WEBVTT 00:00:00.000 --> 00:00:04.519 Welcome to The Debate. Today we're dissecting 00:00:04.519 --> 00:00:06.759 the machinery of the modern mind, or maybe a 00:00:06.759 --> 00:00:09.339 better way to put it is the digital mirror we've 00:00:09.339 --> 00:00:11.679 built to reflect it. We're talking about the 00:00:11.679 --> 00:00:14.820 neural network. Right. And it's a term that has 00:00:14.820 --> 00:00:17.739 become, you know, completely synonymous with 00:00:17.739 --> 00:00:20.039 artificial intelligence, yet it carries this 00:00:20.039 --> 00:00:24.100 biological passport. It suggests that we've successfully 00:00:24.100 --> 00:00:26.719 reverse -engineered the human brain. But when 00:00:26.719 --> 00:00:28.559 you peel back the layers of code, what are you 00:00:28.559 --> 00:00:31.379 actually looking at? a digital cortex, or just 00:00:31.379 --> 00:00:34.200 advanced statistics wearing a biological costume. 00:00:34.460 --> 00:00:37.000 And that is the central tension we're navigating 00:00:37.000 --> 00:00:39.719 today. The neural network is the architecture 00:00:39.719 --> 00:00:43.560 underlying everything from, well, chat GPT to 00:00:43.560 --> 00:00:46.340 self -driving cars. But its origins, and I would 00:00:46.340 --> 00:00:48.799 argue its fundamental nature, are rooted in neuroscience. 00:00:49.299 --> 00:00:52.020 The question is, have we built a model that functions 00:00:52.020 --> 00:00:54.530 like a brain? Or have we simply borrowed the 00:00:54.530 --> 00:00:56.649 terminology to brand a mathematical calculator? 00:00:57.030 --> 00:00:59.689 I'm taking the position that the neural and neural 00:00:59.689 --> 00:01:03.630 network is really a relic of history, not a description 00:01:03.630 --> 00:01:05.969 of function. We're dealing with mathematical 00:01:05.969 --> 00:01:08.849 models designed to approximate nonlinear functions. 00:01:09.250 --> 00:01:11.989 These are tools of engineering, not biology. 00:01:12.650 --> 00:01:15.269 And I'm here to argue that the biological comparison 00:01:15.269 --> 00:01:18.170 is not just a metaphor. It is the blueprint. 00:01:18.450 --> 00:01:20.810 From the architecture of interconnected units 00:01:20.810 --> 00:01:23.670 to the very way they learn from error, artificial 00:01:23.670 --> 00:01:26.689 neural networks are the first technology to successfully 00:01:26.689 --> 00:01:29.450 implement the theories of the mind proposed by 00:01:29.450 --> 00:01:32.189 19th century psychologists. We haven't just built 00:01:32.189 --> 00:01:34.549 a calculator. We have digitized the fundamental 00:01:34.549 --> 00:01:37.390 principles of thought. That is a bold claim. 00:01:38.069 --> 00:01:42.379 Let's verify it. To understand my stance, you 00:01:42.379 --> 00:01:44.700 have to look at where this concept came from. 00:01:45.060 --> 00:01:47.939 We tend to think of AI as a product of the Silicon 00:01:47.939 --> 00:01:52.099 Age, but the theoretical framework actually predates 00:01:52.099 --> 00:01:56.079 the computer. We're going back to 1873, to Alexander 00:01:56.079 --> 00:02:00.040 Bain and later William James in 1890. The fathers 00:02:00.040 --> 00:02:03.040 of modern psychology. Exactly. And what was their 00:02:03.040 --> 00:02:06.040 radical proposition? They suggested that thought 00:02:06.040 --> 00:02:09.599 or consciousness wasn't some singular, indivisible 00:02:09.599 --> 00:02:11.919 thing. They proposed that it was an emergent 00:02:11.919 --> 00:02:14.699 property, something that arises from the