WEBVTT 00:00:00.000 --> 00:00:02.640 Welcome to the deep dive if you're listening 00:00:02.640 --> 00:00:06.179 to this, you're probably someone who Well, you 00:00:06.179 --> 00:00:08.599 love grasping complex topics, right? Do you want 00:00:08.599 --> 00:00:10.640 to understand the world quickly and thoroughly 00:00:10.640 --> 00:00:13.240 but without drowning in all that horrible information 00:00:13.240 --> 00:00:15.699 overload, right? Nobody has time to read every 00:00:15.699 --> 00:00:18.839 single textbook exactly and that's exactly why 00:00:18.839 --> 00:00:21.519 we are pulling from a really comprehensive source 00:00:21.519 --> 00:00:24.100 today We're looking at excerpts from the Wikipedia 00:00:24.100 --> 00:00:26.839 article on meta learning in computer science, 00:00:26.940 --> 00:00:30.660 which I know sounds heavy but it's actually incredibly 00:00:30.660 --> 00:00:33.539 relevant to how we all learn. It really is. The 00:00:33.539 --> 00:00:35.899 mission for today is to explore how artificial 00:00:35.899 --> 00:00:38.659 intelligence is shifting from simply learning 00:00:38.659 --> 00:00:41.719 facts, you know, memorizing things, to actually 00:00:41.719 --> 00:00:44.020 learning how to learn. Right, which is a massive 00:00:44.020 --> 00:00:46.320 paradigm shift. Okay, let's unpack this. Because 00:00:46.320 --> 00:00:47.979 I want you to imagine for a second that you have 00:00:47.979 --> 00:00:51.340 an AI, right? And this system can completely 00:00:51.340 --> 00:00:53.740 crush the world's greatest grandmasters at chess. 00:00:53.820 --> 00:00:56.619 Oh, easily. processing millions of moves a second. 00:00:56.960 --> 00:00:59.200 Yeah. It's by all accounts a certified genius. 00:00:59.719 --> 00:01:02.259 But then you take that exact same super smart 00:01:02.259 --> 00:01:05.299 AI and you ask it to play a simple game of checkers, 00:01:05.659 --> 00:01:08.140 what happens? It completely breaks down. It has 00:01:08.140 --> 00:01:10.420 absolutely no idea what to do. It essentially 00:01:10.420 --> 00:01:13.340 has to be rebuilt, entirely retrained from scratch, 00:01:13.700 --> 00:01:17.000 just to understand how to move a piece diagonally. 00:01:17.120 --> 00:01:19.680 Yeah, and that is, I mean, it's the great paradox 00:01:19.680 --> 00:01:22.219 of modern computer science, really. Why is the 00:01:22.219 --> 00:01:25.099 smartest machine on Earth so incredibly rigid? 00:01:26.019 --> 00:01:28.519 Because we have built these monolithic systems 00:01:28.519 --> 00:01:31.939 that are unbelievably good at one specific thing, 00:01:32.180 --> 00:01:34.879 but they are... while they're incredibly brittle, 00:01:35.319 --> 00:01:38.659 they lack the fundamental ability to adapt. We 00:01:38.659 --> 00:01:40.799 spend massive amounts of time feeding machines 00:01:40.799 --> 00:01:43.260 data just hoping they learn a single task, when 00:01:43.260 --> 00:01:45.379 what we really need is a machine that understands 00:01:45.379 --> 00:01:47.519 the underlying mechanics of learning itself. 00:01:47.920 --> 00:01:50.060 Which brings us to the core definition from our 00:01:50.060 --> 00:01:53.140 source today. Meta -learning is defined as a 00:01:53.140 --> 00:01:55.340 subfield of machine learning. where automatic 00:01:55.340 --> 00:01:57.420 learning algorithms are applied to the metadata 00:01:57.420 --> 00:01:59.819 of machine learning experiments. Right, the metadata. 00:02:00.040 --> 00:02:01.799 But let's put it in the computer science dictionary 00:02:01.799 --> 00:02:03.780 for a second, because what does interacting with 00:02:03.780 --> 00:02:06.780 metadata actually look like in practice? Well... 00:02:06.560 --> 00:02:08.680 To understand that, you kind of have to look 00:02:08.680 --> 00:02:11.219 at what a standard machine learning model ignores. 00:02:11.919 --> 00:02:14.740 Usually, when you train an AI, you feed it thousands 00:02:14.740 --> 00:02:17.680 of images of cats, right? It makes guesses, and 00:02:17.680 --> 00:02:19.979 it gets a score based on its accuracy. So just 00:02:19.979 --> 00:02:22.419 trial and error. Exactly. And the engineers, 00:02:22.479 --> 00:02:24.860 they only care about the final model that can 00:02:24.860 --> 00:02:27.460 spot a cat. The entire journey to get there, 00:02:27.460 --> 00:02:30.840 all the mistakes, is just discarded. But meta 00:02:30.840 --> 00:02:33.460 -learning flips that entirely. It cares immensely 00:02:33.460 --> 00:02:35.199 about the journey. It looks at the metadata. 