WEBVTT 00:00:00.000 --> 00:00:02.220 Imagine trying to hold a conversation with someone 00:00:02.220 --> 00:00:06.639 who has like severe and I mean absolute Short 00:00:06.639 --> 00:00:09.140 -term amnesia right that would be incredibly 00:00:09.140 --> 00:00:11.660 difficult Yeah, every single word you say to 00:00:11.660 --> 00:00:14.279 them they hear it they process it but the moment 00:00:14.279 --> 00:00:18.370 you say the next word the previous one just vanishes 00:00:18.370 --> 00:00:20.690 completely. It's completely wiped. Exactly. They 00:00:20.690 --> 00:00:23.390 experience every fraction of a second completely 00:00:23.390 --> 00:00:25.670 independently. I mean, it would make communication 00:00:25.670 --> 00:00:28.530 basically impossible. You simply cannot understand 00:00:28.530 --> 00:00:30.250 a sentence without holding onto the beginning 00:00:30.250 --> 00:00:32.109 of it by the time you reach the end. And for 00:00:32.109 --> 00:00:34.770 a long time, that was the exact problem we faced 00:00:34.770 --> 00:00:37.609 when trying to get artificial intelligence to 00:00:37.609 --> 00:00:40.270 understand us. Yeah, it really was. Because the 00:00:40.270 --> 00:00:42.890 basic neural networks, they were brilliant at 00:00:42.890 --> 00:00:46.659 looking at a single static picture. and identifying 00:00:46.659 --> 00:00:49.259 a cat, but they were stuck in this isolated present 00:00:49.259 --> 00:00:51.859 moment. They had absolutely no concept of time 00:00:51.859 --> 00:00:55.100 passing or of sequences. Which meant they failed 00:00:55.100 --> 00:00:58.259 completely at tasks involving speech or language 00:00:58.259 --> 00:01:00.420 translation or even just predicting what happens 00:01:00.420 --> 00:01:04.420 next in a time series. Solving that required 00:01:04.420 --> 00:01:06.640 a completely different architecture. It really 00:01:06.640 --> 00:01:10.069 required giving the machine a memory. That is 00:01:10.069 --> 00:01:12.969 exactly our mission for you today. We are doing 00:01:12.969 --> 00:01:17.390 a deep dive into a massive comprehensive Wikipedia 00:01:17.390 --> 00:01:21.329 article on recurrent neural networks, or RNNs. 00:01:21.590 --> 00:01:24.730 It's a fascinating topic. It really is. So if 00:01:24.730 --> 00:01:26.790 you've ever wondered how your phone dictates 00:01:26.790 --> 00:01:29.689 your voice, or how translation apps grasp the 00:01:29.689 --> 00:01:32.170 context of a whole paragraph, or really just 00:01:32.170 --> 00:01:35.370 how AI processes the simple passage of time, 00:01:35.629 --> 00:01:37.370 well, This is the master class. And we're going 00:01:37.370 --> 00:01:39.590 to cover a lot of ground today. We are. OK, let's 00:01:39.590 --> 00:01:41.849 unpack this. When I was prepping for this, the 00:01:41.849 --> 00:01:45.849 source kept contrasting RNNs with standard feedforward 00:01:45.849 --> 00:01:48.189 neural networks. Yes, that's the standard comparison. 00:01:48.310 --> 00:01:50.370 So as I understand it, a feedforward network 00:01:50.370 --> 00:01:52.829 just takes data, pushes it straight through its 00:01:52.829 --> 00:01:55.409 layers once, and spits out an answer. What makes 00:01:55.409 --> 00:01:58.290 an RNN fundamentally different? Well, the crucial 00:01:58.290 --> 00:02:00.689 difference is right in the name itself, recurrent. 00:02:00.849 --> 00:02:03.540 It implies a loop. In a standard feed -forward 00:02:03.540 --> 00:02:06.060 network, the data travels strictly in one direction, 00:02:06.180 --> 00:02:09.979 from input to output. And RNN alters that fundamental 00:02:09.979 --> 00:02:13.280 geometry. The output of a neuron at one step 00:02:13.280 --> 00:02:16.439 is actually fed back into the network as an input 00:02:16.439 --> 00:02:19.460 for the absolute very next step. Oh, wow. So 00:02:19.460 --> 00:02:21.479 it's essentially eating its own output as it 00:02:21.479 --> 00:02:24.000 moves forward. Exactly. That loop is what we 00:02:24.000 --> 00:02:26.479 call a recurrent unit, and it maintains something 00:02:26.479 --> 00:02:29.960 called a hidden state. Hidden state? Right. You 00:02:29.960 --> 00:02:32.580 can think of this hidden state as the network's 00:02:32.580 --> 00:02:36.060 internal working memory. So at any given moment, 00:02:36.479 --> 00:02:38.620 the network isn't just looking at the new piece 00:02:38.620 --> 00:02:40.520 of data coming in. It's looking at the past, 00:02:40.680 --> 00:02:45.060 too. Yes. It is processing that new data mathematically 00:02:45.060 --> 00:02:48.000 combined with its hidden state, which contains, 00:02:48.000 --> 00:02:50.759 well, the echoes of everything it has seen before. 00:02:51.060 --> 00:02:52.979 Let me try to visualize this with an analogy. 00:02:53.340 --> 00:02:55.520 It feels like reading a book. Okay, I like that. 00:02:55.710 --> 00:02:58.449 If I'm a basic feedforward AI, I read the word 00:02:58.449 --> 00:03:01.250 the, process it, and then completely wipe my 00:03:01.250 --> 00:03:03.409 mental slate. Right, gone. Then I read the word 00:03:03.409 --> 00:03:05.990 dog, and I have zero recollection that the came 00:03:05.990 --> 00:03:08.550 before it. It's just a standalone dog. Exactly. 