WEBVTT 00:00:00.000 --> 00:00:03.020 Imagine waking up every single morning with, 00:00:03.020 --> 00:00:07.400 like, total amnesia. Oh, wow. Sounds exhausting. 00:00:07.740 --> 00:00:09.480 Right. I mean, you have to relearn the concept 00:00:09.480 --> 00:00:11.400 of gravity before you can even get out of bed. 00:00:11.480 --> 00:00:13.980 Yeah, you'd be starting from absolute zero. Exactly. 00:00:14.359 --> 00:00:16.899 You have to relearn the physics of friction just 00:00:16.899 --> 00:00:20.280 to, you know, turn a doorknob. And eventually 00:00:20.280 --> 00:00:22.219 you figure it all out. But by the time you do, 00:00:22.320 --> 00:00:24.600 the day is over. You've accomplished basically 00:00:24.600 --> 00:00:27.519 nothing. Right. Then, the next morning, the slate 00:00:27.519 --> 00:00:30.539 is wiped completely clean. And it happens all 00:00:30.539 --> 00:00:34.039 over again. Which is, uh, it's the ultimate computational 00:00:34.039 --> 00:00:36.500 bottleneck, really. Yeah. Yeah, I mean, if you 00:00:36.500 --> 00:00:39.039 were constantly stuck at square one, rebuilding 00:00:39.039 --> 00:00:41.399 the foundational rules of reality from scratch, 00:00:41.880 --> 00:00:44.100 you never actually have the time or the processing 00:00:44.100 --> 00:00:47.359 power to learn anything new or like anything 00:00:47.359 --> 00:00:49.939 complex. And for a long time, that was the underlying 00:00:49.939 --> 00:00:52.060 reality of artificial intelligence. Every time 00:00:52.060 --> 00:00:54.119 you wanted an algorithm to do something new, 00:00:54.200 --> 00:00:56.740 you had to train it from absolute zero. It was 00:00:56.740 --> 00:00:59.750 a massive waste of resources. But today we are 00:00:59.750 --> 00:01:02.390 talking to you, the learner, the person who really 00:01:02.390 --> 00:01:06.069 wants to cut through all the marketing jargon 00:01:06.069 --> 00:01:08.730 of AI and understand the actual mechanics running 00:01:08.730 --> 00:01:11.090 under the hood. The real nuts and bolts. Exactly. 00:01:11.510 --> 00:01:14.569 So today we're doing a deep dive into a comprehensive 00:01:14.569 --> 00:01:17.489 Wikipedia article on a concept called transfer 00:01:17.489 --> 00:01:19.870 learning. It's a fascinating topic. It really 00:01:19.870 --> 00:01:23.129 is. And our mission today is to map out exactly 00:01:23.129 --> 00:01:25.989 how machine learning models finally learn to 00:01:25.989 --> 00:01:28.450 recycle their knowledge. We're going to trace 00:01:28.450 --> 00:01:31.670 it from like a theoretical geometric model in 00:01:31.670 --> 00:01:34.849 the 1970s all the way to a modern breakthrough 00:01:34.849 --> 00:01:37.709 that literally connects human brain waves directly 00:01:37.709 --> 00:01:40.129 to muscle signals. Which is just a fundamental 00:01:40.129 --> 00:01:42.390 shift in the whole architecture of machine intelligence. 00:01:42.549 --> 00:01:44.890 Huge shift. Yeah, we're moving away from these 00:01:44.890 --> 00:01:47.709 isolated single -purpose algorithms and looking 00:01:47.709 --> 00:01:49.569 at systems that can actually carry a learned 00:01:49.569 --> 00:01:52.609 perspective from one environment into a completely 00:01:52.609 --> 00:01:54.469 different one. And I'll tease you with this right 00:01:54.469 --> 00:01:57.439 off the bat. The secret to the future of AI isn't 00:01:57.439 --> 00:01:59.920 just feeding it more data. It's about a highly 00:01:59.920 --> 00:02:03.200 complex mathematical ability to reuse what it 00:02:03.200 --> 00:02:06.140 already knows. And understanding how that works 00:02:06.140 --> 00:02:08.919 is really the only way to understand why an AI 00:02:08.919 --> 00:02:11.000 might suddenly become worse at its job because 00:02:11.000 --> 00:02:13.159 of something called negative transfer. Yeah. 00:02:13.419 --> 00:02:17.259 Negative transfer is a huge hurdle. But to set 00:02:17.259 --> 00:02:21.139 the baseline here, transfer learning or TL, is 00:02:21.139 --> 00:02:23.419 essentially a technique where knowledge learned 00:02:23.419 --> 00:02:27.400 from one task is reused to boost performance 00:02:27.400 --> 00:02:30.419 on a related task. OK, so the classic foundational 00:02:30.419 --> 00:02:32.199 example they usually give is image recognition. 