00:02:35.419 --> 00:02:38.419 Meaning the actual process itself? Yes. It analyzes 00:02:38.419 --> 00:02:40.780 the properties of the problem. It looks at how 00:02:40.780 --> 00:02:43.199 the algorithm's performance fluctuated over time, 00:02:43.259 --> 00:02:45.159 the patterns it tried to use, all of it. And 00:02:45.159 --> 00:02:46.759 what does it do with that? What's fascinating 00:02:46.759 --> 00:02:49.879 here is that it takes all of that context and 00:02:49.879 --> 00:02:53.300 uses it to dynamically alter, select, or even 00:02:53.300 --> 00:02:55.580 combine entirely different learning algorithms 00:02:55.580 --> 00:02:58.500 to solve a new problem. I think the natural reaction 00:02:58.500 --> 00:03:03.620 for someone outside the field is to wonder why 00:03:03.620 --> 00:03:06.180 that extra layer of complexity is even necessary. 00:03:06.740 --> 00:03:08.800 Because if you have a massive neural network 00:03:08.800 --> 00:03:11.560 and just a ton of computing power, shouldn't 00:03:11.560 --> 00:03:13.520 it just naturally be able to learn whatever you 00:03:13.520 --> 00:03:15.610 throw at it? You would think so, right. But it 00:03:15.610 --> 00:03:18.550 hits a wall due to a fundamental limitation. 00:03:18.750 --> 00:03:21.750 It's called inductive bias. Inductive bias. Right. 00:03:22.069 --> 00:03:25.069 Every single standard learning algorithm operates 00:03:25.069 --> 00:03:28.550 on an inductive bias. It's simply a set of built 00:03:28.550 --> 00:03:30.750 -in hard -coded assumptions that it makes about 00:03:30.750 --> 00:03:32.689 the data it's going to process. Like mathematical 00:03:32.689 --> 00:03:35.370 blinders. Exactly. Because it has those blinders 00:03:35.370 --> 00:03:38.129 on, a standard algorithm is only going to perform 00:03:38.129 --> 00:03:41.330 well if its specific assumptions happen to perfectly 00:03:41.330 --> 00:03:43.669 align with the new problem you give it. So it's 00:03:43.669 --> 00:03:46.569 highly specialized. but trapped by its own design. 00:03:47.129 --> 00:03:49.370 Yeah. I mean, think of inductive bias like a 00:03:49.370 --> 00:03:52.509 world -class, ultra -expensive sushi knife. Oh, 00:03:52.629 --> 00:03:54.729 that's a great analogy. Right. If your goal is 00:03:54.729 --> 00:03:57.069 slicing delicate fish, its design assumptions 00:03:57.069 --> 00:03:59.669 are perfect. The blade is exactly what you need. 00:04:00.009 --> 00:04:03.250 But if you take that exact same tool into a forest 00:04:03.250 --> 00:04:06.430 to chop down an oak tree... It's utterly useless. 00:04:06.830 --> 00:04:08.830 The assumptions built into the knife just do 00:04:08.830 --> 00:04:11.169 not match the reality of the wood. Precisely. 00:04:11.449 --> 00:04:14.509 A standard algorithm might dominate text translation, 00:04:14.830 --> 00:04:17.490 but completely fail at image recognition because 00:04:17.490 --> 00:04:19.589 the relationship between the data structure and 00:04:19.589 --> 00:04:22.529 the algorithm is mismatched. So meta -learning 00:04:22.529 --> 00:04:25.370 is sort of like a master craftsman. Okay, how 00:04:25.370 --> 00:04:27.949 so? Instead of just wildly swinging a sushi knife 00:04:27.949 --> 00:04:32.189 at a tree, the AI steps back, analyzes the properties 00:04:32.189 --> 00:04:35.430 of the task, the metadata, and dynamically decides, 00:04:35.670 --> 00:04:38.420 oh wait, this is wood. Right. I need to put away 00:04:38.420 --> 00:04:40.899 the knife, generate a chainsaw, and adjust my 00:04:40.899 --> 00:04:45.019 grip. It shifts its own inductive bias to match 00:04:45.019 --> 00:04:47.079 the reality of the problem in front of it. That's 00:04:47.079 --> 00:04:49.339 exactly it. It essentially rewrites its own rules 00:04:49.339 --> 00:04:51.819 for solving the problem. But practically speaking, 00:04:52.160 --> 00:04:54.699 how did scientists even begin to conceptualize 00:04:54.699 --> 00:04:57.060 code that could rewrite its own rules? Because, 00:04:57.060 --> 00:04:59.439 I mean, writing software that writes its own 00:04:59.439 --> 00:05:01.920 software sounds like a recipe for a total system 00:05:01.920 --> 00:05:04.629 crash. It does, yeah. But they didn't invent 00:05:04.629 --> 00:05:06.850 the concept from scratch. They actually looked 00:05:06.850 --> 00:05:09.430 at the ultimate most robust learning system that 00:05:09.430 --> 00:05:11.829 already exists. Which is? Biological evolution. 