00:03:08.909 --> 00:03:11.530 But an RNN carries a running summary in its head. 00:03:12.009 --> 00:03:14.490 That running summary is the hidden state. Spot 00:03:14.490 --> 00:03:16.590 on. So by the time it reaches the end of the 00:03:16.590 --> 00:03:19.879 sentence and reads the word barked, It understands 00:03:19.879 --> 00:03:23.780 that action in the context of the dog, the mailman, 00:03:24.020 --> 00:03:26.740 and everything that preceded it. That captures 00:03:26.740 --> 00:03:28.800 the dynamic perfectly. And what's fascinating 00:03:28.800 --> 00:03:31.759 here is how this architecture looks when you 00:03:31.759 --> 00:03:33.840 map it out visually. Oh yeah, the diagrams and 00:03:33.840 --> 00:03:35.699 the source material. Right. If you look at diagrams 00:03:35.699 --> 00:03:38.800 of RNNs, they often look like these massive sprawling 00:03:38.800 --> 00:03:41.879 networks with dozens of horizontal layers. But 00:03:41.879 --> 00:03:44.979 that visual is actually a clever illusion. An 00:03:44.979 --> 00:03:46.699 illusion. Wait, what are we actually looking 00:03:46.699 --> 00:03:49.159 at then? You aren't looking at a giant stack 00:03:49.159 --> 00:03:51.639 of different distinct layers. It is the exact 00:03:51.639 --> 00:03:54.960 same network, the same layer, just unfolded in 00:03:54.960 --> 00:03:57.800 time. Unfolded, okay. Yeah. You are seeing a 00:03:57.800 --> 00:04:00.280 snapshot of the same digital brain at second 00:04:00.280 --> 00:04:03.280 one, then second two, then second three, laid 00:04:03.280 --> 00:04:06.120 out side by side on the page. Oh, I see. It transforms 00:04:06.120 --> 00:04:09.020 the input, updates its internal memory, and gets 00:04:09.020 --> 00:04:11.020 ready for the next step using the exact same 00:04:11.020 --> 00:04:13.300 machinery. It just loops the same hardware over 00:04:13.300 --> 00:04:15.060 and over. That makes a lot of sense conceptually. 00:04:15.159 --> 00:04:17.420 It's very elegant. But, you know, when I think 00:04:17.420 --> 00:04:20.779 about the origins of advanced AI, my mind immediately 00:04:20.779 --> 00:04:23.800 goes to Silicon Valley computer labs in the 2000s 00:04:23.800 --> 00:04:26.120 or maybe the 90s. Well, sure. Most people do. 00:04:26.560 --> 00:04:28.560 Reading this source, though... The inspiration 00:04:28.560 --> 00:04:30.639 for this looping computational memory didn't 00:04:30.639 --> 00:04:33.220 come from computer science at all. No, it didn't. 00:04:33.339 --> 00:04:35.420 The origins are much more organic and they stretch 00:04:35.420 --> 00:04:38.740 back much further. The very concept of recurrency 00:04:38.740 --> 00:04:41.459 in this context comes straight from human anatomy 00:04:41.459 --> 00:04:43.759 and early neuroscience. Here's where it gets 00:04:43.759 --> 00:04:46.759 really interesting. We're talking about observations 00:04:46.759 --> 00:04:49.779 of the human brain from over a century ago. Right. 00:04:50.060 --> 00:04:53.339 In 1901, the legendary neuroscientist Santiago 00:04:53.339 --> 00:04:56.500 Ramonica Hall was studying the cerebellar cortex 00:04:56.500 --> 00:04:59.209 of the brain a microscope. Just looking at physical 00:04:59.209 --> 00:05:02.509 brain tissue. Exactly. And he observed these 00:05:02.509 --> 00:05:05.209 literal loop -like anatomical structures, he 00:05:05.209 --> 00:05:07.430 called them recurrent semicircles, and they were 00:05:07.430 --> 00:05:10.009 formed by different types of biological cells. 00:05:10.050 --> 00:05:13.050 Wow. And then moving into the 1930s and 40s, 00:05:13.209 --> 00:05:15.290 researchers began to realize that the brain wasn't 00:05:15.290 --> 00:05:17.709 just a linear pathway where, you know, a stimulus 00:05:17.709 --> 00:05:19.829 goes in and a reaction comes out. Our source 00:05:19.829 --> 00:05:23.230 brings up Donald Hebb in the 1940s who proposed 00:05:23.230 --> 00:05:26.910 reverberating circuits. Yes, Hebb's work is foundational. 00:05:27.149 --> 00:05:29.569 He had this idea that electrical signals in the 00:05:29.569 --> 00:05:31.709 brain could get caught in a biological loop, 00:05:32.110 --> 00:05:34.949 bouncing around continuously, and that this physical 00:05:34.949 --> 00:05:37.750 reverberation might actually be the physical 00:05:37.750 --> 00:05:40.839 mechanism for short -term memory. Following that 00:05:40.839 --> 00:05:44.180 exact thread, in 1943, Warren McCulloch and Walter 00:05:44.180 --> 00:05:47.660 Pitts published a landmark paper mathematically 00:05:47.660 --> 00:05:51.199 modeling how neurons work. And crucially, they 00:05:51.199 --> 00:05:53.379 included artificial networks that had cycles 00:05:53.379 --> 00:05:56.000 or loops in them. They were attempting to use 00:05:56.000 --> 00:05:58.579 these closed loops to explain medical conditions. 