinds

01:08:25.205 --> 01:08:31.265
of analyses to find dimensions whereby a certain property can be decoded.

01:08:31.765 --> 01:08:37.345
We can do that, but what we don't know is what makes the neurons fire that way.

01:08:37.665 --> 01:08:43.225
So from the population response, you can decompose it into different components,

01:08:43.325 --> 01:08:49.745
one of which gives you a good indication of the animal's behavior, its actual response.

01:08:50.125 --> 01:08:56.885
Yes. And so we're assuming some downstream system is able to do that decomposition. Yes.

01:08:57.585 --> 01:09:02.125
But now the standard model of decision-making are drift-diffusion models where

01:09:02.125 --> 01:09:03.685
we just integrate over firing rates.

01:09:04.105 --> 01:09:07.685
And in some sense, if you look at your model or your physiology,

01:09:07.805 --> 01:09:13.585
your data, then you also see some very marked or task-specific modulation of firing rates.

01:09:14.385 --> 01:09:20.165
So is the minimum model that we could apply to this then exactly that,

01:09:20.185 --> 01:09:21.005
the drift diffusion model?

01:09:22.345 --> 01:09:27.905
Yeah, this has something in common with drift diffusion and it's the kind of

01:09:27.905 --> 01:09:32.565
stimulus statistically that has been explored for drift diffusion.

01:09:33.065 --> 01:09:40.925
I think that one of the main differences is that the behavioral arrangement

01:09:40.925 --> 01:09:47.885
in this task does not ask the rat or the human observer to reach some threshold and make a choice.

01:09:48.485 --> 01:09:55.445
We don't give them the option of collecting as much evidence as they want and then taking the action.

01:09:57.742 --> 01:10:06.422
In those cases, it's been argued that the animal or the human collects evidence

01:10:06.422 --> 01:10:11.742
until they reach, and so they drift to a threshold, and then they make the choice.

01:10:11.822 --> 01:10:17.882
In our case, we determine the stimulus duration, and the animal and the rat

01:10:17.882 --> 01:10:21.702
and the human are required to wait for the entire stimulus.

01:10:21.882 --> 01:10:28.002
So that may take them beyond the threshold or they may not be allowed to reach the threshold.

01:10:28.822 --> 01:10:32.642
That's determined by the stimulus duration. Okay, but that would still mean

01:10:32.642 --> 01:10:37.722
that the decision variable could also be reflected just in the firing rate and

01:10:37.722 --> 01:10:39.562
then it's a matter of reading out that firing rate.

01:10:39.782 --> 01:10:46.682
It's not necessarily my favorite model, but you have market correlation if you

01:10:46.682 --> 01:10:50.742
want between the performance and your firing of these neurons.

01:10:50.742 --> 01:10:56.922
Of course, these are good examples, but in all the visual that you have seen,

01:10:57.162 --> 01:11:01.722
in the end, if we're close to the decision moment, you can actually predict

01:11:01.722 --> 01:11:05.062
from the firing rate of the neurons whether we go left or right.

01:11:05.062 --> 01:11:09.842
Yeah, and we've also looked at error trials.

01:11:10.482 --> 01:11:20.702
And on error trials, those neurons which have on correct trials a correlate

01:11:20.702 --> 01:11:25.742
in their neuronal firing rate to the choice of the animal have that same correlate on error trials.

01:11:25.742 --> 01:11:30.622
In other words, if a neuron fires at a high rate when the rat is going to turn

01:11:30.622 --> 01:11:39.102
right on correct trials and a low rate when he turns left on correct trials, then on error trials.

01:11:39.742 --> 01:11:45.062
The neuron will fire for turning right and not fire for turning left even though it's an error.

01:11:45.282 --> 01:11:52.562
So many of the premotor neurons that are choice selective are choice selective

01:11:52.562 --> 01:11:54.702
whether the choice is correct or incorrect. Correct.

01:11:54.822 --> 01:11:57.682
Right. So this is really amazing, right?

01:11:57.742 --> 01:12:02.542
Because you're really now going from a signal-dependent response in the nervous

01:12:02.542 --> 01:12:06.402
system to, let's say, a subjective state-dependent response.

01:12:07.562 --> 01:12:11.702
And with that, actually, you're really close to explaining this task.

01:12:12.522 --> 01:12:16.662
And then at the end of your talk, you also showed how actually it can have some

01:12:16.662 --> 01:12:21.222
diagnostic value if we look at humans, Which was a very surprising result where

01:12:21.222 --> 01:12:26.642
you showed that, well, actually, in performing another variation of this task,

01:12:26.822 --> 01:12:33.282
humans that are at risk of schizophrenia actually show a very different kind of performance.

01:12:33.662 --> 01:12:38.342
Yes. So what is the salient feature there in your mind? Right.

01:12:39.486 --> 01:12:45.566
So we have just a few subjects who are completely healthy,

01:12:45.726 --> 01:12:50.146
yet the family history suggests that they're in the category of what's called

01:12:50.146 --> 01:12:55.366
at-risk for schizophrenia, meaning that there's something in the family genome

01:12:55.366 --> 01:12:58.026
that may give a predisposition.

