WEBVTT 00:00:00.000 --> 00:00:03.359 So if you want to master tennis, you know, you 00:00:03.359 --> 00:00:05.700 just practice tennis. Right, obviously. Yeah, 00:00:05.700 --> 00:00:07.719 you get out there on the court, you work on your 00:00:07.719 --> 00:00:11.740 serve, you drill your backhand, and you figure 00:00:11.740 --> 00:00:14.699 out the footwork. But what if I told you that 00:00:14.699 --> 00:00:18.199 the secret to becoming a world class tennis player 00:00:18.199 --> 00:00:22.460 faster was actually, well, that you had to decide 00:00:22.460 --> 00:00:24.839 to play competitive ping pong at the exact same 00:00:24.839 --> 00:00:27.239 time? I mean, that sounds like a recipe for completely 00:00:27.239 --> 00:00:29.679 scrambling your brain. Right. You would think 00:00:29.679 --> 00:00:32.039 the macro movements of a tennis swing and then 00:00:32.039 --> 00:00:34.979 like the micro flicks of a ping pong paddle would 00:00:34.979 --> 00:00:37.020 just actively interfere with each other. Yeah, 00:00:37.140 --> 00:00:38.859 it does sound like a terrible idea on the surface. 00:00:38.859 --> 00:00:41.320 Totally backwards. But underneath the wildly 00:00:41.320 --> 00:00:43.119 different scales of the tennis court and the 00:00:43.119 --> 00:00:45.079 ping pong table, you are essentially forcing 00:00:45.079 --> 00:00:48.829 your brain to train on the exact same underlying 00:00:48.829 --> 00:00:51.590 mechanics. Like you're building foundational 00:00:51.590 --> 00:00:54.429 spatial awareness, spin prediction, and hand 00:00:54.429 --> 00:00:56.450 -eye coordination. And that counterintuitive 00:00:56.450 --> 00:00:59.090 idea is exactly what we are exploring today. 00:00:59.630 --> 00:01:01.609 So welcome to the Deep Dive. Thanks for having 00:01:01.609 --> 00:01:05.269 me. We are taking a look at a massive, incredibly 00:01:05.269 --> 00:01:07.989 dense set of source material centered around 00:01:07.989 --> 00:01:10.469 a Wikipedia article on this concept called multi 00:01:10.469 --> 00:01:14.150 -task learning, or MTO. Yes, MTL. And our mission 00:01:14.150 --> 00:01:17.469 for you today is to demystify how machine learning 00:01:17.469 --> 00:01:20.290 models actually get significantly smarter by 00:01:20.290 --> 00:01:23.250 juggling multiple tasks at once. Which is fascinating. 00:01:23.469 --> 00:01:25.150 Because for us humans who usually just want to 00:01:25.150 --> 00:01:27.409 focus on one thing at a time, it feels, you know, 00:01:27.430 --> 00:01:29.790 totally backwards. It does. And to establish 00:01:29.790 --> 00:01:33.010 exactly why this matters for you and really the 00:01:33.010 --> 00:01:35.569 technology you use every day, we have to look 00:01:35.569 --> 00:01:38.170 at the context of how artificial intelligence 00:01:38.170 --> 00:01:41.030 traditionally learns. Historically, we train 00:01:41.030 --> 00:01:44.409 in AI. in a complete vacuum. We give it one single 00:01:44.409 --> 00:01:47.329 objective, a massive pile of data, and it learns 00:01:47.329 --> 00:01:49.170 that one specific thing. Right, like a horse 00:01:49.170 --> 00:01:52.370 with blinders on. Exactly. But multi -task learning, 00:01:52.590 --> 00:01:55.870 or MTL, forces the AI to look at multiple different 00:01:55.870 --> 00:01:59.409 tasks simultaneously. It forces the system to 00:01:59.409 --> 00:02:01.450 exploit the commonalities and the differences 00:02:01.450 --> 00:02:04.209 across all those tasks. So it's not getting distracted? 00:02:04.549 --> 00:02:07.500 Not at all. When you compel a system to solve 00:02:07.500 --> 00:02:10.139 multiple problems at once, you aren't distracting 00:02:10.139 --> 00:02:13.080 it. You are actually improving both its learning 00:02:13.080 --> 00:02:15.520 efficiency and its prediction accuracy across 00:02:15.520 --> 00:02:17.520 the board. Okay, let's unpack this, because if 00:02:17.520 --> 00:02:19.300 we look at the source text, it points out that 00:02:19.300 --> 00:02:21.560 the early versions of this concept were actually 00:02:21.560 --> 00:02:24.759 just called hints. Yeah, hints. Which is a surprisingly 00:02:24.759 --> 00:02:27.400 gentle term for computer science, honestly. It 00:02:27.400 --> 00:02:29.699 really is. But then we get to this foundational 00:02:29.699 --> 00:02:33.319 1997 paper by a researcher named Rich Caruana. 