00:05:58.600 --> 00:06:01.220 For instance, why a seizure happens in epilepsy 00:06:01.220 --> 00:06:05.220 or certain types of chronic pain where a nerve 00:06:05.220 --> 00:06:09.680 signal simply will not stop firing. So the idea 00:06:09.680 --> 00:06:12.759 of an AI memory loop starts with literal brain 00:06:12.759 --> 00:06:15.779 anatomy. But the source doesn't stop at biology. 00:06:16.100 --> 00:06:17.819 No, it takes a pretty sharp turn. It takes this 00:06:17.819 --> 00:06:20.459 massive leap into physics. Statistical mechanics 00:06:20.459 --> 00:06:22.120 specifically, which completely threw me for a 00:06:22.120 --> 00:06:24.459 loop. You're referring to the Ising model. This 00:06:24.459 --> 00:06:27.519 dates back to the 1920s, developed by Wilhelm 00:06:27.519 --> 00:06:31.160 Lenz and Ernst Ising. It was originally a statistical 00:06:31.160 --> 00:06:35.000 model to understand magnets. Magnets. Yes. Specifically, 00:06:35.480 --> 00:06:38.300 how magnetic materials reach physical equilibrium. 00:06:38.740 --> 00:06:41.480 OK, hold on a second. I need a bridge here. How 00:06:41.480 --> 00:06:45.079 on earth do 1920s magnets connect to an artificial 00:06:45.079 --> 00:06:47.620 intelligence trying to understand the context 00:06:47.620 --> 00:06:49.579 of a sentence? I know, it sounds totally unrelated 00:06:49.579 --> 00:06:51.579 until you look at the math of energy states. 00:06:51.600 --> 00:06:54.639 Cool. In a material, atoms have a magnetic spin, 00:06:55.019 --> 00:06:57.319 and they constantly influence the spin of their 00:06:57.319 --> 00:06:59.199 immediate neighbors. They're trying to settle 00:06:59.199 --> 00:07:01.680 into the lowest possible energy state, which 00:07:01.680 --> 00:07:04.000 physicists call a local minimum. Right, they 00:07:04.000 --> 00:07:07.079 want to be stable. Exactly. And then in the 1970s 00:07:07.079 --> 00:07:09.980 and 80s, physicists and computer scientists realized 00:07:09.980 --> 00:07:12.519 you could use that exact same mathematical framework 00:07:12.519 --> 00:07:15.399 to model a network of artificial neurons. So 00:07:15.399 --> 00:07:17.959 instead of atoms settling into a magnetic state, 00:07:18.060 --> 00:07:20.500 you have artificial neurons firing and influencing 00:07:20.500 --> 00:07:22.779 each other's weights. The network iterates over 00:07:22.779 --> 00:07:25.259 and over until the whole system mathematically 00:07:25.259 --> 00:07:28.019 settles into a stable pattern. And that stable 00:07:28.019 --> 00:07:31.399 pattern is what we call a memory. That is wild. 00:07:31.860 --> 00:07:34.680 You have neuroscience mapping literal brain loops 00:07:34.680 --> 00:07:37.600 and physics calculating the energy states of 00:07:37.600 --> 00:07:40.899 magnetic atoms. And in the 1980s, these completely 00:07:40.899 --> 00:07:43.620 different disciplines crash into each other to 00:07:43.620 --> 00:07:46.199 form the foundation of modern recurrent neural 00:07:46.199 --> 00:07:49.339 networks. Yeah, by the late 80s and early 90s, 00:07:49.500 --> 00:07:51.560 with architectures like the Jordan and Elman 00:07:51.560 --> 00:07:54.100 networks, researchers were using these early 00:07:54.100 --> 00:07:56.540 RNNs to actually study cognitive psychology. 00:07:56.639 --> 00:07:59.680 They were mapping how human beings process sequential 00:07:59.680 --> 00:08:04.000 tasks. But as incredible as the theory was, the 00:08:04.000 --> 00:08:06.160 reality of building these things hit a massive 00:08:06.160 --> 00:08:09.300 roadblock. They built the RNNs, they gave them 00:08:09.300 --> 00:08:11.819 this looping mathematical memory, but the networks 00:08:11.819 --> 00:08:13.779 had a fatal flaw. They had terrible long -term 00:08:13.779 --> 00:08:16.220 memory. Yes, the vanishing gradient problem. 00:08:16.699 --> 00:08:19.079 This was the great amnesia that almost derailed 00:08:19.079 --> 00:08:22.160 the entire field of sequential AI. Let's break 00:08:22.160 --> 00:08:24.139 this down because Vanishing Gradient sounds like, 00:08:24.139 --> 00:08:26.459 I don't know, a bad sci -fi movie title. To start, 00:08:26.560 --> 00:08:28.660 what actually is the gradient in this context? 00:08:28.800 --> 00:08:30.560 Well, to understand the gradient, you have to 00:08:30.560 --> 00:08:32.899 look at how the network learns through a process 00:08:32.899 --> 00:08:35.659 called backpropagation through time. Let's say 00:08:35.659 --> 00:08:38.679 the network reads a sequence of 100 words and 00:08:38.679 --> 00:08:40.519 then makes a prediction about the next word. 00:08:40.830 --> 00:08:43.529 If the prediction is wrong, it calculates an 00:08:43.529 --> 00:08:45.990 error signal. Makes sense. It then has to go 00:08:45.990 --> 00:08:48.710 backward through time, step by step, through 00:08:48.710 --> 00:08:51.409 the hidden states to figure out which previous 00:08:51.409 --> 00:08:54.470 neuron made the mistake and adjust its mathematical 00:08:54.470 --> 00:08:56.990 weight. That error signal traveling backward, 00:08:57.230 --> 00:08:59.629 that is the gradient. Oh, I see. It's like tracking 00:08:59.629 --> 00:09:03.049 down who started a rumor by going backward through 00:09:03.049 --> 00:09:05.909 a chain of a hundred gossiping people, tweaking 00:09:05.909 --> 00:09:08.659 each person's reliability score as you go. That 00:09:08.659 --> 00:09:10.980 captures the sequence perfectly. But here is 00:09:10.980 --> 00:09:13.580 the specific mathematical wall they hit. Every 00:09:13.580 --> 00:09:15.480 time you take a step backward through time to 00:09:15.480 --> 00:09:17.519 adjust a weight, you use calculus. I'm right. 