00:02:32.580 --> 00:02:34.219 Yeah, that's the standard textbook one. Let's 00:02:34.219 --> 00:02:36.659 say you have an AI. and you spend, I don't know, 00:02:36.840 --> 00:02:38.939 millions of dollars in computing power, teaching 00:02:38.939 --> 00:02:41.500 it to recognize images of cars. Right, so it 00:02:41.500 --> 00:02:43.460 learns edge detection, it learns shape. It also 00:02:43.460 --> 00:02:45.759 learns what wheels look like. Transfer learning 00:02:45.759 --> 00:02:48.039 means taking that foundational architecture and 00:02:48.039 --> 00:02:50.319 applying it as the starting point when you want 00:02:50.319 --> 00:02:52.539 the AI to recognize trucks. Exactly, you don't 00:02:52.539 --> 00:02:55.180 make it relearn what a tire is. You're saving 00:02:55.180 --> 00:02:58.439 massive amounts of compute. The network already 00:02:58.439 --> 00:03:01.139 has the parameters tuned to identify circular 00:03:01.139 --> 00:03:04.099 rubber objects, so it just needs to adjust slightly 00:03:04.099 --> 00:03:06.860 for, you know, the scale and context of a truck. 00:03:07.060 --> 00:03:09.180 Sure, but I have to push back a little on how 00:03:09.180 --> 00:03:12.099 revolutionary this is often framed to be. Oh. 00:03:12.340 --> 00:03:14.979 How so? Well, isn't this just learning? I mean, 00:03:15.099 --> 00:03:16.759 psychologically speaking, if I learn to ride 00:03:16.759 --> 00:03:19.120 a bicycle, I can learn to ride a motorcycle much 00:03:19.120 --> 00:03:20.719 faster than someone who has never balanced on 00:03:20.719 --> 00:03:23.020 two wheels. Right, right. I transfer my knowledge 00:03:23.020 --> 00:03:26.969 of balance. So is the AI just... mimicking human 00:03:26.969 --> 00:03:30.250 psychology here? It's a fair question. And the 00:03:30.250 --> 00:03:32.550 comparison to human psychology is incredibly 00:03:32.550 --> 00:03:36.750 common. The source material even explicitly acknowledges 00:03:36.750 --> 00:03:38.930 the psychological literature on the transfer 00:03:38.930 --> 00:03:41.370 of learning. OK, so they are related. Well, yes 00:03:41.370 --> 00:03:43.729 and no. Here is where we really have to separate 00:03:43.729 --> 00:03:46.590 the metaphor from the machine. OK. The practical 00:03:46.590 --> 00:03:49.990 structural ties between human cognitive psychology 00:03:49.990 --> 00:03:52.469 and machine learning fields are actually quite 00:03:52.469 --> 00:03:56.120 limited. Oh, really? Yeah. A human brain transfers 00:03:56.120 --> 00:03:59.659 knowledge organically, using billions of flexible 00:03:59.659 --> 00:04:02.620 synapses in ways we, frankly, still don't fully 00:04:02.620 --> 00:04:06.120 comprehend. But in machine learning, this transfer 00:04:06.120 --> 00:04:09.460 has to be rigidly mathematically engineered. 00:04:10.099 --> 00:04:12.659 Gotcha. So we aren't dealing with a subconscious 00:04:12.659 --> 00:04:15.639 feeling of balance. No, not at all. We're dealing 00:04:15.639 --> 00:04:18.079 with objective functions and data distribution. 00:04:18.300 --> 00:04:20.360 Which is why understanding the underlying math 00:04:20.360 --> 00:04:23.970 is so critical, right? Exactly. To see why a 00:04:23.970 --> 00:04:25.730 machine recycling knowledge is fundamentally 00:04:25.730 --> 00:04:28.230 different from a human doing it, we have to look 00:04:28.230 --> 00:04:30.790 at how developers artificially separate the world. 00:04:30.930 --> 00:04:32.990 Okay, let's unpack this. They divide reality 00:04:32.990 --> 00:04:36.250 into two distinct mathematical buckets, domains 00:04:36.250 --> 00:04:39.069 and tasks. Right, because looking at the equations 00:04:39.069 --> 00:04:41.769 for domains and tasks, it can feel like hitting 00:04:41.769 --> 00:04:44.149 a brick wall of academic notation. Oh, it gets 00:04:44.149 --> 00:04:46.790 very dense, very fast. Yeah, so let's break down 00:04:46.790 --> 00:04:49.649 the domain first. The notation defines a domain, 00:04:49.689 --> 00:04:52.509 which is denoted as D. as consisting of a feature 00:04:52.509 --> 00:04:55.370 space, which is X, and a marginal probability 00:04:55.370 --> 00:04:58.310 distribution, which is P of X. Right. Let's strip 00:04:58.310 --> 00:05:00.269 away the Greek letters for a second. The feature 00:05:00.269 --> 00:05:04.389 space X is simply the universe of raw inputs 00:05:04.389 --> 00:05:07.449 the AI is looking at. Like the raw data. Exactly. 00:05:07.589 --> 00:05:10.069 If it's analyzing text, the feature space is 00:05:10.069 --> 00:05:12.410 the vocabulary. If it's analyzing images, the 00:05:12.410 --> 00:05:14.490 feature space is the pixels and the colors. OK, 00:05:14.490 --> 00:05:16.949 that makes sense. And the marginal probability 00:05:16.949 --> 00:05:20.350 distribution? That is the frequency or the likelihood 00:05:20.540 --> 00:05:23.339 of those specific features appearing in that 00:05:23.339 --> 00:05:25.519 specific environment. So how often something 00:05:25.519 --> 00:05:28.459 shows up? Right. It's the AI's expectation of 00:05:28.459 --> 00:05:31.860 normal. So if your feature space is the vocabulary 00:05:31.860 --> 00:05:34.699 of a medical journal, the marginal probability 00:05:34.699 --> 00:05:37.720 of seeing the word carcinoma is quite high. Right. 