01:12:58.026 --> 01:13:03.526
Yet at the time of testing, they're absolutely healthy and symptom-free.

01:13:04.406 --> 01:13:15.766
So the task that we gave to the low-risk subjects and the healthy at-risk subjects

01:13:15.766 --> 01:13:18.166
was a time-duration comparison.

01:13:18.286 --> 01:13:25.226
So they received two vibrations sequentially and had to judge whether the first

01:13:25.226 --> 01:13:27.226
or second duration was greater.

01:13:28.026 --> 01:13:32.966
And the confounding factor was the intensity.

01:13:34.026 --> 01:13:39.386
And so this is motivated by the fact that the previous work had found that judgment

01:13:39.386 --> 01:13:41.606
of intensity was confounded by duration.

01:13:41.806 --> 01:13:45.766
We wondered whether perception of duration is confounded by intensity.

01:13:46.166 --> 01:13:52.886
And we found that in both groups of subjects, in fact, the perceived duration

01:13:52.886 --> 01:13:54.646
was confounded by intensity.

01:13:56.185 --> 01:14:02.345
In the sense that the stronger stimulus,

01:14:02.465 --> 01:14:07.225
the higher variance, the higher intensity stimulus was on average perceived

01:14:07.225 --> 01:14:11.365
as with a bias towards a longer duration.

01:14:12.205 --> 01:14:17.545
So to put it in very short and in rhyming words, longer feels stronger.

01:14:18.565 --> 01:14:20.385
Shorter duration feels weaker.

01:14:21.465 --> 01:14:23.965
And that this is the interpretation of the psychometric curves.

01:14:23.965 --> 01:14:33.645
So subjects were not able to purely exclude the intensity of the stimulus from the duration,

01:14:34.525 --> 01:14:38.185
even though the task would require that for perfect performance.

01:14:38.545 --> 01:14:43.825
In the subjects that were healthy but might be at risk due to family history.

01:14:45.105 --> 01:14:48.205
The effect of intensity was much more pronounced.

01:14:48.205 --> 01:14:56.065
In other words, they were less able to exclude the irrelevant stimulus feature from the computation.

01:14:56.505 --> 01:15:04.165
To do the task perfectly, the brain should discard the intensity information,

01:15:04.685 --> 01:15:09.325
not use it in the computation, and do the computation based purely on time.

01:15:09.325 --> 01:15:14.585
That was what the rule was in the task, is measure the time.

01:15:15.425 --> 01:15:19.805
And these subjects showed a much stronger effective intensity.

01:15:20.125 --> 01:15:25.345
So essentially, they were less able to discard the irrelevant feature.

01:15:27.205 --> 01:15:32.925
So we made a long tour now from the whisker system into the clinic.

01:15:32.985 --> 01:15:38.465
And you've been sort of getting us through that every step of the way.

01:15:39.385 --> 01:15:44.445
And so in that sense, you have been really championing this whole more system

01:15:44.445 --> 01:15:47.165
level perspective on the brain and how it relates to behavior.

01:15:47.625 --> 01:15:50.805
So if you would like to follow in that tradition that you represent,

01:15:51.065 --> 01:15:53.745
what should be Matthew's law that we have to adhere to?

01:15:55.024 --> 01:15:58.424
So I don't have any laws.

01:15:58.604 --> 01:16:04.744
I don't have any wisdom to pass on to the younger generation.

01:16:05.524 --> 01:16:10.884
Or even to us, you know? Even to us, I have nothing useful to say.

01:16:10.984 --> 01:16:16.604
The only thing I could add is that the approach that we use in the laboratory,

01:16:16.884 --> 01:16:21.484
and the approach, nothing works unless you have very, very good students.

01:16:21.484 --> 01:16:26.224
I have excellent students, very bright and motivated and hardworking,

01:16:26.424 --> 01:16:31.504
and that's what allows our approach to work.

01:16:31.684 --> 01:16:35.644
And the approach I outlined at the very beginning of the talk,

01:16:35.744 --> 01:16:40.044
which is we construct experiments that have three elements,

01:16:40.364 --> 01:16:46.524
a controlled sensory stimulus, a behavioral output, and neuronal activity,

01:16:46.744 --> 01:16:47.964
measurements of neuronal activity.

01:16:47.964 --> 01:16:53.384
And these three elements we can view as being a triangle and that gives us three

01:16:53.384 --> 01:16:57.584
possible connections the three sides of the triangle the connection between

01:16:57.584 --> 01:16:59.724
the stimulus and the behavior.

01:17:01.264 --> 01:17:06.744
Gives us some insight into how the brain uh how the brain experiences a sensory

01:17:06.744 --> 01:17:12.184
stimulus and this is quantified by psychophysics and we we try to study how

01:17:12.184 --> 01:17:20.244
a sensory stimulus produces a a a neuronal firing pattern within the sensory system,

01:17:20.344 --> 01:17:23.464
which we can call coding or sensory coding or encoding,

01:17:24.684 --> 01:17:30.344
whereby the sensory receptors and then the successive stages of processing put

01:17:30.344 --> 01:17:35.044
the physical properties of stimulus into the language of the brain action potentials.