00:09:17.559 --> 00:09:20.000 The chain rule. Exactly. Which means you are 00:09:20.000 --> 00:09:22.639 multiplying that error signal by a specific number. 00:09:23.139 --> 00:09:25.419 And because of how the network's activation functions 00:09:25.419 --> 00:09:27.879 work, that number is almost always a fraction. 00:09:28.179 --> 00:09:31.120 It's less than one. Oh. And if you multiply a 00:09:31.120 --> 00:09:33.080 fraction by a fraction by a fraction, The number 00:09:33.080 --> 00:09:37.820 shrinks rapidly. If you multiply 0 .9 by itself 00:09:37.820 --> 00:09:40.740 dozens or hundreds of times as you go back through 00:09:40.740 --> 00:09:44.480 a long sequence, the value vanishes exponentially 00:09:44.480 --> 00:09:47.120 quickly. So it just disappears. Right. By the 00:09:47.120 --> 00:09:49.019 time the error signal reaches the beginning of 00:09:49.019 --> 00:09:51.960 the sequence, it's essentially zero. The network 00:09:51.960 --> 00:09:54.639 mathematically cannot learn from events that 00:09:54.639 --> 00:09:57.100 happened too many steps ago. It remembers the 00:09:57.100 --> 00:09:59.779 end of the sentence, but the beginning is effectively 00:09:59.779 --> 00:10:02.519 erased from its learning process. Yes. Okay, 00:10:02.779 --> 00:10:05.500 playing devil's advocate here. If it has a memory 00:10:05.500 --> 00:10:08.080 problem, why not just give it a bigger hard drive? 00:10:08.480 --> 00:10:10.740 Just throw more modern computing power at it, 00:10:10.940 --> 00:10:13.240 make the memory banks larger. Because it's not 00:10:13.240 --> 00:10:16.120 a storage capacity issue, it's a signal degradation 00:10:16.120 --> 00:10:19.230 issue. Oh, okay. Think back to your gossip analogy 00:10:19.230 --> 00:10:21.990 or a game of telephone that stretches for a mile. 00:10:22.629 --> 00:10:24.649 It doesn't matter how big the brains of the people 00:10:24.649 --> 00:10:27.309 in the line are. By the time the whisper goes 00:10:27.309 --> 00:10:29.710 through a thousand people, the original acoustic 00:10:29.710 --> 00:10:32.529 message is degraded to noise. So the underlying 00:10:32.529 --> 00:10:34.850 architecture itself was destroying the signal. 00:10:35.149 --> 00:10:38.470 So how do you fix a game of telephone where the 00:10:38.470 --> 00:10:41.549 signal mathematically degrades? Because, obviously, 00:10:41.889 --> 00:10:43.909 they figured it out, or we wouldn't be using 00:10:43.909 --> 00:10:46.970 these systems today. They solved it with a brilliantly 00:10:46.970 --> 00:10:50.350 counterintuitive approach. In 1997, researchers 00:10:50.350 --> 00:10:53.210 Sepp Hochreiter and Jürgen Schmidhuber invented 00:10:53.210 --> 00:10:56.409 the LSTM. Right. And that stands for long short 00:10:56.409 --> 00:10:58.450 -term memory. Long short -term memory. Kind of 00:10:58.450 --> 00:11:01.690 an oxymoron, but okay. How did this architecture 00:11:01.690 --> 00:11:04.850 fix the vanishing signal? By giving the network 00:11:04.850 --> 00:11:07.509 the explicit ability to intentionally forget. 00:11:07.610 --> 00:11:10.600 Wait. The solution to a memory problem was to 00:11:10.600 --> 00:11:12.679 introduce forgetting. I know, it sounds crazy, 00:11:12.759 --> 00:11:15.320 but an LSTM unit contains internal mechanisms 00:11:15.320 --> 00:11:18.179 called gates. And the most crucial one is the 00:11:18.179 --> 00:11:20.480 forget gate. Okay, what does it do? Instead of 00:11:20.480 --> 00:11:22.879 blindly cramming every single new piece of data 00:11:22.879 --> 00:11:25.320 into the hidden state and letting the old important 00:11:25.320 --> 00:11:28.240 stuff get diluted by the math, the network actively 00:11:28.240 --> 00:11:30.659 decides what is useless and mathematically throws 00:11:30.659 --> 00:11:32.460 it away. How does it actually throw it away, 00:11:32.600 --> 00:11:34.600 mathematically speaking? It uses something called 00:11:34.600 --> 00:11:36.980 a sigmoid function, which basically acts like 00:11:36.980 --> 00:11:40.360 a valve. It takes the incoming data and squishes 00:11:40.360 --> 00:11:43.240 it into a value between zero and one. Okay. So 00:11:43.240 --> 00:11:46.519 if the network decides a word like and or the 00:11:46.519 --> 00:11:49.240 isn't important for the overall context, the 00:11:49.240 --> 00:11:52.600 forget gate outputs a zero. If you multiply any 00:11:52.600 --> 00:11:55.759 piece of memory by zero, it is instantly erased. 00:11:56.000 --> 00:11:59.379 Boom. Gone. Exactly. If it's important, it outputs 00:11:59.379 --> 00:12:02.100 a one and the memory is preserved perfectly. 