interactions 00:02:14.699 --> 00:02:17.539 among a massive number of neurons. This is the 00:02:17.539 --> 00:02:20.199 definition of a neural network, a group of interconnected 00:02:20.199 --> 00:02:22.939 units where the intelligence isn't in the unit, 00:02:23.000 --> 00:02:26.060 but in the connection. I don't dispute the historical 00:02:26.060 --> 00:02:28.919 lineage. I mean, Bain and James were brilliant 00:02:28.919 --> 00:02:31.560 in hypothesizing how the biological brain might 00:02:31.560 --> 00:02:35.990 work. And yes, in 1943... Warren McCulloch and 00:02:35.990 --> 00:02:39.069 Walter Pitts used those ideas to create the first 00:02:39.069 --> 00:02:42.189 computational model of a neural network. They 00:02:42.189 --> 00:02:44.650 were connectionists. They wanted to mimic the 00:02:44.650 --> 00:02:47.129 brain. And they succeeded. They demonstrated 00:02:47.129 --> 00:02:50.169 that simple units acting together could perform 00:02:50.169 --> 00:02:53.349 complex logic. But intention is not the same 00:02:53.349 --> 00:02:56.430 as execution. Just because McCulloch and Pitts 00:02:56.430 --> 00:02:59.569 wanted to build a brain doesn't mean the modern 00:02:59.569 --> 00:03:02.150 result is a brain. If you look at the definition 00:03:02.150 --> 00:03:04.509 of an artificial neural network in machine learning 00:03:04.509 --> 00:03:08.770 today, the biology is notably absent. It's defined 00:03:08.770 --> 00:03:12.349 as a mathematical model used to approximate nonlinear 00:03:12.349 --> 00:03:16.110 functions. Let's pause on that term, nonlinear 00:03:16.110 --> 00:03:18.270 functions, because I think it scares people off, 00:03:18.330 --> 00:03:20.750 but it's actually the strongest argument for 00:03:20.750 --> 00:03:23.770 the biological connection. How so? A linear world 00:03:23.770 --> 00:03:27.710 is simple. If I push a box twice as hard, it 00:03:27.710 --> 00:03:31.189 moves twice as far. That's linear. But the real 00:03:31.189 --> 00:03:34.490 world, and biological decision -making, is non 00:03:34.490 --> 00:03:36.889 -linear. You can tap a person on the shoulder 00:03:36.889 --> 00:03:39.469 and get no reaction. Tap them slightly harder, 00:03:39.629 --> 00:03:42.090 they turn around. Tap them a tiny bit harder 00:03:42.090 --> 00:03:44.889 than that, and they might punch you. The output 00:03:44.889 --> 00:03:48.030 isn't directly proportional to the input. The 00:03:48.030 --> 00:03:51.110 brain is a non -linear processor. Artificial 00:03:51.110 --> 00:03:53.610 neural networks are the only mathematical tools 00:03:53.610 --> 00:03:55.969 that handle this non -linearity the same way 00:03:55.969 --> 00:03:59.479 biology does, through thresholds. I think you're 00:03:59.479 --> 00:04:02.580 conflating the what with the how. Yes, both systems 00:04:02.580 --> 00:04:06.180 handle complex, non -linear tasks. But the mechanism 00:04:06.180 --> 00:04:09.199 matters. In engineering, we say form follows 00:04:09.199 --> 00:04:13.039 function. But here, the form has diverged radically. 00:04:13.560 --> 00:04:16.019 The source material highlights a critical shift. 00:04:16.180 --> 00:04:19.300 Early models, like the Perceptron, built by Frank 00:04:19.300 --> 00:04:23.300 Rosenblatt in 1957, were physical hardware. They 00:04:23.300 --> 00:04:25.870 were machines you could touch. with wires acting 00:04:25.870 --> 00:04:29.389 like axons. And today, there's software. Right. 