EPISODE TRANSCRIPT
Building Machines That Learn and Think Like People (Brenden Lake et al.)
Automatically transcribed from the episode audio. Some words and names may be inaccurate.
Listen to the episodeDownload textDownload captions
Welcome back to the Deep Dive. Today, we are clearing the table. We are taking all the headlines, all the hype, all the panic about robots taking over the world, and we are putting them to the side for a moment. For just a moment. We are going to tackle one of the absolute giants of modern technology, but we're coming at it from a slightly different angle than you might expect. We are talking about artificial intelligence. Indeed. But we aren't just going to marvel at the latest chat bot or the newest image generator. We are going to look under the hood. We're looking at a fundamental critique of the current direction of AI and a roadmap for where it actually needs to go if we want it to be, well, Intelligent.
Exactly. We're basing this deep dive on a fascinating heavyweight paper titled, Building Machines That Learn and Think Like People. It's by Brendan Lake, Tomer Ullman, Joshua Tennenbaum, and Samuel Gershman. These are heavy hitters. MIT, NYU, Harvard. Right. And the title really says it all. We have machines that can do amazing things, right? They can beat us at go. They can recognize a cat in a photo. They can transcribe our speech with uncanny accuracy. Correct. We are currently living through what you might call a boom cycle in AI history, specifically the boom of deep learning and neural networks. It's the engine behind almost everything you see in the news.
But, and here's the hook, the thing that stops you in your tracks. This paper argues that despite all those wins, despite the crushed high scores and the viral videos, there is a massive fundamental piece missing. A huge piece. The central premise here is that while biological inspiration, you know, neurons, brains got us this far, current AI is actually nothing like human intelligence. It's a bold claim, but the evidence they present is compelling. The authors argue that if we really want machines that think we can't just keep building bigger chips and feeding them more data, we need to reverse engineer the human mind. Reverse engineer the mind. We need to look at the software of cognition. not just the hardware of neurons.
I love that distinction. And just to set the stage for everyone listening, the authors break this down into a battle between two heavyweights, pattern recognition versus model building. Pattern recognition is what our current AI does, right? Precisely. Pattern recognition is mapping inputs to outputs. You show the machine a million pixels and it says cat. Wow. It's entirely focused on prediction. OK. But the authors argue that human intelligence is about model building. It's not just predicting, it's explaining. It's constructing a causal model of how the world actually works.
So our mission for this deep dive is to figure out what that model building actually looks like. We're going to look at the startup software we are all born with, how we learn so much from so little data, and how to bridge the gap between fast reactions and deep thinking. It is the roadmap to the next generation of AI. We are moving from the how do we make it faster? phase to the, how do we make it understand phase? OK, let's unpack the landscape first, because we have to give credit where credit is due. Deep learning has been a juggernaut. It's not like these scientists are just saying AI is useless. Absolutely not.
I mean, if you look at the last decade, the progress is undeniable. We're talking about convenience, convolutional neural networks. In tasks like object recognition, they managed to have the error rates of previous technologies almost overnight. Have them? Yeah, that was around 2012. And it changed everything is the ImageNet moment. And we've seen DeepMind's DQN, that's the DeepQ network, playing Atari games at superhuman levels. I remember watching those clips of the computer playing Breakout. Oh, yeah. It figured out that if it tunneled the ball up the side of the wall, the ball would bounce around the top and destroy all the bricks without the paddle having to move.
It was genius, in a way. It is impressive. Yeah. But we need to understand how it does that. At a very high level, these deep neural networks are layers of representations. They use an algorithm called backpropagation. OK, let's slow down and explain backpropagation, because it's the engine of the whole industry. How does it actually work? Can you break that down for us? For sure. Think of it as a correction loop. The system makes a guess. It sees an image and says, that's a toaster. The answer key says, no, that's a cat. The system then calculates the error mathematically, how wrong was I? Yeah. Then it propagates that error backward. through the layers of the network.
Ah, so that's the back part of back propagation. That's the one. It adjusts the connections, the weights between the artificial neurons. It essentially says, hey, next time you see this pattern, don't fire the toaster neuron so hard. Fire the cat neuron. And it does this just once. It does this millions of times, billions even, with millions of pictures of cats and toasters and everything else. And that is the catch. These systems are incredibly data -hungry. Right. To teach a neural network, to recognize a handwritten letter A, you might need to show it thousands of different examples of an A. Whereas if I show my three -year -old an A once, maybe twice, she's got it, she can recognize it, she can draw it, she can probably even spot it in a weird font or carved into a tree. Exactly. And that brings us to the core distinction the paper makes. Humans don't just predict jada, we explain it. We explain it. The paper uses a beautiful analogy. Think of a scientist. A scientist doesn't just want to predict where a planet will be next Tuesday.