00:02:33.439 --> 00:02:35.389 Right, a huge paper. paper in the field. And 00:02:35.389 --> 00:02:38.550 he theorized that MTL uses the training signals 00:02:38.550 --> 00:02:42.090 of related tasks as, and I'm quoting here, inductive 00:02:42.090 --> 00:02:45.590 bias while using a shared representation. Yes. 00:02:45.750 --> 00:02:48.169 Now even for a deep dive, I mean, that is some 00:02:48.169 --> 00:02:50.430 heavy academic phrasing. It is a bit dense. So 00:02:50.430 --> 00:02:52.449 let's break those terms down because the mechanics 00:02:52.449 --> 00:02:54.409 underneath them are really brilliant. Please 00:02:54.409 --> 00:02:57.669 do. So in machine learning, an inductive bias 00:02:57.669 --> 00:03:01.110 is essentially the set of assumptions a learning 00:03:01.110 --> 00:03:04.580 algorithm uses to predict outcomes. for situations 00:03:04.580 --> 00:03:06.500 it has never seen before. OK, so it's like it's 00:03:06.500 --> 00:03:09.020 gut instinct. Kind of, yeah. Inductive means 00:03:09.020 --> 00:03:12.360 it's generalizing from specific examples, and 00:03:12.360 --> 00:03:14.819 bias means it leans towards certain types of 00:03:14.819 --> 00:03:17.199 conclusions. Got it. Think of an AI trying to 00:03:17.199 --> 00:03:20.000 guess the rules of a card game just by watching 00:03:20.000 --> 00:03:23.120 people play. OK. If it only watches one specific 00:03:23.120 --> 00:03:26.889 game, say poker, it might draw the wrong conclusions 00:03:26.889 --> 00:03:29.310 about how all cards work. Like, it might think 00:03:29.310 --> 00:03:31.870 the goal is always to hoard chips. Exactly. But 00:03:31.870 --> 00:03:34.590 if it watches poker, blackjack, and solitaire 00:03:34.590 --> 00:03:37.500 at the same time... The hints it gets from the 00:03:37.500 --> 00:03:39.939 other games guide it toward a much more robust 00:03:39.939 --> 00:03:42.740 understanding. Oh, I see. It learns what a suit 00:03:42.740 --> 00:03:45.699 is or what shuffling does, completely independent 00:03:45.699 --> 00:03:48.180 of the specific game. Let's bring this into the 00:03:48.180 --> 00:03:50.240 real world with the spam filter example from 00:03:50.240 --> 00:03:52.340 the text, because I think this makes Caruana's 00:03:52.340 --> 00:03:54.219 theory instantly click. That's a great example. 00:03:54.360 --> 00:03:56.759 So imagine you are building a spam filter for 00:03:56.759 --> 00:03:59.740 two different users. User A is an English speaker. 00:03:59.860 --> 00:04:02.740 User B is a Russian speaker. Right. Now, if user 00:04:02.740 --> 00:04:05.909 A receives an email entirely in Russian, The 00:04:05.909 --> 00:04:09.009 AI standard single task filter is going to flag 00:04:09.009 --> 00:04:11.349 that immediately. It sees the Cyrillic alphabet 00:04:11.349 --> 00:04:14.490 and says, obviously, spam. Right, because for 00:04:14.490 --> 00:04:17.069 the English speaker, the system just learns a 00:04:17.069 --> 00:04:19.610 lazy rule. Lazy rule. Yeah, it basically just 00:04:19.610 --> 00:04:23.009 says, Cyrillic equals bad. It finds the easiest 00:04:23.009 --> 00:04:25.629 possible shortcut to solve the specific problem 00:04:25.629 --> 00:04:27.879 it was given. Right. But for the Russian speaker, 00:04:28.060 --> 00:04:30.399 that lazy rule would destroy their inbox. Exactly. 00:04:30.500 --> 00:04:32.500 I mean, a Russian email is perfectly normal for 00:04:32.500 --> 00:04:35.220 them. So if you train two separate AI models 00:04:35.220 --> 00:04:38.199 in a vacuum, they learn totally different roles. 00:04:38.259 --> 00:04:41.509 They do. But let's say both users start receiving 00:04:41.509 --> 00:04:44.769 scam emails about an urgent wire transfer. Oh, 00:04:44.769 --> 00:04:48.569 yes. If the AI is using multitask learning, it 00:04:48.569 --> 00:04:51.750 groups these distinct but related classification 00:04:51.750 --> 00:04:54.730 tasks together into a shared representation. 