00:04:29.490 --> 00:04:32.189 And that transition to software is where the 00:04:32.189 --> 00:04:35.589 model becomes an abstraction. When you move to 00:04:35.589 --> 00:04:37.689 software, you're no longer constrained by the 00:04:37.689 --> 00:04:40.009 physics of the brain. You're optimizing for math. 00:04:40.589 --> 00:04:43.670 Modern networks are used for specific utility. 00:04:44.029 --> 00:04:46.970 Predictive modeling, facial recognition, handwriting 00:04:46.970 --> 00:04:50.720 recognition. They're statistical engines. To 00:04:50.720 --> 00:04:53.199 claim they're still biological models is like 00:04:53.199 --> 00:04:55.980 saying a submarine is a model of a fish. Yes, 00:04:56.180 --> 00:04:58.740 they both swim, but one uses a propeller and 00:04:58.740 --> 00:05:01.300 the other uses a tail. The mechanics are fundamentally 00:05:01.300 --> 00:05:03.639 different. I love the submarine analogy, but 00:05:03.639 --> 00:05:05.540 I think it fails because a submarine doesn't 00:05:05.540 --> 00:05:07.699 try to replicate the muscle structure of a fish. 00:05:07.800 --> 00:05:10.300 A neural network does replicate the information 00:05:10.300 --> 00:05:13.139 flow of a brain. Let's look at the actual transmission 00:05:13.139 --> 00:05:15.980 of data. This is the mechanics of transmission. 00:05:16.579 --> 00:05:19.300 Okay, let's get into the weeds. How do you see 00:05:19.300 --> 00:05:21.980 them as identical? In a biological brain, you 00:05:21.980 --> 00:05:25.300 have neurons. A single neuron receives signals 00:05:25.300 --> 00:05:28.100 from its neighbors through dendrites. If the 00:05:28.100 --> 00:05:30.620 total signal is strong enough, the neuron fires, 00:05:30.939 --> 00:05:33.860 sending a signal down the axon to the next neuron. 00:05:34.240 --> 00:05:36.860 Now look at the artificial version. You have 00:05:36.860 --> 00:05:39.899 nodes, or artificial neurons. They're arranged 00:05:39.899 --> 00:05:43.860 in layers. Input, hidden, and output. Information 00:05:43.860 --> 00:05:46.920 comes in, it's processed, and it's passed forward. 00:05:47.480 --> 00:05:49.500 The hidden layers are doing exactly what the 00:05:49.500 --> 00:05:52.600 mass of gray matter in your skull does, processing 00:05:52.600 --> 00:05:56.060 information in stages. Processing is doing a 00:05:56.060 --> 00:05:58.439 lot of heavy lifting in that sentence. Let's 00:05:58.439 --> 00:06:00.279 look at what is actually happening at that node. 00:06:00.579 --> 00:06:03.319 In biology, you're talking about electrochemistry. 00:06:03.639 --> 00:06:06.800 A neuron receives excitatory or inhibitory signals, 00:06:07.060 --> 00:06:10.139 chemicals like glutamate or GABA. These open 00:06:10.139 --> 00:06:12.860 ion channels. The cell membrane voltage changes. 00:06:13.139 --> 00:06:16.160 If it hits a threshold, boom, an action potential. 00:06:16.860 --> 00:06:20.139 It's a spike. It's a dynamic, temporal, physical 00:06:20.139 --> 00:06:23.319 event. And the math simulates that. The math 00:06:23.319 --> 00:06:25.680 replaces that with something totally different. 00:06:25.899 --> 00:06:28.579 The text is very clear on this. The artificial 00:06:28.579 --> 00:06:31.439 neuron calculates a linear combination of the 00:06:31.439 --> 00:06:34.060 outputs of the connected neurons. Which is a 00:06:34.060 --> 00:06:36.639 summation. It's a specific kind of summation. 