They want to understand the laws of gravity that cause the planet to be there. So human thought is a model -building activity. We are constantly creating little theories about the world. If I know the theory of gravity, I can predict the planet, but I can also predict what happens if a meteor hits the planet or if the Sun disappears. Yes, you have a flexible model. Deep learning, largely, is just fitting a curve to data points. It doesn't have the underlying theory. And this brings us to a historical debate that the paper touches on, which is the blank slate myth. Oh, I know this one, the John Locke idea. It's the idea that we are born knowing absolutely nothing and experience writes everything onto our minds.
Alan Turing even compared the child's mind to a notebook with lots of blank sheets. It is a romantic idea. And it aligns well with how current neural networks work. They start with random weights, knowing nothing, and learn everything from the data. But cognitive science tells us this is simply not true for humans. If we were truly blank slates, we would never learn as fast as we do. The math just doesn't work. There's too much to learn. Way too much. The authors argue we possess inductive biases, or what they call startup software. Startup software? I like that. It implies we come preloaded with an operating system that helps us interpret the world immediately.
We aren't just hardware waiting for data. We have a bootloader. Correct. And before we get into the specifics of that software, the paper presents two challenges to illustrate just how big the gap is between machines and humans. They call them the Characters Challenge and the Frostbite Challenge. Let's start with the Characters Challenge. This is about handwriting, right? It is. The standard benchmark in AI for this is the MNIST of data set. It's a massive collection of handwritten digits, zero through nine. I've seen this one. It's like the hello world of machine learning. It basically is. And current deep learning algorithms can hit an error rate of 0 .2 % on this. They're practically perfect at looking at a scribble and saying, that's a seven. So game over. Machines win. Well, if the game is just classifying an image, then yes, it's a tie. But the authors argue that simply correctly labeling an image isn't the same as understanding the concept.
Okay, so what's the difference? How do we prove the machine doesn't understand? The difference is one -shot learning and rich representation. As you mentioned with your daughter, humans can learn from a single example. Right. If I show you a character from the Tibetan alphabet, which you've likely never seen, you can immediately recognize it again. But more than that, you can draw it. Right. Because when I look at a handwritten character, I'm not just seeing a grid of black and white pixels. I'm seeing strokes. I'm seeing the motion. I can see where the pen started and where it lifted. I'm unconsciously reverse engineering the hand that wrote it. Exactly. You are seeing a program. You interpret the image as a series of causal actions. A neural network sees a texture.
It sees a correlation of pixels. That's a great way to put it. A program versus a texture. And this is why humans can pass what the authors call a visual Turing test. If I showed you a Japanese character and asked you to invent a new character that looks like it belongs in the same alphabet, you could do it. You understand the style. And a neural network can't do that. Not easily. Not from one example. It can classify. But asking it to generate a variation based on one example is incredibly difficult for a system that relies on pattern matching thousands of examples. What would it produce? It creates something that looks like a blurry average, not a crisp new logical symbol. It fails to capture the concept behind the character. That makes sense. It's the difference between memorizing a painting and knowing how to paint. Okay, let's move to the second challenge, the frostbite challenge.
This refers to the Atari game. Yes, Frostbite. It's a classic from the 80s. You control an agent, this little guy named Frostbite Bailey, jumping on ice floes to build an igloo. While also avoiding polar bears and managing your body temperature. Sounds stressful. It is. Now, the Deep Q network, Deep QN we mentioned earlier, the one that crushed breakout struggles significantly with Frostbite. Why? Is it just harder? More things on the screen? It's a different kind of hard. Breakout is very reactive. Ball comes down, move paddle. It's a very tight loop. Simple cause and effect. Right.
Frostbite requires temporally extended planning. You have to jump on flows to get ice, but only the active white flows count. And then you have to get back to shore to build the igloo, block by block. So there are steps. You have to have a goal in mind that isn't just survive the next second. Exactly. And to reach human level performance on Atari games, the DQN often needs thousands of hours of gameplay. A human can pick up frostbite and understand the basic goal in minutes. Thousands of hours. That's years of a full -time job just to learn to jump on ice floes. It is immensely inefficient. And the reason for the difference is, again, representation. The neural network sees a screen of pixels changing states. It learns that when this particular pattern of pixels appears, move joystick left. So it doesn't know what an ice floe is. Not at all. It doesn't know what a bear is. It doesn't know it's building an igloo.