00:05:12.410 --> 00:05:15.050 Yeah, early pioneering work in the late 80s and 00:05:15.050 --> 00:05:18.250 early 90s, specifically by researchers like Jürgen 00:05:18.250 --> 00:05:22.259 Schmidhuber. and Yoshua Bengio, they used genetic 00:05:22.259 --> 00:05:24.920 evolution as the blueprint. OK, that makes sense. 00:05:25.120 --> 00:05:27.939 In nature, genetic evolution is the ultimate 00:05:27.939 --> 00:05:30.600 meta learner. It learns the actual learning procedure 00:05:30.600 --> 00:05:34.100 itself, encodes that procedure into DNA, and 00:05:34.100 --> 00:05:36.220 then executes it in the brain of an organism. 00:05:36.639 --> 00:05:39.399 It's an open -ended hierarchy of meta evolution. 00:05:39.639 --> 00:05:41.420 Wait, hold on. I have to push back on this a 00:05:41.420 --> 00:05:43.550 little bit. Go for it. I understand the inspiration, 00:05:44.129 --> 00:05:46.829 but biological evolution takes millions of years 00:05:46.829 --> 00:05:50.209 of horrific trial and error. I mean, entire species 00:05:50.209 --> 00:05:52.649 dying off just to find the right bias for survival. 00:05:52.800 --> 00:05:55.560 Right, it's not exactly efficient. Yeah. So if 00:05:55.560 --> 00:05:58.420 the goal of AI is to be fast and flexible, how 00:05:58.420 --> 00:06:00.660 on earth does it dynamically choose its bias 00:06:00.660 --> 00:06:03.079 in real time without running a million -year 00:06:03.079 --> 00:06:05.279 simulation? Like, what levers is it actually 00:06:05.279 --> 00:06:07.959 pulling to speed this up? That is the exact problem 00:06:07.959 --> 00:06:10.480 researchers had to solve. It requires a crucial 00:06:10.480 --> 00:06:12.959 distinction. When a meta -learning system shifts 00:06:12.959 --> 00:06:16.000 its bias, it isn't blindly mutating like early 00:06:16.000 --> 00:06:18.680 life on Earth. It is meticulously adjusting two 00:06:18.680 --> 00:06:21.620 very specific parameters. And we should probably 00:06:21.620 --> 00:06:24.050 clarify for anyone who has taken a basic data 00:06:24.050 --> 00:06:26.709 science class. This is completely different from 00:06:26.709 --> 00:06:29.209 the standard bias variance dilemma you learn 00:06:29.209 --> 00:06:31.509 about in statistics. Right. This is a different 00:06:31.509 --> 00:06:35.110 kind of bias. Exactly. Meta -learning is specifically 00:06:35.110 --> 00:06:38.170 tweaking what we call declarative bias and procedural 00:06:38.170 --> 00:06:40.290 bias. Okay. Let's unpack those because that sounds 00:06:40.290 --> 00:06:44.009 heavy. What is a declarative bias doing? Declarative 00:06:44.009 --> 00:06:46.990 bias restricts the search space. It mathematically 00:06:46.990 --> 00:06:49.670 dictates the boundaries of what the AI is even 00:06:49.670 --> 00:06:51.970 allowed to consider. Give me an example of that. 00:06:52.629 --> 00:06:55.629 So if the AI looks at the metadata of a new problem, 00:06:55.970 --> 00:06:58.310 it might dynamically alter its declarative bias 00:06:58.310 --> 00:07:01.170 to say, based on these initial properties, I'm 00:07:01.170 --> 00:07:03.209 only going to look at linear functions. Oh, I 00:07:03.209 --> 00:07:05.470 see. By actively restricting the representation 00:07:05.470 --> 00:07:07.850 of the problem, it slashes the amount of searching 00:07:07.850 --> 00:07:10.350 the algorithm has to do. Got it. So declarative 00:07:10.350 --> 00:07:13.089 bias is the AI setting up fences around the playground. 00:07:13.350 --> 00:07:15.199 That's a good way to put it. It looks at the 00:07:15.199 --> 00:07:17.459 task, decides half the playground is a complete 00:07:17.459 --> 00:07:19.819 waste of time, and just blocks it off so it doesn't 00:07:19.819 --> 00:07:22.639 wander into the woods. Exactly. So what is procedural 00:07:22.639 --> 00:07:25.699 bias then? Procedural bias dictates the strategy 00:07:25.699 --> 00:07:28.959 within those fences. It imposes constraints on 00:07:28.959 --> 00:07:31.439 the ordering of the hypotheses the AI tests. 