00:06:36.939 --> 00:06:39.779 It takes a value x, multiplies it by a weight 00:06:39.779 --> 00:06:44.180 w, adds a bias b. It's algebra. y equals wx plus 00:06:44.180 --> 00:06:47.160 b. Then, to your point about non -linearity, 00:06:47.360 --> 00:06:50.060 it runs that number through an activation function, 00:06:50.379 --> 00:06:53.899 like a sigmoid or remove function. This squeezes 00:06:53.899 --> 00:06:56.540 the number to determine the output. You're describing 00:06:56.540 --> 00:06:59.500 the exact mathematical translation of the biological 00:06:59.500 --> 00:07:03.199 process. The weight is the strength of the synapse. 00:07:03.300 --> 00:07:07.120 The summation is the accumulation of neurotransmitters. 00:07:07.199 --> 00:07:10.019 The activation function is the firing threshold. 00:07:10.600 --> 00:07:13.399 Just because we write it in Greek letters instead 00:07:13.399 --> 00:07:16.240 of squishy tissue doesn't mean the logic is different. 00:07:16.420 --> 00:07:20.000 It is a high fidelity abstraction. I strongly 00:07:20.000 --> 00:07:24.420 disagree that it's high fidelity. It's a caricature. 00:07:24.620 --> 00:07:28.660 In the brain, timing matters. The rate of firing 00:07:28.660 --> 00:07:32.319 matters. The chemical soup matters. In the artificial 00:07:32.319 --> 00:07:35.879 network, it's just a static number passing through 00:07:35.879 --> 00:07:38.680 a static function. You're stripping away all 00:07:38.680 --> 00:07:41.920 the chaos and complexity of biology to get a 00:07:41.920 --> 00:07:45.660 clean mathematical equation. That's not a model 00:07:45.660 --> 00:07:48.279 of a brain. That's a spreadsheet that looks like 00:07:48.279 --> 00:07:50.759 a brain. But does the complexity of the substrate 00:07:50.759 --> 00:07:52.860 matter if the emergent behavior is the same? 00:07:53.000 --> 00:07:55.480 If I build a heart out of titanium and plastic 00:07:55.480 --> 00:07:58.500 and it pumps blood, it's a heart. If I build 00:07:58.500 --> 00:08:01.019 a network out of code and math and it recognizes 00:08:01.019 --> 00:08:04.180 a face, it's a neural network. But does it recognize 00:08:04.180 --> 00:08:07.480 the face the way a human does? Or does it just 00:08:07.480 --> 00:08:11.079 find a statistical correlation of pixels? This 00:08:11.079 --> 00:08:14.100 brings us to the most contentious point, learning. 00:08:14.339 --> 00:08:18.199 How does the system actually improve? This is 00:08:18.199 --> 00:08:21.180 my strongest evidence. We have to talk about 00:08:21.180 --> 00:08:26.319 Donald Hebb. Hebbian Learning, 1949. Right. Cells 00:08:26.319 --> 00:08:29.120 that fire together, wire together. This is the 00:08:29.120 --> 00:08:32.159 foundational rule of biological learning. If 00:08:32.159 --> 00:08:35.600 neuron A repeatedly helps fire neuron B, the 00:08:35.600 --> 00:08:38.379 connection between them gets stronger. The synapse 00:08:38.379 --> 00:08:41.860 physically changes. This is exactly, exactly 00:08:41.860 --> 00:08:45.120 how artificial networks learn. Explain that connection. 00:08:45.539 --> 00:08:48.179 In an artificial network, the behavior is determined 00:08:48.179 --> 00:08:51.159 by the weights, those numbers we multiply the 00:08:51.159 --> 00:08:54.059 inputs by. When we train the network, we are 00:08:54.059 --> 00:08:57.039 simply adjusting those weights. We're strengthening 00:08:57.039 --> 00:08:59.120 the connections that lead to the right answer 00:08:59.120 --> 00:09:01.440 and weakening the ones that lead to the wrong 00:09:01.440 --> 00:09:04.659 answer. That is heavy in theory and pure distilled 00:09:04.659 --> 00:09:07.419 code. It sounds like heavy in theory, but the 00:09:07.419 --> 00:09:10.500 mechanism for how those weights change is radically 00:09:10.500 --> 00:09:13.059 different. And the text points this out explicitly. 