It doesn't know that the bear is an enemy or that the ice is a solid platform that will support it. No, it just learns statistical correlations between pixel patterns and high scores. This is the critical part. If you change the color of the ice flows from white to, say, green or move the score counter to the bottom of the screen, the whole thing breaks. The neural networks might completely fail because the pattern it memorized has changed. A human wouldn't blink. We'd say, oh, the ice is green now. Fine. The physics of jumping on a platform is still the same. That's a huge distinction. We see objects and agents.
The machine sees data streams. It's a fragility problem. The machine is brilliant within the exact parameters it was trained on. But if you shift the world slightly, it breaks. And that leads us directly into section three. What is this startup software that allows us to see objects and agents so effortlessly? This is the stuff we're born with, or at least develop extremely early. The paper breaks it down into two main ingredients, intuitive physics and intuitive psychology. Let's look at ingredient A, intuitive physics. Even infants have primitive object concepts.
Developmental psychologists have shown this for decades. Right, like solidity. If I drop my keys, they don't pass through the floor. They hit the floor and stop. A baby is surprised when a magician seems to violate that rule. Exactly. Solidity, persistence, knowing an object still exists when it goes behind a couch in continuity. Objects move on connected paths. They don't just teleport from one place to another. And we don't have to learn this every time we play a new video game. It's just baked in. It's baked in. When you play Frostbite, you assume the ice flow is a solid thing that will hold you up. You assume if the bear goes off the left side of the screen, it might come back or it's at least somewhere off screen. It didn't cease to exist. But the AI has to learn that from scratch. Precisely.
We transfer this vast library of physical intuition to every new task. The neural network has to learn the concept of solid platform from scratch, pixel by pixel, every single time. It is trying to derive Newton's laws of motion from a video game score, which is an absurdly hard way to learn physics. OK, that makes perfect sense. Then we have ingredient B, intuitive psychology. This sounds a bit more complex. It's the understanding of agency. Infants understand that some objects are agents. They have mental states, goals, and beliefs. They act efficiently to achieve those goals. So a rock just falls, but a person walks.
There's intent there. There's intent. And we are wired to look for it. The Pac -Man example in the outline is great for this. If you see Pac -Man running away from a ghost, you don't just see two -colored blobs moving on a screen. No, I infer a story. The ghost is bad. Pac -Man doesn't want to die. I immediately get the relationship. And you infer that efficiently. You assume the agent is rational. This is called the efficiency principle. If the agent wants the food, it will go directly to the food. If it runs away, there must be a reason. So this is a huge learning shortcut. A massive one. It allows us to learn from observation incredibly fast. If you watch a pro play frostbite, you infer their goal. Oh, they're trying to jump on the white ice. You don't need to try and fail a thousand times to figure out that the white ice is good. You infer it through intuitive psychology. So we're basically mind reading the player to understand the game.
We are building a model of the player's intent. In a sense, yes. We are building a causal model of the player's mind. They want points, therefore jumping on white O -seals points. The machine just sees pixels changing color. It misses the Y. It has no model of other minds. Okay, so we have this startup software physics and psychology. It's the foundation. But how do we actually learn new things using it? That brings us to section four, learning as rapid model building. Right. And here the paper proposes a framework called Bayesian Program Learning, or BPL. Bayesian Program Learning. That sounds intense. Let's break that down. It's not as scary as it sounds. The core idea is the program analogy. Instead of thinking of the mind as a network of weights being adjusted, like a giant mixing board in a music studio, think of the mind as a software developer writing code. So when I learn a new concept, I'm writing a little computer program in my head. Metaphorically, yes. To understand the letter A, your mind writes a script. Draw an angled line left, draw an angled line right, cross in the middle. That is a generative model. It's a set of instructions that can produce the data. It's not a picture of an A, it's the recipe for an A. Perfect analogy. And because it's a program, it's flexible. You can run that program with a thick pen, a thin pen, on a napkin, on a whiteboard. the underlying code is the same.
OK, I get it. And the Bayesian part, what's that about? The Bayesian part just means it deals with probability. The model is uncertain. Maybe the line is a little wobbly. Maybe the crossbar is high or low. The model represents all these possibilities and which are more likely. I see. So it's a probabilistic program. Exactly. And this brings in three sub -ingredients that make this mental programming so powerful. The first is compositionality. This is one of my favorite concepts. It's basically the LEGO principle, right? It is. Compositionality is the ability to build complex ideas out of simpler, reusable parts. Take the visual example of vehicles. Right.