00:07:31.560 --> 00:07:34.360 So within that restricted playground. How does 00:07:34.360 --> 00:07:37.240 it systematically try out solutions? Right. A 00:07:37.240 --> 00:07:39.699 procedural bias might be dynamically set. So 00:07:39.699 --> 00:07:42.540 the AI always tests the smallest, simplest mathematical 00:07:42.540 --> 00:07:45.620 hypotheses first before moving on to incredibly 00:07:45.620 --> 00:07:48.819 complex ones. Oh, wow. Yeah. So by intelligently 00:07:48.819 --> 00:07:51.220 shifting these two biases together based on the 00:07:51.220 --> 00:07:54.339 problem's metadata, the AI zeros in on a solution 00:07:54.339 --> 00:07:57.360 exponentially faster than just random trial and 00:07:57.360 --> 00:07:59.600 error. Here's where it gets really interesting 00:07:59.600 --> 00:08:02.399 because going from theory to actual software 00:08:02.399 --> 00:08:04.800 architecture, is where the magic happens. Oh, 00:08:04.800 --> 00:08:07.480 definitely. The source outlines three primary 00:08:07.480 --> 00:08:09.600 ways engineers are currently building these meta 00:08:09.600 --> 00:08:11.879 -learning systems, and they function very differently. 00:08:12.000 --> 00:08:15.259 Let's walk through them. The first is model -based 00:08:15.259 --> 00:08:17.740 meta -learning. Right, and the core goal of the 00:08:17.740 --> 00:08:20.399 model -based approach is extremely rapid adaptation. 00:08:20.800 --> 00:08:24.079 Okay. These systems often use cyclic networks 00:08:24.079 --> 00:08:27.160 and rely on external or internal memory banks. 00:08:27.500 --> 00:08:30.220 The intention is to update the model's parameters, 00:08:30.459 --> 00:08:32.879 its internal settings, with just a handful of 00:08:32.879 --> 00:08:35.419 training steps. Rather than the thousands of 00:08:35.419 --> 00:08:38.379 steps a normal neural network requires. Exactly. 00:08:38.419 --> 00:08:40.580 And a standout example of this from the text 00:08:40.580 --> 00:08:44.639 is memory augmented neural networks, or MENs. 00:08:44.759 --> 00:08:47.580 Yes, MENs are fascinating. From what I understand, 00:08:47.840 --> 00:08:51.610 standard deep learning has to slowly adjust millions 00:08:51.610 --> 00:08:54.929 of internal weights every single time it sees 00:08:54.929 --> 00:08:56.850 a new piece of data. It's a tedious process. 00:08:57.110 --> 00:08:59.580 Very slow. But a make -in essentially has a built 00:08:59.580 --> 00:09:02.600 -in scratch pad. It can rapidly write new information 00:09:02.600 --> 00:09:04.440 to an external memory bank and then read from 00:09:04.440 --> 00:09:06.940 it immediately. It can adapt to an entirely new 00:09:06.940 --> 00:09:08.960 task after seeing only three or four examples 00:09:08.960 --> 00:09:10.899 because it's storing the knowledge differently. 00:09:11.100 --> 00:09:13.840 Exactly. It bypasses the slow, gradual weight 00:09:13.840 --> 00:09:16.740 -updating process of standard networks by leaning 00:09:16.740 --> 00:09:18.899 on that fast access memory. Which is brilliant. 00:09:19.159 --> 00:09:21.879 It is. Which leads us to the second major architectural 00:09:21.879 --> 00:09:24.840 approach, metric -based meta -learning. OK. How 00:09:24.840 --> 00:09:26.940 is this one different? Well, this operates on 00:09:26.940 --> 00:09:29.919 a completely different philosophy. The core idea 00:09:29.919 --> 00:09:32.559 here is related to nearest neighbors' algorithms. 00:09:33.120 --> 00:09:35.360 The system generates a weight using a kernel 00:09:35.360 --> 00:09:39.059 function. Ultimately, it is trying to learn an 00:09:39.059 --> 00:09:41.779 effective metric, a mathematical distance function 00:09:41.779 --> 00:09:44.299 between different objects. Okay, kernel function 00:09:44.299 --> 00:09:46.399 and distance function. Let me try to translate 00:09:46.399 --> 00:09:47.980 that into something we use every day. Please 00:09:47.980 --> 00:09:50.399 do. Think of a highly advanced matchmaking app. 00:09:50.440 --> 00:09:54.909 Okay. A basic, rigid app. just looks for exact 00:09:54.909 --> 00:09:57.710 overlaps, right? Like you both selected dogs 00:09:57.710 --> 00:10:00.549 and hiking, therefore you are a match. Right, 00:10:00.889 --> 00:10:02.909 standard pattern matching. But a metric -based 00:10:02.909 --> 00:10:06.159 meta learner does something much deeper. It doesn't 00:10:06.159 --> 00:10:08.659 just figure out who is similar, it dynamically 00:10:08.659 --> 00:10:11.600 learns what metric actually matters in a given 00:10:11.600 --> 00:10:13.980 context. It learns how to mathematically measure 00:10:13.980 --> 00:10:16.860 the distance between two user profiles depending 00:10:16.860 --> 00:10:19.240 on the task. Because the criteria for finding 00:10:19.240 --> 00:10:21.440 a romantic partner is completely different from 00:10:21.440 --> 00:10:23.460 finding a tennis buddy or a business co -founder. 