00:12:02.240 --> 00:12:04.279 Oh, I see. It's like a central conveyor belt 00:12:04.279 --> 00:12:06.389 running through a factory. Right. Instead of 00:12:06.389 --> 00:12:08.070 dumping every new piece of trash under the belt 00:12:08.070 --> 00:12:10.889 and causing a massive jam, the forget gates act 00:12:10.889 --> 00:12:14.009 as robotic arms, actively sweeping useless items 00:12:14.009 --> 00:12:16.370 off the belt. This ensures the core components 00:12:16.370 --> 00:12:18.389 can travel down the line completely undisturbed. 00:12:18.509 --> 00:12:20.509 That is a much more accurate way to visualize 00:12:20.509 --> 00:12:23.450 it. By carefully regulating what flows onto the 00:12:23.450 --> 00:12:26.070 conveyor belt and what gets swept off, the LSTM 00:12:26.070 --> 00:12:28.769 creates a protected central channel. A protected 00:12:28.769 --> 00:12:31.129 channel. Yes, what they call a constant error 00:12:31.129 --> 00:12:33.769 carousel. In this channel, the gradient doesn't 00:12:33.769 --> 00:12:36.649 vanish. The error signal can flow backward through 00:12:36.649 --> 00:12:39.269 thousands, even millions of time steps without 00:12:39.269 --> 00:12:42.230 losing its potency. Wow. And the source also 00:12:42.230 --> 00:12:45.789 mentions GRUs, gated recurrent units. Where do 00:12:45.789 --> 00:12:48.340 they fit into the conveyor belt analogy? Introduced 00:12:48.340 --> 00:12:53.019 much later, in 2014, a GRU is basically a streamlined, 00:12:53.100 --> 00:12:55.860 more efficient version of the LSTM. Like a model 00:12:55.860 --> 00:12:58.019 upgrade? Yeah, it combines a couple of the gates 00:12:58.019 --> 00:13:00.799 to compute faster, but it performs almost identically 00:13:00.799 --> 00:13:03.919 well on most tasks. But the LSTM, that was the 00:13:03.919 --> 00:13:06.980 true paradigm shift. It solved the great amnesia. 00:13:07.259 --> 00:13:09.539 So they cured the memory problem. They built 00:13:09.539 --> 00:13:12.039 a network that can remember the past and prioritize 00:13:12.039 --> 00:13:15.100 the important data. How did researchers actually 00:13:15.100 --> 00:13:16.940 start putting these building blocks together 00:13:16.940 --> 00:13:18.879 to do useful things? Well, this is where it gets 00:13:18.879 --> 00:13:21.259 really creative. Because reading through the 00:13:21.259 --> 00:13:23.840 configurations in the source, the architecture 00:13:23.840 --> 00:13:26.759 gets wild pretty fast. It does. Once you have 00:13:26.759 --> 00:13:29.100 a reliable memory cell like an LSTM, you can 00:13:29.100 --> 00:13:31.120 start stacking them and arranging them in fascinating 00:13:31.120 --> 00:13:33.720 ways. One of the most powerful configurations 00:13:33.720 --> 00:13:37.899 is the bi -directional RNN. processing data forwards 00:13:37.899 --> 00:13:40.039 and backwards at the same time. I remember reading 00:13:40.039 --> 00:13:42.200 this and thinking, why would you ever read a 00:13:42.200 --> 00:13:43.659 sentence backward? Why do you think? Well, if 00:13:43.659 --> 00:13:48.120 I say the bank, you don't know if I mean a financial 00:13:48.120 --> 00:13:50.419 institution or the side of a river until I finish 00:13:50.419 --> 00:13:53.320 the sentence with was muddy from the rain. Exactly. 00:13:53.779 --> 00:13:56.960 A standard RNN moving strictly from left to right 00:13:56.960 --> 00:14:00.059 has to guess what bank means before it sees the 00:14:00.059 --> 00:14:04.220 word muddy. But a bi -directional RNN uses two 00:14:04.220 --> 00:14:06.659 separate networks. Right. One reads left to right, 00:14:06.740 --> 00:14:08.659 the other reads right to left. Hold on, let me 00:14:08.659 --> 00:14:11.259 push back on that. If the backward reading network 00:14:11.259 --> 00:14:14.000 has to see the end of the sentence to understand 00:14:14.000 --> 00:14:16.700 the beginning, how can that possibly work in 00:14:16.700 --> 00:14:18.980 real time? It can't. Because if I was speaking 00:14:18.980 --> 00:14:21.240 to a voice assistant, it doesn't know the end 00:14:21.240 --> 00:14:23.659 of my sentence until I actually say it. Isn't 00:14:23.659 --> 00:14:26.000 reading backward cheating? That's a crucial distinction. 00:14:26.570 --> 00:14:29.490 Bidirectional RNNs are not used for real -time, 00:14:29.529 --> 00:14:31.269 instantaneous prediction. You are absolutely 00:14:31.269 --> 00:14:32.990 right. They need the whole sequence available 00:14:32.990 --> 00:14:35.629 beforehand. They are used for offline processing, 00:14:36.169 --> 00:14:38.350 like translating a document you've already uploaded 00:14:38.350 --> 00:14:41.090 or analyzing the sentiment of a completely written 00:14:41.090 --> 00:14:43.669 review. The two networks meet in the middle, 00:14:44.070 --> 00:14:46.850 combine their hidden states, and the model understands 00:14:46.850 --> 00:14:49.889 the ambiguous word perfectly within its complete, 00:14:50.129 --> 00:14:52.769 finalized context. That makes perfect sense. 00:14:52.769 --> 00:14:54.429 It only works when the data is already sitting 00:14:54.429 --> 00:14:56.480 there. Yeah. Then there's the encoder -decoder 00:14:56.480 --> 00:14:59.200 configuration, which the source notes was the 00:14:59.200 --> 00:15:01.419 secret sauce for early neural machine translation. 