00:05:37.720 --> 00:05:39.699 You'd expect to see that a lot. But if your feature 00:05:39.699 --> 00:05:41.560 space is the vocabulary of a children's book, 00:05:41.819 --> 00:05:43.980 the probability of seeing carcinoma is basically 00:05:43.980 --> 00:05:46.180 zero, but the probability of seeing the word 00:05:46.180 --> 00:05:49.660 apple is high. OK. So the domain is the environment. 00:05:49.930 --> 00:05:51.889 the raw materials and how often they show up. 00:05:52.009 --> 00:05:54.589 You got it. Then we have the second bucket. The 00:05:54.589 --> 00:05:57.550 task, which is denoted as T. Right. The task 00:05:57.550 --> 00:06:00.850 consists of a label space called Y and an objective 00:06:00.850 --> 00:06:03.170 predictive function, which is denoted as F of 00:06:03.170 --> 00:06:05.629 X. Right. So the label space is basically the 00:06:05.629 --> 00:06:07.769 list of categories the AI is allowed to sort 00:06:07.769 --> 00:06:10.750 things into. Like. Spam or not spam. Exactly. 00:06:10.910 --> 00:06:13.829 Or benign or malignant. And the objective predictive 00:06:13.829 --> 00:06:17.149 function is the actual rule the AI develops to 00:06:17.149 --> 00:06:19.589 look at the features and accurately assign those 00:06:19.589 --> 00:06:21.949 labels. Let me try to map this out with an original 00:06:21.949 --> 00:06:23.769 analogy just to make sure we aren't getting bogged 00:06:23.769 --> 00:06:25.709 down in the abstraction here. Yeah, analogies 00:06:25.709 --> 00:06:27.810 help a lot with this. Let's say we are building 00:06:27.810 --> 00:06:31.629 an AI to analyze a live sports broadcast. OK. 00:06:31.970 --> 00:06:34.889 The domain, that feature space and the probabilities, 00:06:35.230 --> 00:06:39.060 is the entire visual universe of tennis. It's 00:06:39.060 --> 00:06:41.699 the green court, the yellow ball, the white lines, 00:06:42.019 --> 00:06:44.319 and the statistical likelihood of where those 00:06:44.319 --> 00:06:46.680 pixels usually appear on the screen. That covers 00:06:46.680 --> 00:06:49.180 the domain perfectly, yes. Okay. Then the task, 00:06:49.319 --> 00:06:50.899 the label space, and the predictive function 00:06:50.899 --> 00:06:53.759 is the specific narrow job of predicting whether 00:06:53.759 --> 00:06:55.540 the yellow tennis ball is going out of bounds. 00:06:55.800 --> 00:06:58.720 Right. So the labels are simply in or out, and 00:06:58.720 --> 00:07:00.720 the predictive function is the math figuring 00:07:00.720 --> 00:07:03.420 out which label to apply based on the trajectory. 00:07:03.779 --> 00:07:06.240 That's a great way to look at it. So taking that 00:07:06.240 --> 00:07:09.170 analogy, Transfer learning is the process of 00:07:09.170 --> 00:07:11.430 using the knowledge from a source domain and 00:07:11.430 --> 00:07:14.089 source task to improve the predictive function 00:07:14.089 --> 00:07:16.949 in a completely different target domain or target 00:07:16.949 --> 00:07:20.470 task. So keeping with the sports broadcast, maybe 00:07:20.470 --> 00:07:23.990 the source domain is tennis. The AI learns the 00:07:23.990 --> 00:07:26.410 physics of a bouncing ball. It learns the swing 00:07:26.410 --> 00:07:29.629 of a racket. Then we move to a target domain, 00:07:29.970 --> 00:07:32.550 say ping pong. Right. So now the visual environment 00:07:32.550 --> 00:07:35.459 changes. The feature space shrinks. The table 00:07:35.459 --> 00:07:38.160 is blue instead of a green court. The ball is 00:07:38.160 --> 00:07:40.519 white instead of yellow. Everything looks different. 00:07:40.660 --> 00:07:42.819 And the marginal probability distribution is 00:07:42.819 --> 00:07:44.779 entirely different because the velocity and the 00:07:44.779 --> 00:07:47.399 scale have changed. But the underlying physics 00:07:47.399 --> 00:07:50.920 like gravity, spin, trajectory, those remain 00:07:50.920 --> 00:07:53.779 related. Exactly. We are transferring the predictive 00:07:53.779 --> 00:07:55.699 function across different mathematical domains. 00:07:55.879 --> 00:07:58.120 That's wild. Alternatively, you could keep the 00:07:58.120 --> 00:08:00.579 domain exactly the same. Why? Say we are still 00:08:00.579 --> 00:08:03.519 watching tennis, but change the task. OK, how 00:08:03.519 --> 00:08:05.540 so? Instead of predicting if the ball is out 00:08:05.540 --> 00:08:08.560 of bounds, the new task is classifying whether 00:08:08.560 --> 00:08:10.519 the player is going to serve to the left or the 00:08:10.519 --> 00:08:13.500 right. Oh, so the label space changes. Right. 