00:10:24.179 --> 00:10:27.159 Exactly. The AI dynamically learns how to measure 00:10:27.159 --> 00:10:29.679 the space between data points based on what it's 00:10:29.679 --> 00:10:31.840 trying to achieve. That is a brilliant way to 00:10:31.840 --> 00:10:34.250 visualize it. Thanks. It isn't learning the data 00:10:34.250 --> 00:10:36.610 itself. It is learning the relationship between 00:10:36.610 --> 00:10:39.649 the inputs. A perfect example of this in practice 00:10:39.649 --> 00:10:42.490 is convolutional Siamese neural networks. Siamese 00:10:42.490 --> 00:10:45.009 neural networks. Yeah. As the name implies, this 00:10:45.009 --> 00:10:48.370 architecture uses two twin networks. They share 00:10:48.370 --> 00:10:51.049 the exact same internal weights and parameters. 00:10:51.149 --> 00:10:54.070 OK. You feed one image into the left twin and 00:10:54.070 --> 00:10:56.669 a different image into the right twin. The networks 00:10:56.669 --> 00:10:59.750 aren't trying to output a label like, this is 00:10:59.750 --> 00:11:02.379 a cat. What are they doing instead? They're jointly 00:11:02.379 --> 00:11:04.899 trained to output the mathematical distance, 00:11:05.080 --> 00:11:06.860 or the difference between the two inputs. Oh, 00:11:06.860 --> 00:11:08.919 wow. So if you show it a new type of machinery 00:11:08.919 --> 00:11:11.559 it has never seen before, it doesn't need to 00:11:11.559 --> 00:11:13.460 know what the machine is called. Nope. It just 00:11:13.460 --> 00:11:15.539 needs to look at a baseline photo, look at the 00:11:15.539 --> 00:11:19.019 new photo, and say these are mathematically 99 00:11:19.019 --> 00:11:22.960 % similar. It learns how to compare things, not 00:11:22.960 --> 00:11:25.279 how to name them. Exactly. Another variation 00:11:25.279 --> 00:11:27.740 mentioned in the text is prototypical networks. 00:11:28.059 --> 00:11:31.200 How do those work? In this setup, the AI creates 00:11:31.200 --> 00:11:33.860 a generalized average prototype representation 00:11:33.860 --> 00:11:36.840 for a class of items. When a new piece of data 00:11:36.840 --> 00:11:39.820 comes in, the network just computes the distance 00:11:39.820 --> 00:11:42.440 to these prototypes. Like comparing a new dog 00:11:42.440 --> 00:11:45.039 to its ultimate concept of a dog. Precisely. 00:11:45.259 --> 00:11:48.220 It is a highly effective, simplified inductive 00:11:48.220 --> 00:11:50.919 bias that yields phenomenal results for what 00:11:50.919 --> 00:11:53.600 we call few -shot learning. Which is when you 00:11:53.600 --> 00:11:55.720 only have a tiny amount of data to work with, 00:11:55.799 --> 00:11:58.159 right? Exactly. Okay, so we have model -based 00:11:58.250 --> 00:12:01.269 using memory scratch pads and metric -based learning 00:12:01.269 --> 00:12:04.350 how to measure the distance between things. That 00:12:04.350 --> 00:12:06.450 brings us to the third approach, which sounds 00:12:06.450 --> 00:12:10.429 the most aggressive. Optimization -based. Metal 00:12:10.429 --> 00:12:12.629 learning. Yeah, this approach goes straight for 00:12:12.629 --> 00:12:15.690 the engine. Optimization based algorithms explicitly 00:12:15.690 --> 00:12:18.509 adjust the underlying optimization process itself. 00:12:18.730 --> 00:12:21.210 Okay. A massive milestone here was introduced 00:12:21.210 --> 00:12:25.190 in 2017. It's known as MM, a model agnostic meta 00:12:25.190 --> 00:12:27.169 learning. I want to break down that term model 00:12:27.169 --> 00:12:29.389 agnostic because that implies it doesn't actually 00:12:29.389 --> 00:12:31.190 care what kind of neural network you're running 00:12:31.190 --> 00:12:33.070 it on, right? That is exactly what it means. 