01:17:35.284 --> 01:17:37.504
Then we want to look at,

01:17:38.298 --> 01:17:43.918
how a sensory representation can lead to a decision, which we can refer to as

01:17:43.918 --> 01:17:47.078
decision-making or decoding or any number of things.

01:17:47.558 --> 01:17:54.458
So I think that if you have the right members of your research team,

01:17:55.358 --> 01:17:59.718
then putting these three elements into an experiment can provide some insights.

01:18:00.198 --> 01:18:04.618
Now, we want to put this law on a T-shirt, and we don't want to make the print too small. Yeah.

01:18:04.698 --> 01:18:08.958
So what kind of, what's Matthew's law we can actually print on a t-shirt now?

01:18:10.458 --> 01:18:14.718
I think the law is, again, I don't want to call it a law.

01:18:14.778 --> 01:18:20.118
I'd say something that we have fun doing. Let's call it what we have fun doing instead of the law.

01:18:21.078 --> 01:18:29.578
It's Matthew's triangle. It's the golden triangle, which I would say is a systems approach.

01:18:29.818 --> 01:18:36.658
I'd say the approach is that it's more interesting when you actually see the

01:18:36.658 --> 01:18:37.738
organism doing something.

01:18:38.058 --> 01:18:44.258
So it's important to have behavior, it's important to know the input,

01:18:44.498 --> 01:18:47.838
and it's important to know the neuronal basis of this.

01:18:47.838 --> 01:18:54.298
So to put it in a word, I'd say the golden triangle of cognitive neuroscience.

01:18:54.758 --> 01:18:57.378
Wouldn't it be better to find a fourth point and call it a diamond?

01:18:59.538 --> 01:19:00.218
Matthew's diamond!

01:19:03.218 --> 01:19:07.758
So I can't be the one that does that. We'll do it for you. Don't you worry.

01:19:08.058 --> 01:19:09.978
But the last thing is, look, Tony likes traveling.

01:19:10.698 --> 01:19:14.038
And so he wants to definitely come and visit you in Trieste five years from

01:19:14.038 --> 01:19:16.438
now because he's not so quick booking these kinds of things.

01:19:16.438 --> 01:19:23.658
So, but five years from now it's going to be at your lab to test whether a prediction

01:19:23.658 --> 01:19:25.958
you made today was actually confirmed or rejected.

01:19:26.238 --> 01:19:30.218
So what's the key hypothesis that you want to see tested in that timeframe?

01:19:31.951 --> 01:19:36.731
Well, goodness gracious, that's a tough one. But I think that what I think would

01:19:36.731 --> 01:19:43.611
be fun, really fun to be able to do within five years is to have rats.

01:19:43.811 --> 01:19:49.011
We know we could do it in humans. It's not difficult, but to have our animals

01:19:49.011 --> 01:19:56.091
able to apply two different decision rules to sensory inputs.

01:19:56.451 --> 01:20:03.131
So in the work that I talked about today, we have a comparison task where the

01:20:03.131 --> 01:20:06.011
stimulus is characterized by intensity and duration,

01:20:06.231 --> 01:20:08.071
and we can train rats to do an

01:20:08.071 --> 01:20:12.931
intensity comparison, or other rats we can train to do a time comparison.

01:20:13.291 --> 01:20:19.011
The stimulus set is the same, so wouldn't it be fun and interesting to be able

01:20:19.011 --> 01:20:25.191
to train rats in one session or one group of trials to receive the stimuli and make one judgment,

01:20:25.311 --> 01:20:28.471
and then in the next group of trials to make the other judgment,

01:20:28.591 --> 01:20:38.651
and to see where in the brain the processing diverges according to what the animal is doing with it.

01:20:38.751 --> 01:20:42.371
The sensory input would be the same in both groups of trials,

01:20:42.531 --> 01:20:48.871
but what the brain has to extract to make its choice and its action is different,

01:20:48.991 --> 01:20:56.171
and it would it would be intriguing to see where the task rule takes its effect

01:20:56.171 --> 01:20:57.591
on neuronal processing.

01:20:58.031 --> 01:21:01.591
Right. Great. Matthew Diamond, thank you very much for this conversation. Thank you.

01:21:04.051 --> 01:21:09.731
The CSN Podcast was produced by the Convergent Science Network of Biometrics

01:21:09.731 --> 01:21:16.151
and Biohybrid Systems, a project funded by the European 7th Research Framework Program.

01:21:17.631 --> 01:21:22.991
For more interviews, recorded lectures, or upcoming conferences in the field

01:21:22.991 --> 01:21:29.271
of biometrics and bio-hybrid systems, go to csnnetwork.eu.

01:21:30.320 --> 01:21:38.160
Music.