00:15:01.580 --> 00:15:03.740 Yes, like translating French to English. Right. 00:15:03.840 --> 00:15:07.080 This was a massive leap in the early 2010s. An 00:15:07.080 --> 00:15:10.740 encoder -decoder uses two separate RNNs. The 00:15:10.740 --> 00:15:13.600 first one, the encoder, reads the entire French 00:15:13.600 --> 00:15:16.860 sentence. OK. It processes it step by step until 00:15:16.860 --> 00:15:18.980 it compresses the meaning of that entire sentence 00:15:18.980 --> 00:15:21.940 into a single, highly dense mathematical vector. 00:15:22.379 --> 00:15:25.509 It's essentially a pure language concept of the 00:15:25.509 --> 00:15:27.789 sentence. It digests the French into a pure thought 00:15:27.789 --> 00:15:31.370 vector. Beautifully put. Then it hands that dense 00:15:31.370 --> 00:15:34.090 concept over to the second RNN, the decoder. 00:15:34.149 --> 00:15:36.450 And what does the decoder do? The decoder takes 00:15:36.450 --> 00:15:39.389 that concept and starts unfolding it, generating 00:15:39.389 --> 00:15:42.009 the English sentence word by word based on that 00:15:42.009 --> 00:15:44.830 pure thought. It doesn't translate word for word, 00:15:45.210 --> 00:15:47.710 it translates the entire idea. And we interact 00:15:47.710 --> 00:15:49.799 with these architectures constantly. The source 00:15:49.799 --> 00:15:52.899 points out that these LSTM networks powered Google 00:15:52.899 --> 00:15:55.559 Voice Search and dictation on Android devices 00:15:55.559 --> 00:15:58.330 for years. They were the industry standard. But 00:15:58.330 --> 00:16:00.070 some of the other applications listed in this 00:16:00.070 --> 00:16:02.490 Wikipedia article prove it's not just about text. 00:16:02.610 --> 00:16:05.169 Not at all. Sequence data is everywhere in the 00:16:05.169 --> 00:16:07.649 physical world. Right. For example, they used 00:16:07.649 --> 00:16:10.850 LSTMs to learn the long -term structure of blues 00:16:10.850 --> 00:16:14.110 music. Yes, to compose original blues songs. 00:16:14.330 --> 00:16:16.830 Because music at its core is just a sequence 00:16:16.830 --> 00:16:19.669 of notes over time. Exactly. And it scales up 00:16:19.669 --> 00:16:21.809 to incredibly high stakes physical environments 00:16:21.809 --> 00:16:25.169 as well. The article mentions an RNN system that 00:16:25.169 --> 00:16:28.009 learned how to tie knots for robotic heart surgery. 00:16:28.149 --> 00:16:30.690 Wow. But it makes sense. Tying a knot is entirely 00:16:30.690 --> 00:16:33.389 sequential. You can't skip step two and somehow 00:16:33.389 --> 00:16:36.070 arrive at step three. The precise order of movements 00:16:36.070 --> 00:16:38.370 is critical. Furthermore, they are being used 00:16:38.370 --> 00:16:40.809 to predict fusion plasma disruptions in nuclear 00:16:40.809 --> 00:16:44.309 reactors. Wait, really? Nuclear reactors? Yes. 00:16:44.529 --> 00:16:47.149 A nuclear reactor state is a continuous time 00:16:47.149 --> 00:16:50.809 series of sensor data. And RNN can analyze that 00:16:50.809 --> 00:16:53.169 ongoing sequence and predict a highly dangerous 00:16:53.169 --> 00:16:55.889 disruption before it actually occurs. That is 00:16:55.889 --> 00:16:58.610 incredible and maybe the most futuristic application 00:16:58.610 --> 00:17:02.549 listed. Decoding speech directly from the brain 00:17:02.549 --> 00:17:05.470 of a paralyzed person. A truly profound application. 00:17:05.650 --> 00:17:07.730 They take the sequential firing of neurons in 00:17:07.730 --> 00:17:10.470 the brain, feed that biological sequence into 00:17:10.470 --> 00:17:13.369 an RNN, and it outputs the sequence of words 00:17:13.369 --> 00:17:16.349 the person is trying to say. That is just staggering. 00:17:16.690 --> 00:17:19.230 It highlights the raw power of sequential processing. 00:17:19.549 --> 00:17:22.190 But taking a step back, training a system to 00:17:22.190 --> 00:17:24.849 do something that complex, without it cascading 00:17:24.849 --> 00:17:27.789 into a mess of compounding errors, requires some 00:17:27.789 --> 00:17:29.710 clever engineering tricks during the learning 00:17:29.710 --> 00:17:32.410 phase. Which naturally leads us to teacher forcing. 00:17:32.569 --> 00:17:35.430 Yes. When an RNN loops its own output back into 00:17:35.430 --> 00:17:37.769 itself, what happens if it hallucinates the wrong 00:17:37.769 --> 00:17:39.710 word at the very start of the sentence? Doesn't 00:17:39.710 --> 00:17:41.630 it just poison its own memory pool for the rest 00:17:41.630 --> 00:17:43.970 of the sequence? That is exactly the danger. 