00:12:33.769 --> 00:12:36.549 MAML is compatible with virtually any model that 00:12:36.549 --> 00:12:38.759 learns through gradient descent. And for anyone 00:12:38.759 --> 00:12:41.360 who isn't deep into the math, gradient descent 00:12:41.360 --> 00:12:43.820 is basically the AI trying to find the lowest 00:12:43.820 --> 00:12:46.600 point of error. Right. Imagine the AI is standing 00:12:46.600 --> 00:12:49.759 on a foggy mountain and its goal is to find the 00:12:49.759 --> 00:12:52.820 bottom of the deepest valley. It takes a step, 00:12:53.200 --> 00:12:56.440 feels the slope, and walks downward. That's gradient 00:12:56.440 --> 00:12:59.100 descent. That's a great visual. But normally, 00:12:59.220 --> 00:13:01.460 finding that valley takes thousands and thousands 00:13:01.460 --> 00:13:04.039 of steps. Right. A standard model trains its 00:13:04.039 --> 00:13:06.059 parameters to reach the bottom of one specific 00:13:06.059 --> 00:13:09.659 valley for one specific task. MAML, however, 00:13:09.940 --> 00:13:12.179 doesn't train for one valley. It trains the parameters 00:13:12.179 --> 00:13:15.259 across a huge sequence of different tasks. So 00:13:15.259 --> 00:13:17.659 what is its actual goal? Its goal is to find 00:13:17.659 --> 00:13:19.840 a starting point on the mountain that is mathematically 00:13:19.840 --> 00:13:22.279 close to all the valleys. Oh, wow. So if you're 00:13:22.279 --> 00:13:24.629 listening to this on your commute, Think of Mamm 00:13:24.629 --> 00:13:26.690 -ML like your brain when you jump into a rental 00:13:26.690 --> 00:13:30.210 car. You already know how to drive. You don't 00:13:30.210 --> 00:13:32.210 need to relearn the physics of a steering wheel 00:13:32.210 --> 00:13:34.909 or the concept of a brake pedal. Because of your 00:13:34.909 --> 00:13:37.789 prior experience across many cars, your brain 00:13:37.789 --> 00:13:39.990 is pre -optimized. Right, you're not starting 00:13:39.990 --> 00:13:42.529 from scratch. Exactly. You just need two minutes 00:13:42.529 --> 00:13:44.389 to fine -tune your environment to figure out 00:13:44.389 --> 00:13:46.710 where the wipers are and how sensitive the gas 00:13:46.710 --> 00:13:50.490 pedal is. And Mamm -ML trains the AI to be the 00:13:50.490 --> 00:13:53.340 driver of the rental car. As the original researchers 00:13:53.340 --> 00:13:55.919 stated, it explicitly trains the model to be 00:13:55.919 --> 00:13:58.500 easy to fine tune. It is an incredibly powerful 00:13:58.500 --> 00:14:01.759 concept. It essentially primes the network so 00:14:01.759 --> 00:14:04.639 that when an entirely new unseen task arrives, 00:14:05.019 --> 00:14:07.460 it only takes a tiny amount of training data 00:14:07.460 --> 00:14:10.519 and just a few steps of gradient descent to reach 00:14:10.519 --> 00:14:12.659 expert level performance. It really is. It sounds 00:14:12.659 --> 00:14:14.899 like the holy grail of artificial intelligence. 00:14:15.039 --> 00:14:17.419 It does. You build a foundation that is inherently 00:14:17.419 --> 00:14:19.799 prepped to learn anything on the fly. But we 00:14:19.799 --> 00:14:22.399 have to talk about the reality of deploying this 00:14:22.399 --> 00:14:24.519 outside of a laboratory. Yes, we do. Because 00:14:24.519 --> 00:14:27.860 the literature points out a massive glaring flaw 00:14:27.860 --> 00:14:30.279 in how these optimization algorithms have historically 00:14:30.279 --> 00:14:32.720 worked. What happens when the AI gets it wrong? 00:14:32.909 --> 00:14:35.129 Yeah, there is a very dangerous catch when you 00:14:35.129 --> 00:14:37.409 implement this in the real world. Historically, 00:14:37.450 --> 00:14:39.769 when a meta -learning system is juggling a diverse 00:14:39.769 --> 00:14:42.929 set of tasks, it optimizes for the average score 00:14:42.929 --> 00:14:45.590 across all of them. Right. It wants the highest 00:14:45.590 --> 00:14:47.909 mean performance. And this is where statistics 00:14:47.909 --> 00:14:50.330 can lie to you. Oh, absolutely. If you are only 00:14:50.330 --> 00:14:53.230 looking at the average, you are inevitably masking 00:14:53.230 --> 00:14:56.529 some catastrophic individual failures. Yep. Like, 00:14:56.529 --> 00:14:59.610 if I take five exams and I score 100 on four 00:14:59.610 --> 00:15:01.990 of them, but I get a zero