00:08:13.959 --> 00:08:16.279 And because transfer learning navigates multiple 00:08:16.279 --> 00:08:18.860 objective functions across these differing domains 00:08:18.860 --> 00:08:22.660 and tasks, it relies heavily on multi -objective 00:08:22.660 --> 00:08:25.319 optimization and cost -sensitive machine learning. 00:08:25.579 --> 00:08:28.740 That sounds complicated. It is. The algorithm 00:08:28.740 --> 00:08:31.279 is constantly calculating the mathematical cost 00:08:31.279 --> 00:08:34.820 of unlearning an old rule versus the benefit 00:08:34.820 --> 00:08:36.840 of applying it to the new data. You know, when 00:08:36.840 --> 00:08:39.799 we talk about multi -objective optimization and 00:08:39.799 --> 00:08:43.259 shifting marginal probabilities, it sounds intensely 00:08:43.259 --> 00:08:46.039 modern. It really does. It sounds like the byproduct 00:08:46.039 --> 00:08:48.460 of the massive deep learning broom of the last 00:08:48.460 --> 00:08:51.850 decade. But the historical timeline reveals these 00:08:51.850 --> 00:08:54.169 ideas have been brewing since the era of disco. 00:08:54.490 --> 00:08:57.210 Yeah, machine learning is deeply iterative. The 00:08:57.210 --> 00:08:59.309 algorithms dominating today are standing on the 00:08:59.309 --> 00:09:01.450 shoulders of theoretical math polished decades 00:09:01.450 --> 00:09:03.850 ago. Let's trace that causal chain, because the 00:09:03.850 --> 00:09:06.389 sheer age of some of this research is just staggering 00:09:06.389 --> 00:09:08.490 to me. It is pretty surprising. The timeline 00:09:08.490 --> 00:09:12.509 notes that way back in 1976, researchers Buzinovsky 00:09:12.509 --> 00:09:15.549 and Fulgosi published a mathematical and geometrical 00:09:15.549 --> 00:09:18.070 model for transfer learning in neural network 00:09:18.070 --> 00:09:22.909 training. 1976. 1976. We're talking about an 00:09:22.909 --> 00:09:25.669 era where personal computers were barely a concept, 00:09:26.029 --> 00:09:28.330 yet they were mapping out how neural networks 00:09:28.330 --> 00:09:31.029 could transfer knowledge using geometry. And 00:09:31.029 --> 00:09:34.370 the use of geometry is key there. They were conceptualizing 00:09:34.370 --> 00:09:36.629 how data points relate to each other in high 00:09:36.629 --> 00:09:39.149 dimensional space. Right. Theorizing that if 00:09:39.149 --> 00:09:41.909 a network learns a geometric boundary for one 00:09:41.909 --> 00:09:44.929 concept, it could mathematically shift that boundary 00:09:44.929 --> 00:09:47.450 to understand a new one. But it was pure theory 00:09:47.450 --> 00:09:50.649 until 1981 when we see an actual experimental 00:09:50.649 --> 00:09:53.649 application. Right, the first real test. A report 00:09:53.649 --> 00:09:56.450 applied transfer learning to a data set of images 00:09:56.450 --> 00:09:59.250 representing letters on computer terminals. They 00:09:59.250 --> 00:10:01.509 were moving from abstract geometry to trying 00:10:01.509 --> 00:10:03.970 to get a machine to actually recognize visual 00:10:03.970 --> 00:10:06.590 data across different formats. And that transition 00:10:06.590 --> 00:10:08.929 from pure theory to visual application exposed 00:10:08.929 --> 00:10:11.110 the practical bottlenecks. It wasn't smooth. 00:10:11.269 --> 00:10:13.870 No, it took another decade to refine how we measure 00:10:13.870 --> 00:10:15.889 what should actually be transferred. Which brings 00:10:15.889 --> 00:10:19.710 us to 1992, when Lorian Pratt formulated the 00:10:19.710 --> 00:10:22.549 discriminability -based transfer algorithm, or 00:10:22.549 --> 00:10:26.049 DBT. DBT was a crucial stepping stone. Before 00:10:26.049 --> 00:10:27.649 you transfer the weights of a neural network 00:10:27.649 --> 00:10:31.570 from a source task to a target task, DBT essentially 00:10:31.570 --> 00:10:34.129 measures how well those old weights can discriminate 00:10:34.129 --> 00:10:37.350 or separate the new data classes. Like a compatibility 00:10:37.350 --> 00:10:39.450 test. Exactly. It was a mathematical way of asking, 00:10:39.649 --> 00:10:41.769 is this old knowledge actually going to be useful 00:10:41.769 --> 00:10:44.230 here? And that structural framework led to a 00:10:44.230 --> 00:10:47.519 major explosion in 1998, advancing the field 00:10:47.519 --> 00:10:50.860 into a multi -task learning, cemented by an influential 00:10:50.860 --> 00:10:53.299 book titled, Learning to Learn. There weren't 00:10:53.299 --> 00:10:55.559 just programming facts anymore. Right, they