00:17:44.250 --> 00:17:46.990 If the decoder in a translation task generates 00:17:46.990 --> 00:17:49.799 the wrong second word, it's This hidden state 00:17:49.799 --> 00:17:52.319 is now corrupted by that error. So the whole 00:17:52.319 --> 00:17:55.920 thing is ruined? Basically. Every word it generates 00:17:55.920 --> 00:17:58.339 after that will likely be wrong, and the learning 00:17:58.339 --> 00:18:00.299 signal becomes useless because the whole sentence 00:18:00.299 --> 00:18:02.579 is garbage. Here's my analogy for this training 00:18:02.579 --> 00:18:05.400 method. Imagine you're teaching a kid to spell 00:18:05.400 --> 00:18:08.670 a 10 -letter word, like dictionary. OK. If they 00:18:08.670 --> 00:18:10.809 get the second letter wrong, they write a U instead 00:18:10.809 --> 00:18:13.809 of an I. If you just let them keep going, the 00:18:13.809 --> 00:18:15.950 next eight letters are going to be a confused 00:18:15.950 --> 00:18:17.990 mess because they are trying to build up a fundamental 00:18:17.990 --> 00:18:20.609 mistake. Right. The errors compound exponentially. 00:18:20.789 --> 00:18:23.109 So teacher forcing is the teacher stepping in 00:18:23.109 --> 00:18:26.589 immediately. The kid writes U. The teacher intervenes 00:18:26.589 --> 00:18:29.589 and says, no, it's an I. Now try the third letter 00:18:29.589 --> 00:18:31.970 based on the correct I. That's a great way to 00:18:31.970 --> 00:18:34.250 explain it. During training, the network uses 00:18:34.250 --> 00:18:37.130 the actual correct previous word from the training 00:18:37.130 --> 00:18:40.630 data to predict the next word, rather than relying 00:18:40.630 --> 00:18:43.690 on its own potentially flawed guesses. It keeps 00:18:43.690 --> 00:18:46.390 the training process firmly on the rails. It 00:18:46.390 --> 00:18:48.890 allows the network to learn the true statistical 00:18:48.890 --> 00:18:51.529 relationship between the second and third word, 00:18:51.930 --> 00:18:53.910 rather than getting lost in the weeds of its 00:18:53.910 --> 00:18:56.369 own early hallucination. OK, so we've mapped 00:18:56.369 --> 00:18:59.390 out this incredible looping brain. We've seen 00:18:59.390 --> 00:19:03.049 how LSTMs cured its amnesia. We've explored how 00:19:03.049 --> 00:19:05.569 it translates languages and ties surgical knots. 00:19:05.569 --> 00:19:07.890 You've covered a lot. But if we've solved all 00:19:07.890 --> 00:19:10.650 these problems, why are we suddenly hearing about 00:19:10.650 --> 00:19:13.650 a completely different technology? Anyone following 00:19:13.650 --> 00:19:16.910 AI recently knows that large language models 00:19:16.910 --> 00:19:19.230 are completely dominating the space, and they 00:19:19.230 --> 00:19:21.990 aren't using RNNs. If we connect this to the 00:19:21.990 --> 00:19:25.029 bigger picture of AI today, the natural language 00:19:25.029 --> 00:19:27.369 processing field has been overtaken by a different 00:19:27.369 --> 00:19:29.769 architecture called transformers. The source 00:19:29.769 --> 00:19:32.569 is very clear about this shift. Transformers 00:19:32.569 --> 00:19:34.910 don't use recurrence, they don't loop. No, they 00:19:34.910 --> 00:19:37.049 don't. They use a mechanism called self -attention. 00:19:37.470 --> 00:19:40.789 To explain the difference, an RNN processes a 00:19:40.789 --> 00:19:43.230 sequence one word at a time, updating its hidden 00:19:43.230 --> 00:19:46.200 state. A transformer's self -attention mechanism 00:19:46.200 --> 00:19:48.960 works by looking at the entire sequence simultaneously. 00:19:49.160 --> 00:19:52.480 All at once. Yes. It mathematically weights the 00:19:52.480 --> 00:19:55.099 importance of every single word against every 00:19:55.099 --> 00:19:57.259 other word in the document all at once. So it 00:19:57.259 --> 00:19:59.000 doesn't have to step through it chronologically. 00:19:59.259 --> 00:20:02.099 Exactly. And because they don't process sequentially, 00:20:02.599 --> 00:20:04.779 transformers are much easier to compute in parallel 00:20:04.779 --> 00:20:08.640 on massive arrays of graphics cards. Plus, they 00:20:08.640 --> 00:20:11.279 handle incredibly long -range dependencies even 00:20:11.279 --> 00:20:14.410 better than the conveyor belts of L .A. So, serious 00:20:14.410 --> 00:20:18.009 question for you. If transformers exist and they 00:20:18.009 --> 00:20:20.849 are dominating text and language right now with 00:20:20.849 --> 00:20:24.180 massive parallel processing, Why should anyone 00:20:24.180 --> 00:20:26.700 listening to this care about RNNs? Are they just 00:20:26.700 --> 00:20:28.640 a stepping stone in computer science history? 00:20:28.740 --> 00:20:31.619 Oh, they are far from obsolete. Right. Yes. Transformers 00:20:31.619 --> 00:20:33.700 are incredibly powerful, but that look at everything 00:20:33.700 --> 00:20:36.160 at once. Self -attention mechanism is extremely 00:20:36.160 --> 00:20:38.819 computationally expensive. It requires massive 00:20:38.819 --> 00:20:41.259 amounts of memory and processing power. Right. 