on the fifth, my average 00:15:01.990 --> 00:15:05.769 is an 80. Which isn't bad. Right. On paper, to 00:15:05.769 --> 00:15:08.070 an algorithm looking at the mean, I look like 00:15:08.070 --> 00:15:11.990 a solid B student. But if that zero was my driver's 00:15:11.990 --> 00:15:15.029 license exam, I absolutely should not be operating 00:15:15.029 --> 00:15:17.169 a vehicle. Oh no, you should not. If you apply 00:15:17.169 --> 00:15:20.669 this to real -world AI like medical diagnostics 00:15:20.669 --> 00:15:23.250 predicting diseases or autonomous vehicles navigating 00:15:23.250 --> 00:15:27.330 streets, sacrificing the performance of one specific 00:15:27.330 --> 00:15:29.870 task just to keep the overall average high is 00:15:29.870 --> 00:15:32.470 lethal. It's totally unacceptable. It is a critical 00:15:32.470 --> 00:15:35.309 vulnerability and researchers realized that they 00:15:35.309 --> 00:15:38.190 had to address it immediately. This led to a 00:15:38.190 --> 00:15:40.950 very recent development from 2023 called Rommel. 00:15:41.120 --> 00:15:43.879 Which stands for? Robust meta -reinforcement 00:15:43.879 --> 00:15:47.340 learning. OK. Rommelau completely flips the objective. 00:15:47.539 --> 00:15:50.360 Instead of chasing a high average score, it focuses 00:15:50.360 --> 00:15:52.960 entirely on the worst -case scenarios. Really? 00:15:53.120 --> 00:15:55.940 Yeah. It identifies the tasks where the AI is 00:15:55.940 --> 00:15:58.679 performing the poorest and dynamically shifts 00:15:58.679 --> 00:16:01.240 computing resources to improve those specific 00:16:01.240 --> 00:16:04.120 low scores. So it guarantees a baseline of safety. 00:16:04.220 --> 00:16:07.600 Exactly. It refuses to let any single task drop 00:16:07.600 --> 00:16:09.919 below a certain threshold, ensuring the system 00:16:09.919 --> 00:16:12.879 is actually robust, regardless of what environment 00:16:12.879 --> 00:16:15.059 it's dropped into. And the brilliant part is 00:16:15.059 --> 00:16:17.840 that Ramalel is a meta algorithm itself. You 00:16:17.840 --> 00:16:20.340 can basically stack it on top of Mamalel. That's 00:16:20.340 --> 00:16:22.519 right. It forces the underlying system to care 00:16:22.519 --> 00:16:25.659 about the outliers. And this layering of algorithms, 00:16:25.779 --> 00:16:28.460 you know, AI managing AI, is really the defining 00:16:28.460 --> 00:16:30.899 characteristic of this entire field. It's moving 00:16:30.899 --> 00:16:33.500 so fast. We have moved incredibly fast from the 00:16:33.500 --> 00:16:35.779 theoretical evolutionary blueprints of the 80s 00:16:35.779 --> 00:16:39.620 to tangible world -altering software. Back in 00:16:39.620 --> 00:16:42.720 2017, Google Brain launched its AutoML project, 00:16:43.000 --> 00:16:45.240 which was literally described as AI building 00:16:45.240 --> 00:16:48.080 AI. I remember that. That system actually managed 00:16:48.080 --> 00:16:50.500 to design a neural network architecture that 00:16:50.500 --> 00:16:53.860 briefly exceeded the best, most optimized networks 00:16:53.860 --> 00:16:57.059 designed by human experts on standard benchmarks. 00:16:57.539 --> 00:17:00.399 AI successfully designing an AI that beats human 00:17:00.399 --> 00:17:04.460 engineers. That is wild. It is. This raises an 00:17:04.460 --> 00:17:06.960 important question, though. If we project this 00:17:06.960 --> 00:17:09.299 exact trajectory into the future, where does 00:17:09.299 --> 00:17:11.559 it end? Right. What's the end game? Well, the 00:17:11.559 --> 00:17:14.440 mathematics of this field point toward an extreme 00:17:14.440 --> 00:17:16.710 endpoint. It's a theory. a theoretical construct 00:17:16.710 --> 00:17:19.210 known in the literature as the Godel machine. 00:17:19.509 --> 00:17:21.569 The Godel machine. It sounds like something out 00:17:21.569 --> 00:17:23.730 of a sci -fi novel. What exactly is it? It is 00:17:23.730 --> 00:17:26.529 a theoretical ultimate meta -learning system. 00:17:26.950 --> 00:17:30.170 It contains a general theorem prover. And what 00:17:30.170 --> 00:17:32.329 makes the Godel machine unique is that it has 00:17:32.329 --> 00:17:34.789 full access to its own source code. Meaning? 