were 00:10:55.559 --> 00:10:57.559 trying to optimize the acquisition of facts, 00:10:57.940 --> 00:11:00.139 which ultimately snowballs into a massive wave 00:11:00.139 --> 00:11:03.450 of hype. Oh, the hype was real. By 2016, at the 00:11:03.450 --> 00:11:06.570 major AI conference in IPS, Andrew Ng, one of 00:11:06.570 --> 00:11:09.049 the most prominent voices in the field, declared 00:11:09.049 --> 00:11:10.889 that transfer learning would become the next 00:11:10.889 --> 00:11:13.049 driver of machine learning commercial success, 00:11:13.450 --> 00:11:16.389 right after supervised learning. And a prediction 00:11:16.389 --> 00:11:20.549 like that from Andrew Ng signals a massive influx 00:11:20.549 --> 00:11:23.950 of capital in research. It was really the moment 00:11:23.950 --> 00:11:26.169 transfer learning went from an academic pursuit 00:11:26.169 --> 00:11:29.850 to the presumed backbone of commercial AI architecture. 00:11:30.129 --> 00:11:31.570 But I want to pull a thread here because there 00:11:31.570 --> 00:11:34.730 is a glaring detail from that 1981 experiment 00:11:34.730 --> 00:11:37.570 with the computer terminal letters that completely 00:11:37.570 --> 00:11:40.870 undercuts the idea of a smooth upward trajectory. 00:11:41.330 --> 00:11:44.370 Uh -oh. Yeah. The experiment didn't just demonstrate 00:11:44.370 --> 00:11:46.950 that transfer learning worked. It experimentally 00:11:46.950 --> 00:11:49.649 demonstrated both positive and negative transfer 00:11:49.649 --> 00:11:51.610 learning. The dark side of transfer learning. 00:11:51.559 --> 00:11:54.840 Which means the AI learned something in the source 00:11:54.840 --> 00:11:57.100 domain that fundamentally made it worse at the 00:11:57.100 --> 00:11:59.419 target task. It's the computational equivalent 00:11:59.419 --> 00:12:01.879 of a bad habit. If the knowledge from the source 00:12:01.879 --> 00:12:04.620 domain is misapplied to the target domain, the 00:12:04.620 --> 00:12:06.940 model doesn't just fail to learn, it actively 00:12:06.940 --> 00:12:09.639 degrades. So if you learn to drive in the U .S. 00:12:09.700 --> 00:12:11.639 where the domain rules dictate driving on the 00:12:11.639 --> 00:12:13.279 right side of the road and you transfer that 00:12:13.279 --> 00:12:15.740 exact muscle memory to the U .K. You steer right 00:12:15.740 --> 00:12:18.419 into oncoming traffic. The old predictive function 00:12:18.419 --> 00:12:21.580 is catastrophically wrong for the new marginal 00:12:21.580 --> 00:12:25.159 probability distribution. The AI would have actually 00:12:25.159 --> 00:12:27.980 achieved a higher accuracy if it had simply started 00:12:27.980 --> 00:12:31.200 from absolute zero with randomized weights, rather 00:12:31.200 --> 00:12:33.019 than carrying over all the baggage of the source 00:12:33.019 --> 00:12:35.470 domain. And that specific reality of negative 00:12:35.470 --> 00:12:38.289 transfer leads perfectly into the modern scientific 00:12:38.289 --> 00:12:40.870 debate. It proved that even with all the commercial 00:12:40.870 --> 00:12:43.970 hype in 2016, the industry hadn't actually solved 00:12:43.970 --> 00:12:46.830 the foundational issues discovered in 1981. Not 00:12:46.830 --> 00:12:49.870 even close. The accepted dogma for a long time 00:12:49.870 --> 00:12:52.940 was this idea of pre -training. The standard 00:12:52.940 --> 00:12:56.500 operating procedure was to take a massive generalized 00:12:56.500 --> 00:12:59.679 data set like millions of scraped internet images 00:12:59.679 --> 00:13:02.100 and pre -train your model on that source domain. 00:13:02.259 --> 00:13:04.919 Throw everything at it. Exactly. Then you transfer 00:13:04.919 --> 00:13:06.740 those weights and fine -tune the model on your 00:13:06.740 --> 00:13:09.450 smaller specific target task. The assumption 00:13:09.450 --> 00:13:12.470 was that more initial data always equals a smarter 00:13:12.470 --> 00:13:14.350 starting point. The car to truck pipeline, but 00:13:14.350 --> 00:13:18.230 on a massive scale. Right. But a major 2020 paper 00:13:18.230 --> 00:13:21.029 by Zoff and colleagues, provocatively titled 00:13:21.029 --> 00:13:23.230 Rethinking Pre -Training and Self -Training, 00:13:23.470 --> 00:13:26.269 challenged that entire paradigm. They reported 00:13:26.269 --> 00:13:29.009 that pre -training can actually hurt accuracy. 