00:04:55.209 --> 00:04:57.550 It suddenly realizes, wait, whether the text 00:04:57.550 --> 00:05:00.970 is in English or Russian, this specific manipulative 00:05:00.970 --> 00:05:03.529 phrasing about a wire transfer is a universal 00:05:03.529 --> 00:05:05.810 indicator of a scam. And that reveals the core 00:05:05.810 --> 00:05:08.290 mechanism of why this is so powerful. It's what 00:05:08.290 --> 00:05:10.790 we call regularization. Regularization. OK. Yeah. 00:05:10.850 --> 00:05:12.250 In machine learning, you always run the risk 00:05:12.250 --> 00:05:14.509 of your model overfitting. Overfitting. Right. 00:05:14.509 --> 00:05:17.750 Overfitting happens when an AI becomes obsessively 00:05:17.750 --> 00:05:20.910 fixated on the specific random quirks of its 00:05:20.910 --> 00:05:23.290 training data rather than the actual underlying 00:05:23.290 --> 00:05:25.529 patterns. Like assuming all Russian is spam. 00:05:25.629 --> 00:05:29.250 Exactly. By jointly solving each user's spam 00:05:29.250 --> 00:05:32.569 problem, the AI solutions actively regularize 00:05:32.569 --> 00:05:34.910 each other. So they balance each other out. Yes. 00:05:35.329 --> 00:05:38.370 The Russian task acts as a guardrail for the 00:05:38.370 --> 00:05:41.089 English task, effectively telling it, hey, don't 00:05:41.089 --> 00:05:43.149 focus so much on the alphabet, focus on the intent. 00:05:43.170 --> 00:05:45.750 That makes a lot of sense. Regularization mathematically 00:05:45.750 --> 00:05:49.029 forces the system to zoom out and learn the actual 00:05:49.029 --> 00:05:51.810 universal rules of what makes an email spam. 00:05:52.230 --> 00:05:55.269 So grouping related tasks makes perfect logical 00:05:55.269 --> 00:05:57.670 sense. They share a fundamental goal. They do. 00:05:57.899 --> 00:05:59.519 But here's where it gets really interesting, 00:05:59.800 --> 00:06:01.459 because the source material takes a hard left 00:06:01.459 --> 00:06:03.459 turn here. It does take a leap. It introduces 00:06:03.459 --> 00:06:06.819 a concept called auxiliary learning, and it claims 00:06:06.819 --> 00:06:09.819 that teaching an AI completely entirely unrelated 00:06:09.819 --> 00:06:12.980 tasks alongside its main goal actually improves 00:06:12.980 --> 00:06:15.639 its performance even more. Yes. Wait, so teaching 00:06:15.639 --> 00:06:18.399 an AI unrelated things actually helps. It really 00:06:18.399 --> 00:06:20.759 does. Isn't that like a professional race car 00:06:20.759 --> 00:06:23.779 driver? training for a championship by taking 00:06:23.779 --> 00:06:26.199 up juggling? I mean, juggling has absolutely 00:06:26.199 --> 00:06:28.420 nothing to do with steering a car. It sounds 00:06:28.420 --> 00:06:32.240 completely paradoxical. Totally. Logically, forcing 00:06:32.240 --> 00:06:36.000 a system to learn unrelated tasks should just 00:06:36.000 --> 00:06:39.199 introduce chaos. But the race car driver analogy 00:06:39.199 --> 00:06:43.040 is actually perfect. Yeah. Juggling doesn't teach 00:06:43.040 --> 00:06:45.600 the driver how to shift gears, but it forces 00:06:45.600 --> 00:06:47.959 their brain to build hyper -efficient spatial 00:06:47.959 --> 00:06:50.819 awareness and peripheral reaction times. Oh, 00:06:50.939 --> 00:06:53.360 wow. And those underlying neurological upgrades 00:06:53.360 --> 00:06:55.639 then make them a better driver. I hadn't thought 00:06:55.639 --> 00:06:57.500 of it like that. The research shows that when 00:06:57.500 --> 00:07:00.959 you joint learn a principal task alongside completely 00:07:00.959 --> 00:07:04.740 unrelated auxiliary tasks, crucially using the 00:07:04.740 --> 00:07:07.480 exact same input data, it provides a massive 00:07:07.480 --> 00:07:09.639 improvement over standard multitask learning. 00:07:09.790 --> 00:07:11.649 But how is that practically happening inside 00:07:11.649 --> 00:07:14.350 the computer? If I'm trying to predict the stock 00:07:14.350 --> 00:07:16.730 market, why would I also ask the AI to count 00:07:16.730 --> 00:07:18.870 the number of vowels in the financial reports? 