00:20:41.579 --> 00:20:44.559 RNNs, by contrast, are highly computationally 00:20:44.559 --> 00:20:46.180 efficient. Because they only have to look at 00:20:46.180 --> 00:20:48.779 one step at a time, carrying that single hidden 00:20:48.779 --> 00:20:51.000 state forward, rather than holding the entire 00:20:51.000 --> 00:20:54.160 document in active memory. Precisely. So if you 00:20:54.160 --> 00:20:57.339 are running an AI on an edge device, like a smartwatch 00:20:57.339 --> 00:21:00.059 or a tiny environmental sensor, where you have 00:21:00.059 --> 00:21:02.059 strictly limited battery and limited memory, 00:21:02.579 --> 00:21:04.859 an RNN is often much more practical. That makes 00:21:04.859 --> 00:21:07.299 sense. Furthermore, RNNs remain highly relevant 00:21:07.299 --> 00:21:09.880 for continuous data that is inherently sequential 00:21:09.880 --> 00:21:12.160 and endless. Like the nuclear reactor sensors 00:21:12.160 --> 00:21:15.000 we mentioned. Yes. Or stock market data. Continuous 00:21:15.000 --> 00:21:18.380 time series forecasting. Exactly. A transformer 00:21:18.380 --> 00:21:21.819 needs a defined window of data, say a thousand 00:21:21.829 --> 00:21:25.390 context window to look at all at once. And RNN 00:21:25.390 --> 00:21:28.309 can just sit there processing a continuous never 00:21:28.309 --> 00:21:31.549 -ending stream of data tick by tick, updating 00:21:31.549 --> 00:21:33.950 its internal state indefinitely without running 00:21:33.950 --> 00:21:36.589 out of memory. They are fundamentally built for 00:21:36.589 --> 00:21:38.549 the flow of time. So they aren't going anywhere. 00:21:38.650 --> 00:21:40.630 We just have a specialized tool for sequential 00:21:40.630 --> 00:21:42.829 data. That's the best way to view it. Well, this 00:21:42.829 --> 00:21:44.609 has been an incredible journey through the source 00:21:44.609 --> 00:21:47.809 material. We started with Santiago Ramon y Cajal, 00:21:48.089 --> 00:21:50.430 looking at the biological loops in brain tissue 00:21:50.430 --> 00:21:53.930 in 1901. A long time ago. We took a detour through 00:21:53.930 --> 00:21:56.490 the physics of magnetic atoms settling into low 00:21:56.490 --> 00:21:59.069 energy states. We confronted the mathematical 00:21:59.069 --> 00:22:01.789 wall of the vanishing gradient and saw how the 00:22:01.789 --> 00:22:04.829 robotic sorting arms of the LSTM, forget gates, 00:22:05.289 --> 00:22:07.710 saved the day. That's a quiet story. And we explored 00:22:07.710 --> 00:22:10.029 how these time traveling networks paved the way 00:22:10.029 --> 00:22:12.900 for modern applications. from real -time translation 00:22:12.900 --> 00:22:15.890 to robotic surgery. It really is a testament 00:22:15.890 --> 00:22:18.809 to how many different fields of human inquiry, 00:22:19.109 --> 00:22:21.410 neuroscience, statistical mechanics, computer 00:22:21.410 --> 00:22:24.069 science, had to seamlessly merge to make this 00:22:24.069 --> 00:22:26.750 technology work. It is. And for you listening, 00:22:27.190 --> 00:22:29.269 I hope the next time you hit the microphone button 00:22:29.269 --> 00:22:31.750 on your phone and dictate a text message, you 00:22:31.750 --> 00:22:33.890 take a second to appreciate the mechanics behind 00:22:33.890 --> 00:22:36.910 it. Definitely. Because you are relying on decades 00:22:36.910 --> 00:22:39.569 of physicists, neuroscientists, and engineers 00:22:39.569 --> 00:22:42.130 who figured out how to mathematically teach a 00:22:42.130 --> 00:22:45.119 machine the human cons... of time and memory. 00:22:45.359 --> 00:22:47.519 And that actually brings up a fascinating final 00:22:47.519 --> 00:22:49.779 thought to consider. Oh, yeah. Think about the 00:22:49.779 --> 00:22:51.799 fundamental difference we just discussed between 00:22:51.799 --> 00:22:55.599 an RNN and the newer transformers. An RNN processes 00:22:55.599 --> 00:22:59.259 information strictly sequentially. It moves from 00:22:59.259 --> 00:23:02.000 the past into the present, carrying its memories 00:23:02.000 --> 00:23:04.900 with it, step by step. Just like human conscious 00:23:04.900 --> 00:23:07.660 experience, we can only experience the present 00:23:07.660 --> 00:23:10.240 moment informed by the past. Right. It literally 00:23:10.240 --> 00:23:13.230 experiences time. but a transformer looks at 00:23:13.230 --> 00:23:16.150 all the data at once. It sits outside the flow 00:23:16.150 --> 00:23:18.309 of time, looking at the beginning, middle, and 00:23:18.309 --> 00:23:21.329 end of a sequence simultaneously. As we continue 00:23:21.329 --> 00:23:23.430 to build more advanced artificial intelligence, 00:23:23.650 --> 00:23:26.250 we are shifting from models that process reality, 00:23:26.789 --> 00:23:29.809 like our biological time -bound brains, toward 00:23:29.809 --> 00:23:32.410 models that process information like omnipresent 00:23:32.410 --> 00:23:35.410 observers. Wow. It is profound. It makes you 00:23:35.410 --> 00:23:38.029 wonder, what does that mean for the kind of intelligence 00:23:38.029 --> 00:23:40.569 we are actually creating? If it doesn't experience 00:23:40.569 --> 00:23:42.890 the linear flow of time, is its understanding 00:23:42.890 --> 00:23:45.869 of reality fundamentally alien to our own? It's 00:23:45.869 --> 00:23:47.650 certainly a long way from the amnesiac AI we 00:23:47.650 --> 00:23:50.349 started with, trapped in a single isolated moment. 00:23:51.049 --> 00:23:52.849 Something to chew on the next time you ask a 00:23:52.849 --> 00:23:54.990 chat bot a question. Thank you for joining us 00:23:54.990 --> 00:23:56.809 on this deep dive. We'll see you next time.