00:13:29.169 --> 00:13:32.070 Wait, pre -training hurting accuracy is exactly 00:13:32.070 --> 00:13:34.929 negative transfer. Yes, exactly. If I train an 00:13:34.929 --> 00:13:37.950 AI on millions of general internet images, it 00:13:37.950 --> 00:13:40.350 builds incredible edge detectors and texture 00:13:40.350 --> 00:13:43.149 recognition. How could transferring that massive 00:13:43.149 --> 00:13:45.590 library of knowledge possibly hurt it when it 00:13:45.590 --> 00:13:48.750 looks at something specific like, say, an x -ray? 00:13:49.009 --> 00:13:52.019 This raises an important question. It hurts it 00:13:52.019 --> 00:13:54.600 because it forces the AI to view the new data 00:13:54.600 --> 00:13:57.379 entirely through the rigid lens of the old data. 00:13:58.360 --> 00:14:00.659 If the target domain has a fundamentally different 00:14:00.659 --> 00:14:03.860 feature space or probability distribution, the 00:14:03.860 --> 00:14:06.740 massive pre -trained model stubbornly clings 00:14:06.740 --> 00:14:09.129 to its original worldview. Like it's stuck in 00:14:09.129 --> 00:14:11.870 its ways. Yes. It looks for patterns that don't 00:14:11.870 --> 00:14:14.090 exist in the new domain, which ends up creating 00:14:14.090 --> 00:14:17.429 blind spots. So what did Zoff's paper advocate 00:14:17.429 --> 00:14:19.690 for instead of pre -training? They advocated 00:14:19.690 --> 00:14:22.610 for self -training. Instead of dragging the baggage 00:14:22.610 --> 00:14:24.850 of a massive external source domain into your 00:14:24.850 --> 00:14:27.870 project, you start with your specific target 00:14:27.870 --> 00:14:30.730 domain. You train a baseline model on whatever 00:14:30.730 --> 00:14:33.049 small amount of label data you have for your 00:14:33.049 --> 00:14:36.779 specific task. Then... You use that baseline 00:14:36.779 --> 00:14:40.039 model to generate pseudo labels for a massive 00:14:40.039 --> 00:14:43.080 amount of unlabeled data within that same domain. 00:14:43.360 --> 00:14:46.639 Finally, you train a new final model on both 00:14:46.639 --> 00:14:50.240 sets. Oh, I see. So self -training entirely bypasses 00:14:50.240 --> 00:14:52.299 the risk of negative transfer from an outside 00:14:52.299 --> 00:14:55.679 source domain because it forces the AI to bootstrap 00:14:55.679 --> 00:14:57.940 its understanding from within its own specific 00:14:57.940 --> 00:15:00.299 environment. Exactly. It's a localized evolution 00:15:00.299 --> 00:15:02.860 rather than an imported one. And it proved that 00:15:02.860 --> 00:15:05.679 AI development is just a messy science. Very 00:15:05.679 --> 00:15:07.580 messy. The people building these systems are 00:15:07.580 --> 00:15:10.039 constantly warring over the mathematical architecture. 00:15:10.580 --> 00:15:12.820 Accepted best practices like massive pre -training 00:15:12.809 --> 00:15:14.909 are constantly being audited and found lacking 00:15:14.909 --> 00:15:17.710 in specific contexts. Yet, despite these intense 00:15:17.710 --> 00:15:19.610 architectural debates over pre -training versus 00:15:19.610 --> 00:15:21.990 self -training, the actual deployed applications 00:15:21.990 --> 00:15:24.409 of transfer learning are pushing into incredibly 00:15:24.409 --> 00:15:27.730 diverse, almost sci -fi territories. Oh, the 00:15:27.730 --> 00:15:30.289 underlying principle of recycling learned parameters 00:15:30.289 --> 00:15:33.470 is incredibly versatile, assuming you can successfully 00:15:33.470 --> 00:15:36.269 map the domains. Just looking at the broad spectrum 00:15:36.269 --> 00:15:38.950 mentioned in the source, we have cancer subtype 00:15:38.950 --> 00:15:41.649 discovery, building room occupancy prediction, 00:15:42.029 --> 00:15:45.070 general game playing, text classification, and 00:15:45.070 --> 00:15:47.409 spam filtering. It's everywhere. It mentions 00:15:47.409 --> 00:15:49.929 that algorithms are available in Markov logic 00:15:49.929 --> 00:15:52.789 networks and Bayesian networks. Right. And to 00:15:52.789 --> 00:15:54.710 briefly look under the hood of how that works 00:15:54.710 --> 00:15:57.470 conceptually, think about a Markov logic network. 00:15:57.490 --> 00:16:00.950 OK. It applies first order logic formulas to 00:16:00.950 --> 00:16:04.299 a domain. In transfer learning, you might transfer 00:16:04.299 --> 00:16:06.639 the logical rules from the source domain, the 00:16:06.639 --> 00:16:08.879 structure of how things relate, but force the 00:16:08.879 --> 00:16:11.419 network to relearn the specific weights and probabilities 00:16:11.419 --> 00:16:14.279 for the target domain. Oh, so it keeps the framework 00:16:14.279 --> 00:16:16.840 but updates the math. Exactly. But here's where 00:16:16.840 --> 00:16:19.480 it gets really interesting. Moving beyond text 00:16:19.480 --> 00:16:22.700 and spam, there is a 2020 biological breakthrough 00:16:22.700 --> 00:16:25.200 detailed here that