00:07:19.310 --> 00:07:21.449 How does the juggling help the driving in data? 00:07:21.769 --> 00:07:23.949 What's fascinating here is how the system is 00:07:23.949 --> 00:07:26.509 forced to handle the concept of noise. Noise, 00:07:26.569 --> 00:07:29.689 OK. In any data set, there's the core truth you're 00:07:29.689 --> 00:07:32.850 looking for, the signal. Right. And then there 00:07:32.850 --> 00:07:35.110 are the idiosyncrasies of the data distribution. 00:07:35.149 --> 00:07:37.889 Right. The random useless fluctuations, which 00:07:37.889 --> 00:07:41.410 we call noise. OK. So signal and noise. Exactly. 00:07:41.569 --> 00:07:45.069 Yeah. When you give an AI two totally unrelated 00:07:45.069 --> 00:07:48.529 tasks to solve using the exact same input data, 00:07:49.050 --> 00:07:51.269 you are essentially forcing it to build a much 00:07:51.269 --> 00:07:54.649 sparser, far more informative foundational representation 00:07:54.649 --> 00:07:56.649 of that data. Because it can't rely on those 00:07:56.649 --> 00:07:59.420 lazy shortcuts anymore. Exactly. If it uses a 00:07:59.420 --> 00:08:02.220 cheap trick or memorizes the noise to solve task 00:08:02.220 --> 00:08:05.480 A, that trick will almost certainly fail catastrophically 00:08:05.480 --> 00:08:08.540 on the completely unrelated task B. Right. The 00:08:08.540 --> 00:08:10.899 system literally imposes a mathematical penalty 00:08:10.899 --> 00:08:13.800 on itself for relying on noise. The architecture 00:08:13.800 --> 00:08:15.879 encourages the mathematical representations of 00:08:15.879 --> 00:08:18.100 these different groups of tasks to be orthogonal. 00:08:18.319 --> 00:08:20.500 Orthogonal. Meaning they operate at right angles 00:08:20.500 --> 00:08:22.600 to each other. Yes. But what does a right angle 00:08:22.600 --> 00:08:24.939 mean for data? In mathematical and data terms, 00:08:25.660 --> 00:08:28.920 if two variables are orthogonal, It means changing 00:08:28.920 --> 00:08:31.139 one has absolutely zero effect on the other. 00:08:31.560 --> 00:08:33.159 They are completely independent. So they don't 00:08:33.159 --> 00:08:35.559 touch at all. Right. It's like the volume knob 00:08:35.559 --> 00:08:38.600 and the channel dial on a radio. Okay. Twisting 00:08:38.600 --> 00:08:40.740 the volume doesn't change the station, and changing 00:08:40.740 --> 00:08:42.879 the station doesn't change the volume. I like 00:08:42.879 --> 00:08:46.659 that analogy. By forcing the AI to study the 00:08:46.659 --> 00:08:49.799 stock market prices and count the vowels simultaneously, 00:08:50.440 --> 00:08:53.000 the AI is forced to separate the core signal 00:08:53.279 --> 00:08:56.720 the actual financial data from the specific unrelated 00:08:56.720 --> 00:08:58.700 tasks. Right, because vowels don't drive the 00:08:58.700 --> 00:09:01.320 stock price. Exactly. It realizes the only way 00:09:01.320 --> 00:09:04.299 to solve both is to strip away all the surface 00:09:04.299 --> 00:09:06.960 level noise and extract only the absolute most 00:09:06.960 --> 00:09:09.539 fundamental features of the raw data before it 00:09:09.539 --> 00:09:11.720 splits off to solve the specific problems. That 00:09:11.720 --> 00:09:14.039 is mind -bending. The distraction is actually 00:09:14.039 --> 00:09:16.460 the cure for the noise. It really is. The system 00:09:16.460 --> 00:09:19.500 is forced to become profound rather than superficial. 00:09:19.720 --> 00:09:22.159 And it builds a foundation so solid that it can 00:09:22.159 --> 00:09:24.370 apply that underlying understanding to almost 00:09:24.370 --> 00:09:26.549 anything. Okay, so this makes sense if the world 00:09:26.549 --> 00:09:29.049 stands still, you know. You have a static set 00:09:29.049 --> 00:09:31.330 of tasks, you feed them to the AI and it strips 00:09:31.330 --> 00:09:33.450 away the noise. Right. But I'm thinking about 00:09:33.450 --> 00:09:35.289 something like my Netflix account. All right. 