I know what a wheel is. I know what a handlebar is. I know what a chassis is. If I see a motorcycle, I don't memorize the whole image as a unique blob. I see Wheels plus chassis plus handlebars plus engine. And the power comes from combinatorics rearranging those parts. If I show you a segue, you've maybe never seen one before, but you recognize wheels and platform and handle. You can construct a model of this new object instantly by recombining old parts you already know. And deep networks struggle with this. They often struggle to reconfigure parts they haven't seen together before. If a network is trained on cars and bicycles, and then you show it a unicycle, it might be completely baffled. Because the pattern of a single wheel with a seat doesn't match its statistical database.
It looks for the bike pattern and fails. But a human says, oh, it's like a bike, but with only one wheel. We compose and decompose. That leads to the second sub ingredient, causality. We touched on this with handwriting. Yes. Causality is understanding the process that generates the data. It's not just the shape of the letter. It's the motor program of the hand that drew it. This is why we can spot a fake signature, right? Even if it looks good superficially. Often, yes. You can tell if a line was drawn slowly and shakily or quickly and confidently. You are reverse -engineering the cause. The paper argues that this causal understanding is what allows us to pass that visual Turing test we mentioned.
Because I know how it's made. If you know how a character is drawn, you can invent a new one that follows the same rules of construction. You're running your little mental program with new parameters. And finally, the third sub -ingredient. Learning to learn. This sounds like meta -learning. This is crucial. It's how we get so good at learning. We use knowledge from previous classes of objects to learn new ones faster. So if I learn the Greek alphabet, it helps me learn the Hebrew alphabet, even though the letters are different. In a way. You learn what an alphabet is. You learn the characters are usually made of simple strokes.
They have standard size. They are distinct from each other. You build a higher level model. The paper calls this hierarchical knowledge. Right. Level one is learning a specific letter. like alpha, level two is learning what do letters look like in general. Neural networks typically start from scratch every time. They don't have that level two knowledge saying, hey, this is a handwritten character, so expect strokes and lines. They don't learn the context. BPL and humans transfer that high level knowledge to speed up the specific learning. It's like, I know how distinct strokes usually connect. So when I see a new character, I can immediately start breaking it down into parts I understand.
So if we have BPL. this amazing, flexible, programmable way of thinking, why do we need neural networks at all? Or rather, why is the brain not just a slow, logical computer? Because sometimes I don't reason things out, sometimes I just react. That is the perfect segue to section five, real -time thinking. The problem with Bayesian inference and model building is that it is computationally expensive. It's slow. It's very slow. If a tiger jumps out at you, you don't have time to construct a causal model of the tiger's biology and for its goals using intuitive psychology and calculate the physics of its trajectory. You just run.
You need to run. That's pattern recognition. A big, stripy, toothy pattern means danger. So it's a trade -off. Exactly. This is the classic system one and system two idea from Daniel Kahneman. System one is fast, intuitive pattern matching. System two is slow, deliberative, model -based reasoning. And the argument in the paper is that the brain uses both, and they work together. Yes. And they introduce a fascinating concept called amortized inference. Amortized inference. OK, unpack that. It sounds like something from an accounting class. It sort of is. Think of it as muscle memory for thought. Let's say you are learning to solve a specific type of math problem. At first, you use model -based thinking, system two, step by step, logical, slow. You are running the program. Painfully slow. Like when you first learn to drive a stick shift, you're thinking, OK, clutch in, move the stick, gas down, clutch out. Exactly. But after you have done it 100 times, you start to just see the answer or feel the gear shift. You recognize the pattern. You don't think about it anymore.
You are no longer running the full causal model. You are simply mapping the input to the output based on past experience. You have amortized the cost of that thinking into a fast neural reflex. System 1 takes over. So the neural network, the pattern recognizer, is actually trained by the model builder. The slow, thoughtful part of me teaches the fast, reactive part of me. That is the theory. The paper mentions the Helmholtz machine idea from the 90s. Imagine your brain uses its causal models to dream or imagine situations. It generates its own data and then trains its fast pattern recognition networks on that imagined data. Wow, that sounds like what happens when we sleep. Consolidation. We replay the day's events. It connects beautifully to sleep research.
Offline processing might be the brain transferring knowledge from the slow, flexible, model -based systems to the fast, efficient reflex systems. That is a light bulb moment for me. It's a powerful idea. So the authors aren't saying neural networks are useless. They are saying neural networks are the output of a deeper learning process, not the whole process itself. Exactly. If you only build the fast reactions, like current AI, you get brittle systems that break when the game changes. What if you have the model? If you have the model, you can retrain the reflex. You can adapt. OK. We've covered a lot of ground. We've bashed deep learning a little bit, praised the human mind, and built a roadmap. But surely there are people screaming at their speakers right now with objections. Section 6 addresses the critiques. Oh, for sure. And the most common critique is the unfair comparison. Right. The argument goes, hey, you're comparing a newborn AI to an adult human. Humans have years of experience.