00:17:34.990 --> 00:17:37.809 It can inspect and modify any part of its own 00:17:37.809 --> 00:17:41.049 software at will. It is designed to achieve recursive 00:17:41.049 --> 00:17:43.750 self -improvement. Well, lots of AI systems tweak 00:17:43.750 --> 00:17:46.039 their weights, though. What makes this the ultimate 00:17:46.039 --> 00:17:47.779 endpoint? What makes it the endpoint is that 00:17:47.779 --> 00:17:50.339 its self -improvement is mathematically provably 00:17:50.339 --> 00:17:53.420 optimal. Wait, provably optimal? Yes. It doesn't 00:17:53.420 --> 00:17:55.720 use trial and error. It doesn't guess what might 00:17:55.720 --> 00:17:58.680 make it better. Before it rewrites a single line 00:17:58.680 --> 00:18:02.160 of its own code, it uses its theorem prover to 00:18:02.160 --> 00:18:04.579 mathematically guarantee that the change will 00:18:04.579 --> 00:18:08.059 increase its ability to solve problems. It completely 00:18:08.059 --> 00:18:11.079 masters the metal learning process, continually 00:18:11.079 --> 00:18:13.519 and optimally accelerating its own intelligence 00:18:13.519 --> 00:18:17.500 over a single lifelong run. It is the absolute 00:18:17.500 --> 00:18:20.339 theoretical ceiling of an algorithm shifting 00:18:20.339 --> 00:18:23.420 its own inductive bias. So what does this all 00:18:23.420 --> 00:18:25.099 mean? Let's bring this all the way back to you, 00:18:25.200 --> 00:18:27.299 the person listening right now. Every single 00:18:27.299 --> 00:18:30.000 day, you are tweaking your own internal procedural 00:18:30.000 --> 00:18:32.420 bias. Yeah, you are. You choose to listen to 00:18:32.420 --> 00:18:34.519 deep dives. You figure out whether you absorb 00:18:34.519 --> 00:18:36.240 information better through audio or reading. 00:18:36.680 --> 00:18:38.400 You aggressively try to filter out the noise 00:18:38.400 --> 00:18:40.480 of the internet so you can learn faster and adapt 00:18:40.480 --> 00:18:42.680 to new challenges at your job. Yeah. You are, 00:18:42.819 --> 00:18:46.099 by definition, a meta learner. Absolutely. And 00:18:46.099 --> 00:18:48.700 computer scientists are painstakingly engineering 00:18:48.700 --> 00:18:51.099 artificial intelligence to do exactly what you 00:18:51.099 --> 00:18:53.779 do naturally. For decades, machine learning was 00:18:53.779 --> 00:18:56.190 essentially about giving an AI a fish. feeding 00:18:56.190 --> 00:18:58.549 it millions of pixels so it could recognize a 00:18:58.549 --> 00:19:01.190 cat or a stop sign. Exactly, but metal learning 00:19:01.190 --> 00:19:04.150 is the painstaking process of teaching the AI 00:19:04.150 --> 00:19:07.029 how to build the perfect fishing rod for whatever 00:19:07.029 --> 00:19:09.329 ocean it suddenly finds itself in. And it goes 00:19:09.329 --> 00:19:11.630 even further than that. It is teaching the machine 00:19:11.630 --> 00:19:14.160 to understand the physics of the water. to analyze 00:19:14.160 --> 00:19:16.920 the behavior of the fish, and to physically alter 00:19:16.920 --> 00:19:19.460 the rod's design in real time to guarantee a 00:19:19.460 --> 00:19:21.740 catch no matter the conditions. Which leaves 00:19:21.740 --> 00:19:24.539 us with a rather profound and honestly a slightly 00:19:24.539 --> 00:19:27.839 haunting thought to mull over. Yeah. If we follow 00:19:27.839 --> 00:19:30.960 this meta -learning path all the way to its logical 00:19:30.960 --> 00:19:33.819 conclusion, to that theoretical Goethe machine 00:19:33.819 --> 00:19:37.660 we just discussed, An AI that perfectly, flawlessly, 00:19:37.900 --> 00:19:40.380 and provably optimizes its own source code and 00:19:40.380 --> 00:19:42.960 learning procedures across every possible domain, 00:19:43.599 --> 00:19:46.039 does human -led machine learning research eventually 00:19:46.039 --> 00:19:48.480 make itself completely obsolete? It's a heavy 00:19:48.480 --> 00:19:51.559 question. If an AI becomes the ultimate self 00:19:51.559 --> 00:19:54.140 -improving meta -learner, the very last thing 00:19:54.140 --> 00:19:56.440 human engineers may ever need to teach a machine 00:19:56.440 --> 00:19:59.720 is simply the desire to learn. Thank you for 00:19:59.720 --> 00:20:01.579 joining us on this deep dive into the architecture 00:20:01.579 --> 00:20:03.890 of intelligence. Keep questioning the systems 00:20:03.890 --> 00:20:05.509 around you and keep learning how to learn.