00:09:35.710 --> 00:09:37.929 What I want to watch on a Tuesday morning before 00:09:37.929 --> 00:09:41.289 work is totally different from what I want to 00:09:41.289 --> 00:09:44.409 watch on a Friday night. Oh, absolutely. My preferences 00:09:44.409 --> 00:09:47.899 are a constantly moving target. If my data is 00:09:47.899 --> 00:09:50.399 constantly shifting, doesn't this shared memory 00:09:50.399 --> 00:09:53.759 become outdated immediately? Like, how does MTL 00:09:53.759 --> 00:09:57.879 adapt to time and evolution? That is the exact 00:09:57.879 --> 00:10:00.659 problem researchers had to solve next. And it 00:10:00.659 --> 00:10:02.539 requires us to distinguish between concurrent 00:10:02.539 --> 00:10:04.600 learning, which is the traditional MTL we've 00:10:04.600 --> 00:10:07.059 been discussing, and the sequential transfer 00:10:07.059 --> 00:10:10.000 of knowledge. Sequential transfer? Yeah. In concurrent 00:10:10.000 --> 00:10:12.000 learning, the shared representation is developed 00:10:12.000 --> 00:10:14.580 across all tasks at the exact same time. Right. 00:10:14.820 --> 00:10:16.950 But in transfer learning, you develop a massive 00:10:16.950 --> 00:10:19.850 robust representation first, and then pass it 00:10:19.850 --> 00:10:22.769 down the line to new tasks. The Wikipedia article 00:10:22.769 --> 00:10:25.190 cites GoogleNet as a prime example of this. Which 00:10:25.190 --> 00:10:28.330 is this massive deep convolutional neural network 00:10:28.330 --> 00:10:30.889 used for image classification. Right. It takes 00:10:30.889 --> 00:10:33.649 an enormous amount of computational power to 00:10:33.649 --> 00:10:36.649 train GoogleNet to understand what a cat looks 00:10:36.649 --> 00:10:38.250 like from a million different angles. Now I can 00:10:38.250 --> 00:10:40.909 imagine. But in doing so, the initial layers 00:10:40.909 --> 00:10:43.029 of the network aren't actually learning cat. 00:10:43.340 --> 00:10:45.039 Wait, what are they learning, though? They are 00:10:45.039 --> 00:10:47.659 learning how to detect an edge. They are learning 00:10:47.659 --> 00:10:50.779 how to perceive a curve or a shadow or a texture. 00:10:51.039 --> 00:10:54.500 Oh, wow. Once it has developed that incredibly 00:10:54.500 --> 00:10:58.159 robust mathematical representation of basic geometry, 00:10:59.360 --> 00:11:02.179 that knowledge isn't locked away. OK. It can 00:11:02.179 --> 00:11:04.960 be used to initialize an entirely different model 00:11:04.960 --> 00:11:09.009 meant to detect, say, tumors. in medical scans. 00:11:09.429 --> 00:11:11.950 Really? From cats to tumors? Yes. You sequentially 00:11:11.950 --> 00:11:14.570 transfer that pre -trained foundation so the 00:11:14.570 --> 00:11:16.350 new algorithm doesn't have to start from scratch. 00:11:16.570 --> 00:11:18.250 It's like sending someone to art school for four 00:11:18.250 --> 00:11:20.289 years to learn everything about light, shadow, 00:11:20.429 --> 00:11:22.789 and anatomy, right? Yes. And then on graduation 00:11:22.789 --> 00:11:25.289 day, handing them a tattoo gun? Yeah. Like, they 00:11:25.289 --> 00:11:28.039 don't know how to tattoo yet? But their foundational 00:11:28.039 --> 00:11:30.480 understanding of how shapes work is so strong 00:11:30.480 --> 00:11:33.440 that they will learn the specific art of tattooing 00:11:33.440 --> 00:11:35.620 infinitely faster than someone off the street. 00:11:36.100 --> 00:11:38.899 That is a brilliant way to visualize it. But 00:11:38.899 --> 00:11:41.500 the researchers pushed it even further to address 00:11:41.500 --> 00:11:44.220 your Netflix problem. The moving target problem. 00:11:44.299 --> 00:11:47.139 Right. What happens if the environment is continuously 00:11:47.139 --> 00:11:50.179 changing while the AI is operating? OK. They 00:11:50.179 --> 00:11:52.759 developed something called GoAL, which stands 00:11:52.759 --> 00:11:56.320 for Group Online Adaptive Learning. GoAL. This 00:11:56.320 --> 00:11:59.000 applies the principles of MTL to non -stationary 00:11:59.000 --> 00:12:02.240 environments. With GoAL, a learner operating 00:12:02.240 --> 00:12:04.700 in a shifting environment, like a financial time 00:12:04.700 --> 00:12:07.600 series or a recommendation algorithm, can actually 00:12:07.600 --> 00:12:09.840 benefit from the previous experience of another 00:12:09.840 --> 00:12:12.340 learner to quickly adapt to his new surroundings. 