The neural net starts from nothing. If the net had 20 years of pre -training, it would match the human. It's a fair observation. But the authors have a really sharp rebuttal. I say. That is exactly the point. That's the feature, not the bub. Right. The challenge isn't just more data. The challenge is how to represent that prior experience so it can be used effectively. Humans don't just have a hard drive full of random video clips from their childhood. They have structured compositional causal knowledge, that startup software we talked about. The critique that humans have more prior knowledge just reinforces the argument that we need to figure out how to give machines that same kind of rich prior knowledge.
It's not about the quantity of the data. It's about the quality of the structure holding it. Perfectly said. OK. Critique number two, biological plausibility. This one comes from the neuroscientists, right? The brain is a neural network, so we should only use neural networks to model it. This is the hardware fallacy. And the authors use the aerodynamics versus imitating a bird analogy, which is just fantastic. I love this one. The Wright brothers didn't succeed by building a machine that flapped its wings and had feathers. No, they tried that and it failed. They succeeded by understanding the principles of aerodynamics, lift, drag, thrust. They understood the physics, not just the surface level biology. OK. The argument here is that we need to understand The software algorithms of intelligent symbolic processing, causal models, before we obsess over the hardware implementation of neurons. Just because the brain uses neurons doesn't mean the best way to explain intelligence is to simulate neurons one by one without a higher level theory. We need to understand the code, not just the transistors it runs on.
Exactly. What computations are those neurons performing? That's the key question. Okay, final critique. What about language? The paper doesn't talk much about language and isn't that the pinnacle of human intelligence. You see things like GPT -3 and it feels like we're close. Language is huge, but the authors argue that it sits on top of the startup software. It's built upon it. How so? You can't learn the word ball until you understand the concept of an object and the physics of solidity. You can't learn the word give until you understand agency and goals and other minds.
So language is the walls and the roof. But intuitive physics and intuitive psychology are the foundation. You need the foundation first. Yes. You need to understand the world before you can talk about it. Language maps onto these pre -existing conceptual structures. This brings us to the outro. What is the synthesis here? It feels like we are moving towards a hybrid model, not a competition. That is the vision. The future isn't neural nets versus cognitive science. It's the integration. The paper points towards something they call differentiable programming. Which sounds like the best of both worlds. It is. It's about combining the powerful gradient -based learning of deep learning with the structure, compositionality, and reasoning of program induction. We want machines that can learn from massive data and reason from first principles. And why does this matter to the listener?
Why should we care if machines think like this? What's the practical application? Imagine robotics. Right now, if you want a robot to fold laundry, you might have to train it on millions of towels. If you give it a shirt, it fails. It has no concept of laundry. Exactly. We want robots that can learn a new tool or a new environment instantly, like a human using a new drill for the first time. We want data efficiency. We want systems that don't need the entire internet to learn what a cat is. That's the dream. A robot you can show once, here's how you load the dishwasher, and it gets it. Because it understands dishes, water, And cleaning, not just pixel coordinates and motor commands, it has a model of the world.
Before we wrap up, there is one final slightly provocative thought at the end of the paper that I found fascinating. It's about evolution. Ah, yes, the stochastic hill climbing concept. The authors suggest that if we can't code these complex structures by hand, this startup software, and if gradient descent, which is standard deep learning, can't find them, then maybe we are missing the biggest algorithm of all, evolution. It's a bit of a mic drop moment. Evolution is a messy, slow, random search process. It took millions of years of trial and error to build the human brain's architecture. So, what are they suggesting? They ask a profound question. Are we trying to skip the millions of years of evolution that gave us our startup software in the first place? We want the result. The baby's mind. Without the process. without the millions of years of death and selection that forged it. It suggests that maybe the real path forward is to simulate that evolutionary process to build the right architectures, which then can learn rapidly like a child. We might be trying to make the child learn things that the species learned over eons. That is a heavy thought. We might be asking our AI to do too much learning and not enough inheriting.
We're giving it a blank slate when it needs a genetic history. It's a humbling perspective on the challenge ahead. This has been a massive deep dive. We've gone from Atari to handwritten digits, from neurons to Bayesian programs. It really changes how you look at the AI revolution, doesn't it? It feels like we are just at the beginning of the real challenge. I think so. We are past... the magic trick phase of AI. Now we have to do the hard work of understanding the mind. Thank you so much for breaking this down with me. This was incredible. And to our listeners, thanks for tuning in. As always, keep asking the hard questions. Keep thinking. We'll see you next time on the Deep Dive.