00:09:13.700 --> 00:09:16.299 Artificial networks are trained using empirical 00:09:16.299 --> 00:09:19.299 risk minimization and backpropagation. Now those 00:09:19.299 --> 00:09:21.399 are just the algorithms we use to adjust the 00:09:21.399 --> 00:09:25.879 weights? Just the algorithms. is a massive understatement. 00:09:26.700 --> 00:09:29.820 Backpropagation is the engine of modern AI, and 00:09:29.820 --> 00:09:32.519 it has no biological equivalent. Think about 00:09:32.519 --> 00:09:35.799 how it works. The network makes a guess. It calculates 00:09:35.799 --> 00:09:37.840 the error, the difference between its guess and 00:09:37.840 --> 00:09:40.840 the right answer. Then it uses calculus to go 00:09:40.840 --> 00:09:43.179 backwards through the network from output to 00:09:43.179 --> 00:09:45.559 input, calculating the gradient of the error 00:09:45.559 --> 00:09:48.340 and adjusting every single weight to minimize 00:09:48.340 --> 00:09:51.360 that error. It's an optimization technique. It's 00:09:51.360 --> 00:09:53.399 a way to find the right weights faster. It's 00:09:53.399 --> 00:09:56.379 a cheat code that biology doesn't have. Your 00:09:56.379 --> 00:09:59.559 brain cannot freeze time, calculate the mathematical 00:09:59.559 --> 00:10:02.259 error of a thought, and then send a correction 00:10:02.259 --> 00:10:04.799 signal backwards through your axons to adjust 00:10:04.799 --> 00:10:08.299 the synapses. Biological learning is local. It 00:10:08.299 --> 00:10:11.080 happens at the synapse, in real time, based on 00:10:11.080 --> 00:10:14.139 local signals. Artificial learning is global 00:10:14.139 --> 00:10:17.409 optimization based on a static dataset. I think 00:10:17.409 --> 00:10:19.110 you're getting hung up on the implementation 00:10:19.110 --> 00:10:22.850 details again. Backpropagation is simply the 00:10:22.850 --> 00:10:26.029 most efficient mathematical way to achieve the 00:10:26.029 --> 00:10:28.909 state of learning. The end result is what matters, 00:10:29.049 --> 00:10:31.789 a network where the connection strengths encode 00:10:31.789 --> 00:10:34.970 knowledge. The process dictates the capabilities. 00:10:35.509 --> 00:10:38.389 Because we use backpropagation, we need labeled 00:10:38.389 --> 00:10:41.690 data. We need a pre -existing dataset. The text 00:10:41.690 --> 00:10:45.049 says we train these networks to fit. a data set. 00:10:45.250 --> 00:10:47.769 We show it 10 ,000 pictures of cats and say, 00:10:47.889 --> 00:10:50.850 minimize the error in identifying these. That 00:10:50.850 --> 00:10:53.330 is statistical regression. That is curve fitting. 00:10:53.649 --> 00:10:56.870 But humans learn from data sets too. We call 00:10:56.870 --> 00:11:00.789 it experience. But we don't need 10 ,000 labeled 00:11:00.789 --> 00:11:04.590 examples to know what a cat is. We learn adaptively, 00:11:04.730 --> 00:11:08.129 physically, continuously. We don't perform empirical 00:11:08.129 --> 00:11:11.289 risk minimization on a static batch of data. 00:11:11.509 --> 00:11:14.730 We survive in an environment. The text makes 00:11:14.730 --> 00:11:17.710 this distinction clear. Biological networks are 00:11:17.710 --> 00:11:20.649 large -scale brain networks integrated into a 00:11:20.649 --> 00:11:23.110 nervous system that drives muscle cells in motion. 