00:12:12.559 --> 00:12:14.519 And this leads right into a phrase from the text 00:12:14.519 --> 00:12:16.360 that totally stopped me on my tracks. Oh yeah. 00:12:16.480 --> 00:12:18.759 Yeah, under the optimization section, it talks 00:12:18.759 --> 00:12:21.600 about evolutionary multitasking. I mean, are 00:12:21.600 --> 00:12:24.200 we talking about literal AI genetics here? Are 00:12:24.200 --> 00:12:26.480 we breeding algorithms to see which ones survive? 00:12:26.720 --> 00:12:29.860 We actually are. Really? Yeah. Instead of just 00:12:29.860 --> 00:12:33.220 tweaking one algorithm, the researchers map all 00:12:33.220 --> 00:12:36.360 of these distinct optimization tasks into a single 00:12:36.360 --> 00:12:39.639 unified search space. OK. And within that space, 00:12:39.840 --> 00:12:42.320 you have a whole population of candidate solutions. 00:12:42.639 --> 00:12:45.539 It relies on continuous genetic transfer. OK, 00:12:45.559 --> 00:12:48.340 but how does an algorithm transfer genetics? 00:12:48.440 --> 00:12:50.620 I mean, it doesn't have DNA. Through a process 00:12:50.620 --> 00:12:53.740 that mirrors biological crossover. As this evolving 00:12:53.740 --> 00:12:56.100 population of candidate solutions tries to solve 00:12:56.100 --> 00:12:59.279 these different tasks, they harness hidden relationships 00:12:59.279 --> 00:13:02.240 through something called implicit parallelism. 00:13:02.580 --> 00:13:05.039 Implicit parallelism. Implicit parallelism basically 00:13:05.039 --> 00:13:08.509 means that by evaluating one single child the 00:13:08.509 --> 00:13:11.210 system is secretly testing dozens of underlying 00:13:11.210 --> 00:13:13.870 mathematical traits at the same time without 00:13:13.870 --> 00:13:16.100 doing any extra math. Can you give me an example 00:13:16.100 --> 00:13:18.759 of that? Sure. If I test a car's top speed, I'm 00:13:18.759 --> 00:13:20.899 implicitly testing its engine, its aerodynamics, 00:13:21.080 --> 00:13:23.259 and its tires all at once. So if candidate A 00:13:23.259 --> 00:13:25.299 is really good at predicting stock volatility 00:13:25.299 --> 00:13:27.659 and candidate B is really good at sentiment analysis 00:13:27.659 --> 00:13:29.860 of news articles, they literally cross over their 00:13:29.860 --> 00:13:33.620 traits. Yes. The system physically swaps segments 00:13:33.620 --> 00:13:36.419 of their underlying code. That's wild. It induces 00:13:36.419 --> 00:13:39.399 a leap in performance. Suddenly... Their offspring 00:13:39.399 --> 00:13:42.899 is this hyper -aware financial savant that understands 00:13:42.899 --> 00:13:45.379 both the math of the market and the mood of the 00:13:45.379 --> 00:13:48.279 news. And because it's happening across a whole 00:13:48.279 --> 00:13:51.100 population simultaneously, the evolutionary system 00:13:51.100 --> 00:13:53.519 progresses all the distinct tasks at the same 00:13:53.519 --> 00:13:56.220 time. The experience of the entire population 00:13:56.409 --> 00:13:59.090 lifts every individual solver. So we have this 00:13:59.090 --> 00:14:02.370 beautiful system. It groups users. It filters 00:14:02.370 --> 00:14:05.429 spam. It sees through noise by juggling unrelated 00:14:05.429 --> 00:14:08.710 things. And it breeds super algorithms through 00:14:08.710 --> 00:14:11.009 genetic crossover. That sounds perfect. Yeah. 00:14:11.070 --> 00:14:12.909 But what happens if the tasks we group together 00:14:12.909 --> 00:14:14.889 actually hate each other? What if they have opposing 00:14:14.889 --> 00:14:17.070 goals? We have to look at the system's vulnerability. 00:14:17.190 --> 00:14:19.230 We do. And to understand that vulnerability, 00:14:19.669 --> 00:14:21.789 the source material dives into some very heavy 00:14:21.789 --> 00:14:24.990 mathematics. It brings up something called RKHSVV. 