genuinely bends the mind. 00:16:25.519 --> 00:16:27.259 Ah, you're looking at the research connecting 00:16:27.259 --> 00:16:30.799 biological signals. Yes. In 2020, researchers 00:16:30.799 --> 00:16:32.679 discovered that transfer learning is possible 00:16:32.679 --> 00:16:36.559 between electromyographic signals, EMG, which 00:16:36.559 --> 00:16:39.039 are the electrical impulses generated by your 00:16:39.039 --> 00:16:42.559 physical muscles when you move, and electroencephalographic 00:16:42.559 --> 00:16:45.539 signals, EEG, which are your brain waves. It's 00:16:45.539 --> 00:16:47.759 incredible. They successfully transferred knowledge 00:16:47.759 --> 00:16:50.700 from the gesture recognition domain, the muscles, 00:16:50.799 --> 00:16:53.480 to the mental state recognition domain, the brain 00:16:53.480 --> 00:16:56.679 waves. How on earth do you mathematically map 00:16:56.679 --> 00:16:59.600 a physical muscle to an invisible brainwave. 00:16:59.740 --> 00:17:01.779 It feels like algorithmic telepathy, doesn't 00:17:01.779 --> 00:17:04.259 it? It really does. But it relies on a very grounded 00:17:04.259 --> 00:17:07.140 physical reality. The domains can be mapped because 00:17:07.140 --> 00:17:09.000 of the similar physical natures of the signals. 00:17:09.319 --> 00:17:12.099 But a bicep curling and a frontal lobe processing 00:17:12.099 --> 00:17:14.599 a thought are completely different biological 00:17:14.599 --> 00:17:17.220 functions. Biological functions, yes. But to 00:17:17.220 --> 00:17:19.440 the AI, they are both simply measuring voltage 00:17:19.440 --> 00:17:23.660 changes over time. EMG and EEG are both electrical 00:17:23.660 --> 00:17:26.259 activity. The amplitude is different, muscle 00:17:26.259 --> 00:17:28.460 signals are stronger, and the frequency ranges 00:17:28.460 --> 00:17:30.779 are different. But the underlying language of 00:17:30.779 --> 00:17:33.599 electrical fluctuation is shared. The AI maps 00:17:33.599 --> 00:17:36.200 the topography of the data, finding a shared 00:17:36.200 --> 00:17:39.380 latent space where the patterns of a muscle contraction 00:17:39.380 --> 00:17:41.700 correlate to the neural patterns that preceded 00:17:41.700 --> 00:17:43.809 it. And this wasn't just a one -way street, right? 00:17:44.109 --> 00:17:46.670 The data explicitly shows this relationship worked 00:17:46.670 --> 00:17:49.170 in both directions. That's the craziest part. 00:17:50.089 --> 00:17:52.309 Electroencephalographic brainwave data can likewise 00:17:52.309 --> 00:17:56.329 be used to classify physical EMG data. The AI 00:17:56.329 --> 00:17:58.450 can learn about thought to understand physical 00:17:58.450 --> 00:18:00.750 movement, or learn about physical movement to 00:18:00.750 --> 00:18:02.690 understand thought. And the performance metrics 00:18:02.690 --> 00:18:05.170 were concrete. They measured the improvement 00:18:05.170 --> 00:18:09.329 in two distinct phases. First, prior to any learning. 00:18:09.490 --> 00:18:12.009 Meaning the baseline starting point. Exactly. 00:18:12.430 --> 00:18:14.450 Normally, a neural network starts with completely 00:18:14.450 --> 00:18:17.420 random weights. It just guesses blindly. But 00:18:17.420 --> 00:18:19.420 by initializing the network with the transferred 00:18:19.420 --> 00:18:21.900 weights from the other biological domain, it 00:18:21.900 --> 00:18:24.000 started with a significantly higher accuracy 00:18:24.000 --> 00:18:26.680 than random chance. That is wild. And second, 00:18:26.759 --> 00:18:28.539 they measured it at the end of the learning process, 00:18:28.859 --> 00:18:32.640 the asymptote. The final trained model was demonstrably 00:18:32.640 --> 00:18:35.220 more accurate simply because it had been exposed 00:18:35.220 --> 00:18:37.970 to the other biological domains' data. It's like 00:18:37.970 --> 00:18:40.890 the AI games an intuition about the entire human 00:18:40.890 --> 00:18:43.509 nervous system by studying just one part of it. 00:18:43.569 --> 00:18:45.829 Precisely. And it notes that end users don't 00:18:45.829 --> 00:18:49.410 just have to accept the pre -trained model as 00:18:49.410 --> 00:18:51.809 is. They can change the structure of something 00:18:51.809 --> 00:18:54.690 called fully connected layers to improve performance 00:18:54.690 --> 00:18:58.190 even more. It specifically points to an architecture 00:18:58.190 --> 00:19:01.069 called SpinalNet. Right, and this gets to the 00:19:01.069 --> 00:19:03.309 heart of how we physically tweak these models 00:19:03.309 --> 00:19:06.509 for real -world application. Okay. In a typical 00:19:06.509 --> 00:19:09.109 neural network, the