00:11:23.690 --> 00:11:26.669 Artificial networks are isolated software loops 00:11:26.669 --> 00:11:29.429 minimizing a loss function. You're acting as 00:11:29.429 --> 00:11:31.649 if these networks are stuck doing simple regression. 00:11:32.009 --> 00:11:34.409 We need to talk about the evolution of purpose. 00:11:34.690 --> 00:11:37.070 We aren't just building simple perceptrons anymore. 00:11:37.269 --> 00:11:40.830 We are building deep neural networks. Deep just 00:11:40.830 --> 00:11:43.919 means more than three layers. usually two or 00:11:43.919 --> 00:11:48.059 more hidden layers. But that depth creates emergence. 00:11:48.639 --> 00:11:52.179 The text mentions generative AI and general game 00:11:52.179 --> 00:11:55.059 playing. When you add those hidden layers, the 00:11:55.059 --> 00:11:57.299 network starts doing things that look less like 00:11:57.299 --> 00:12:00.379 statistics and more like cognition. It creates 00:12:00.379 --> 00:12:03.759 art. It writes poetry. It plays Go better than 00:12:03.759 --> 00:12:06.679 any human. It's impressive, I grant you. It connects 00:12:06.679 --> 00:12:10.139 back to a discovery by Sminyat Chakan in 1956 00:12:10.139 --> 00:12:13.259 regarding retinal cells. He found that you couldn't 00:12:13.259 --> 00:12:14.940 understand the function of the eye by looking 00:12:14.940 --> 00:12:17.299 at a single cell. You had to understand the network 00:12:17.299 --> 00:12:20.200 of horizontal cells. The interaction created 00:12:20.200 --> 00:12:22.919 the capability. That is what deep learning is 00:12:22.919 --> 00:12:25.799 doing. By stacking these layers, we're mimicking 00:12:25.799 --> 00:12:28.259 the deep hierarchical structure of the cortex. 00:12:28.559 --> 00:12:31.320 We're moving away from curve fitting toward feature 00:12:31.320 --> 00:12:33.580 extraction and representation. The constraints. 00:12:34.019 --> 00:12:37.340 Even with deep learning, The text notes the divergence 00:12:37.340 --> 00:12:40.279 from biology. These systems are often brittle. 00:12:40.399 --> 00:12:42.559 They can be fooled by noise that wouldn't fool 00:12:42.559 --> 00:12:45.419 a human. They require massive energy to train. 00:12:45.779 --> 00:12:49.000 They are approximating nonlinear functions, just 00:12:49.000 --> 00:12:51.919 extremely complex ones. I think you're underestimating 00:12:51.919 --> 00:12:54.519 the general in general game playing. A system 00:12:54.519 --> 00:12:56.600 that can learn the rules of chess, then shogi, 00:12:56.759 --> 00:12:59.480 then go, without being reprogrammed, that is 00:12:59.480 --> 00:13:01.940 approaching generalized intelligence. That isn't 00:13:01.940 --> 00:13:04.620 just a calculator. That is a malleable learning 00:13:04.620 --> 00:13:07.519 substrate. just like the brain. But it learns 00:13:07.519 --> 00:13:10.620 those games to maximize a score. It's still an 00:13:10.620 --> 00:13:13.200 optimization problem. It doesn't know it's playing 00:13:13.200 --> 00:13:15.840 a game. It doesn't have agency. The biological 00:13:15.840 --> 00:13:18.360 network is designed for survival. It's chemically 00:13:18.360 --> 00:13:21.100 connected to a body. The artificial network is 00:13:21.100 --> 00:13:23.639 designed for accuracy on a test set. What about 00:13:23.639 --> 00:13:26.139 self -driving cars? That is adaptive control, 00:13:26.299 --> 00:13:29.039 which the text mentions. That is a network navigating 00:13:29.039 --> 00:13:31.259 the physical world, making life -or -death decisions 00:13:31.259 --> 00:13:34.639 in real time. That is the closest parallel, I 00:13:34.639 --> 00:13:37.639 admit. But even there, the car is seeing numbers, 00:13:37.799 --> 00:13:40.240 not the road. It's calculating probabilities 00:13:40.240 --> 00:13:43.500 based on LIDAR point clouds. It's a simulation 00:13:43.500 --> 00:13:46.620 of perception. You keep using the word simulation 00:13:46.620 --> 00:13:50.360 as a pejorative, but all models are simulations. 