00:14:25.090 --> 00:14:27.870 Right. Yes, the reproducing kernel Hilbert space 00:14:27.870 --> 00:14:30.549 of vector -valued functions along with separable 00:14:30.549 --> 00:14:33.470 kernels, which is quite the mouthful. It is, 00:14:33.470 --> 00:14:36.649 but the mechanism is crucial. And there's a political 00:14:36.649 --> 00:14:38.850 example in the text to explain this. Just to 00:14:38.850 --> 00:14:40.789 be completely clear to you listening, we're not 00:14:40.789 --> 00:14:42.649 taking any political sides here. Right. We're 00:14:42.649 --> 00:14:45.049 purely looking at the math of how data is distributed. 00:14:45.129 --> 00:14:47.309 Exactly. Just looking at the original source 00:14:47.309 --> 00:14:50.629 material impartially. So the example talks about 00:14:50.629 --> 00:14:53.409 predicting a politician's favorability rating 00:14:53.409 --> 00:14:56.519 based on political party. Yes. Think of an RK 00:14:56.519 --> 00:15:00.240 HSVV, like a massive multi -dimensional sorting 00:15:00.240 --> 00:15:03.500 facility. If you just have a standard 2D map 00:15:03.500 --> 00:15:06.460 of politicians and voters based on favorability, 00:15:07.200 --> 00:15:09.059 everyone is jumbled together in a muddy mess. 00:15:09.179 --> 00:15:11.899 Just a big blob of data. Right. Taking an average 00:15:11.899 --> 00:15:14.740 tells you nothing. But a Hilbert space uses complex 00:15:14.740 --> 00:15:17.399 matrices to mathematically toss all that data 00:15:17.399 --> 00:15:20.679 into a 3D or 4D room. Oh, interesting. Suddenly, 00:15:20.820 --> 00:15:22.860 the boundaries between how a Republican views 00:15:22.860 --> 00:15:24.940 the politician and how a Democrat views them 00:15:24.940 --> 00:15:27.639 become obvious. Because they're physically separated 00:15:27.639 --> 00:15:30.600 in that 4D space. Exactly. The math allows the 00:15:30.600 --> 00:15:32.659 system to regularize the predictions by grouping 00:15:32.659 --> 00:15:35.240 people by party, controlling the variance with 00:15:35.240 --> 00:15:37.659 respect to a group mean. It acknowledges that 00:15:37.659 --> 00:15:40.100 the underlying rules for predicting a Democrat's 00:15:40.100 --> 00:15:43.220 view are structurally different from a Republican's, 00:15:43.220 --> 00:15:46.240 but maps them within the same societal matrix. 00:15:46.480 --> 00:15:49.700 So what does this all mean? The system is great 00:15:49.700 --> 00:15:52.279 at mapping things out, but right after establishing 00:15:52.279 --> 00:15:55.360 this 4D sorting room, the text introduces the 00:15:55.360 --> 00:15:58.179 concept of negative transfer. Yes, the dark side. 00:15:58.419 --> 00:16:00.500 From reading the source, it sounds like a tug 00:16:00.500 --> 00:16:03.580 of war. Let's say you have a shared module trying 00:16:03.580 --> 00:16:07.200 to learn two tasks. If those tasks seek conflicting 00:16:07.200 --> 00:16:09.940 features, their mathematical gradients point 00:16:09.940 --> 00:16:12.899 in completely opposing directions. And to clarify 00:16:12.899 --> 00:16:15.360 for you, a gradient is essentially the direction 00:16:15.360 --> 00:16:17.519 the AI needs to go to find the right answer. 00:16:17.580 --> 00:16:20.440 Right. In gradient descent optimization, which 00:16:20.440 --> 00:16:22.559 is the engine driving these deep neural networks, 00:16:23.340 --> 00:16:25.879 the AI is basically blindfolded, taking small 00:16:25.879 --> 00:16:28.399 steps down a mathematical mountain to find the 00:16:28.399 --> 00:16:30.759 lowest possible point of error. The valley. Right, 00:16:30.820 --> 00:16:33.000 the lowest point is the optimal solution. But 00:16:33.000 --> 00:16:36.100 if task A says, the lowest point is to the left, 00:16:36.240 --> 00:16:38.399 and task B says, no, the lowest point is to the 00:16:38.399 --> 00:16:40.919 right, you are literally asking the AI to go 00:16:40.919 --> 00:16:44.539 north and south at the exact same time. The shared 00:16:44.539 --> 00:16:47.019 module just gets paralyzed. That is negative 00:16:47.019 --> 00:16:49.580 transfer in a nutshell. The gradients cancel 00:16:49.580 --> 00:16:52.659 each other out