early convolutional layers 00:19:09.109 --> 00:19:12.650 do the raw feature extraction, pulling the electrical 00:19:12.650 --> 00:19:15.569 frequencies from the EEG. But the fully connected 00:19:15.569 --> 00:19:18.309 layers at the very end are where the AI makes 00:19:18.309 --> 00:19:20.650 its final decision, taking all those extracted 00:19:20.650 --> 00:19:22.710 features and assigning the label. So how does 00:19:22.710 --> 00:19:24.869 SpinalNet change that? Instead of dumping all 00:19:24.869 --> 00:19:27.150 the extracted features into the final fully connected 00:19:27.150 --> 00:19:29.750 layers all at once, SpinalNet feeds the inputs 00:19:29.750 --> 00:19:32.609 in gradually across multiple hidden layers. much 00:19:32.609 --> 00:19:35.089 like how a human's spine wouts nerve signals 00:19:35.089 --> 00:19:37.730 incrementally up to the brain. Exactly. The analogy 00:19:37.730 --> 00:19:41.289 is right in the name. And by tweaking that specific 00:19:41.289 --> 00:19:44.529 architectural layout, end users can take a generalized 00:19:44.529 --> 00:19:47.509 cross -domain model and hyper -refine it for 00:19:47.509 --> 00:19:49.769 a specific medical or technological application. 00:19:50.000 --> 00:19:52.180 So what does this all mean for you, the learner? 00:19:52.420 --> 00:19:54.880 Let's recap the complex terrain we just navigated 00:19:54.880 --> 00:19:56.819 today. We covered a lot of ground. We really 00:19:56.819 --> 00:19:59.759 did. We started with the foundational idea of 00:19:59.759 --> 00:20:02.400 transferring edge detection from cars to trucks 00:20:02.400 --> 00:20:05.200 just to save computing power. We broke down how 00:20:05.200 --> 00:20:08.099 developers artificially slice reality into domains 00:20:08.099 --> 00:20:11.140 with marginal probabilities and tasks with predictive 00:20:11.140 --> 00:20:13.700 functions. The heavy math. The heavy math. We 00:20:13.700 --> 00:20:17.440 tracked the causal chain from abstract 1976 geometry 00:20:17.440 --> 00:20:20.740 to the dbt algorithms of the 90s, all the way 00:20:20.740 --> 00:20:23.160 to the 2020 debate where self -training challenged 00:20:23.160 --> 00:20:25.559 the massive pre -training paradigm. And then 00:20:25.559 --> 00:20:28.200 into biology. Right. We ended up looking at AI 00:20:28.200 --> 00:20:30.359 architectures like SpinalNet translating the 00:20:30.359 --> 00:20:32.380 shared electrical language of human brainwaves 00:20:32.380 --> 00:20:34.599 and muscle spasms. And I think the through line 00:20:34.599 --> 00:20:36.640 here is that machine learning is a discipline 00:20:36.640 --> 00:20:39.799 of constant optimization. Always evolving. Yeah. 00:20:39.980 --> 00:20:42.339 So the next time your email filter seamlessly 00:20:42.339 --> 00:20:45.160 adapts to a new type of phishing scam, or you 00:20:45.160 --> 00:20:47.539 read about an AI detecting anomalies in medical 00:20:47.539 --> 00:20:50.000 imaging, you understand the mechanics at play. 00:20:50.359 --> 00:20:53.059 It's not magic. No. The AI isn't waking up with 00:20:53.059 --> 00:20:56.119 amnesia. It is standing on the shoulders of carefully 00:20:56.119 --> 00:20:59.019 engineered, transferred knowledge, constantly 00:20:59.019 --> 00:21:01.779 weighing the mathematical cost of every new piece 00:21:01.779 --> 00:21:04.309 of data. It is a profound structural advantage. 00:21:04.309 --> 00:21:06.150 But I want to leave you with one final thought 00:21:06.150 --> 00:21:08.450 to mull over. Oh, boy. We spent a lot of time 00:21:08.450 --> 00:21:11.490 on the 1981 discovery of negative transfer and 00:21:11.490 --> 00:21:15.190 how the 2020 Zoff paper proved that rigid pre 00:21:15.190 --> 00:21:18.190 -training can force an AI to view new data through 00:21:18.190 --> 00:21:21.529 a biased, stubborn lens, actively hurting its 00:21:21.529 --> 00:21:25.210 accuracy. Right. So if an AI has the power to 00:21:25.210 --> 00:21:27.690 map the mathematical topography of the electrical 00:21:27.690 --> 00:21:30.029 impulses in your physical muscles directly to 00:21:30.029 --> 00:21:32.750 the mental states of your brain waves, What happens 00:21:32.750 --> 00:21:35.250 when a massive generalized model transfers a 00:21:35.250 --> 00:21:37.210 fundamental misunderstanding from one domain 00:21:37.210 --> 00:21:39.630 into the other? What happens when an AI learns 00:21:39.630 --> 00:21:42.309 the wrong lesson from your muscles and uses it 00:21:42.309 --> 00:21:44.630 as the baseline to interpret your thoughts? That 00:21:44.630 --> 00:21:46.890 is a terrifying question. Something to think 00:21:46.890 --> 00:21:48.529 about. Until next time.