00:13:51.080 --> 00:13:54.080 My argument, and the argument of the biological 00:13:54.080 --> 00:13:56.980 roots perspective, is that we have found the 00:13:56.980 --> 00:14:00.250 fundamental algorithm of intelligence. It just 00:14:00.250 --> 00:14:02.309 so happens that you can run that algorithm on 00:14:02.309 --> 00:14:06.049 wet biological tissue or on silicon chips. The 00:14:06.049 --> 00:14:09.450 weight is the universal unit of memory. The layer 00:14:09.450 --> 00:14:12.750 is the universal unit of processing. And my argument 00:14:12.750 --> 00:14:15.409 is that you've found a mathematical trick that 00:14:15.409 --> 00:14:18.309 produces results analogous to intelligence, but 00:14:18.309 --> 00:14:20.970 through a fundamentally different route. The 00:14:20.970 --> 00:14:23.289 linear combination is not an action potential. 00:14:23.730 --> 00:14:26.629 Back propagation is not heavy in plasticity. 00:14:26.990 --> 00:14:29.570 And empirical risk minimization is not survival. 00:14:29.889 --> 00:14:33.009 But we're getting closer. Every year, the networks 00:14:33.009 --> 00:14:35.629 get deeper. The architectures get more complex. 00:14:35.850 --> 00:14:38.509 We're adding attention mechanisms which function 00:14:38.509 --> 00:14:41.990 like human focus. And yet, the text reminds us 00:14:41.990 --> 00:14:44.590 that even as they get more complex, they become 00:14:44.590 --> 00:14:46.850 increasingly different from their biological 00:14:46.850 --> 00:14:49.850 counterparts. To solve the engineering problems 00:14:49.850 --> 00:14:52.889 of AI, we've had to abandon the biological constraints. 00:14:53.519 --> 00:14:55.480 we stopped trying to build a brain and started 00:14:55.480 --> 00:14:57.980 trying to build a machine that works. Maybe that's 00:14:57.980 --> 00:15:00.519 the ultimate irony. To build a machine that thinks 00:15:00.519 --> 00:15:02.879 like a human, we had to stop copying the human 00:15:02.879 --> 00:15:05.440 anatomy and start focusing on the human mathematics. 00:15:05.899 --> 00:15:08.240 Or maybe we're just seeing what we want to see. 00:15:08.340 --> 00:15:10.580 We look into the black box of a neural network 00:15:10.580 --> 00:15:13.340 and we see a reflection of our own minds, when 00:15:13.340 --> 00:15:15.820 really it's just a very shiny mirror made of 00:15:15.820 --> 00:15:18.480 calculus. That is the question we leave on the 00:15:18.480 --> 00:15:21.440 table. We've traced the arc from Bain and James' 00:15:21.639 --> 00:15:24.360 theories of the 1870s, to McCulloch and Pitts' 00:15:24.519 --> 00:15:27.379 circuits of the 1940s, to the massive deep learning 00:15:27.379 --> 00:15:30.379 models of today. A journey from biology to math, 00:15:30.519 --> 00:15:33.139 and perhaps back again. We encourage you to look 00:15:33.139 --> 00:15:35.580 at the source material and decide for yourself, 00:15:35.879 --> 00:15:39.080 is the deep neural network a sibling to the human 00:15:39.080 --> 00:15:42.019 mind, or is it simply the world's most impressive 00:15:42.019 --> 00:15:44.779 approximation machine? Thank you for joining 00:15:44.779 --> 00:15:46.779 The Debate. See you next time.