or they violently swing the AI 00:16:52.659 --> 00:16:54.779 back and forth across the mountains. Ripping 00:16:54.779 --> 00:16:57.879 it apart. Yeah. The simultaneous training of 00:16:57.879 --> 00:17:00.539 these supposedly related tasks actually hinders 00:17:00.539 --> 00:17:02.700 the overall performance compared to just letting 00:17:02.700 --> 00:17:05.380 them learn on their own single task models. Okay. 00:17:05.900 --> 00:17:07.940 The joint feature representation breaks down 00:17:07.940 --> 00:17:11.220 because it cannot serve two masters with fundamentally 00:17:11.220 --> 00:17:13.700 opposing needs. It's like tying the tennis player 00:17:13.700 --> 00:17:15.720 and the ping pong player together at the waist 00:17:15.720 --> 00:17:17.539 and telling them to play their respective games. 00:17:17.920 --> 00:17:19.339 They're just going to drag each other to the 00:17:19.339 --> 00:17:22.359 floor. Precisely. But the researchers came up 00:17:22.359 --> 00:17:24.740 with a deeply elegant solution to this problem, 00:17:25.599 --> 00:17:27.920 game theoretic optimization. Game theoretic. 00:17:27.980 --> 00:17:30.980 Yeah. They propose viewing the optimization problem 00:17:30.980 --> 00:17:34.400 as a literal game where every single task is 00:17:34.400 --> 00:17:37.390 a self -interested player. And all of these players 00:17:37.390 --> 00:17:40.069 are competing through a reward matrix, trying 00:17:40.069 --> 00:17:43.069 to reach a state that satisfies everyone. I, 00:17:43.069 --> 00:17:45.789 uh, I have to push back on this. Okay. Wait. 00:17:46.190 --> 00:17:49.170 Math doesn't have a conscience. How do you force 00:17:49.170 --> 00:17:51.730 a mathematical equation to compromise if its 00:17:51.730 --> 00:17:54.150 only directive is to find its own lowest error 00:17:54.150 --> 00:17:56.470 rate? I mean, selfish algorithms don't just sit 00:17:56.470 --> 00:17:59.089 around a table and negotiate. If we connect this 00:17:59.089 --> 00:18:01.529 to the bigger picture, we could look at a concept 00:18:01.529 --> 00:18:04.089 from economics called the Nash Cooperative Dargaining 00:18:04.089 --> 00:18:07.380 System. Like John Nash. Yes. In traditional gradient 00:18:07.380 --> 00:18:10.380 descent, the problem is that each task just aggressively 00:18:10.380 --> 00:18:13.259 shouts its own loss rate and demands the whole 00:18:13.259 --> 00:18:15.200 system move in its direction. It's a room full 00:18:15.200 --> 00:18:17.380 of people screaming their own demands. Right. 00:18:17.819 --> 00:18:20.380 But the game theoretic approach changes the mathematical 00:18:20.380 --> 00:18:23.500 incentives. It defines a game matrix where the 00:18:23.500 --> 00:18:26.079 reward for each task is no longer just getting 00:18:26.079 --> 00:18:28.730 its own way. Okay, then what is the reward? The 00:18:28.730 --> 00:18:30.849 reward is calculated based on the agreement of 00:18:30.849 --> 00:18:32.930 its own gradient with the common gradient of 00:18:32.930 --> 00:18:35.589 the group. So how does the math actually calculate 00:18:35.589 --> 00:18:38.809 that agreement? It calculates a vector that minimizes 00:18:38.809 --> 00:18:42.109 the maximum loss among all the tasks. Instead 00:18:42.109 --> 00:18:44.490 of just averaging the directions, which might 00:18:44.490 --> 00:18:47.710 send one task completely off a cliff, the algorithm 00:18:47.710 --> 00:18:50.349 finds a specific angle of descent that guarantees 00:18:50.349 --> 00:18:54.880 no single task. goes backwards. Wow. You mathematically 00:18:54.880 --> 00:18:58.059 mandate cooperation by setting the update direction 00:18:58.059 --> 00:19:00.440 to be the Nash cooperative bargaining solution. 00:19:00.480 --> 00:19:03.200 Okay. It forces the algorithms to realize that 00:19:03.200 --> 00:19:05.440 the only way they can optimize their own specific 00:19:05.440 --> 00:19:08.380 task is if they find a vector that simultaneously 00:19:08.380 --> 00:19:11.259 benefits the whole. That is incredible. It transforms 00:19:11.259 --> 00:19:13.799 a destructive tug of war into a synchronized 00:19:13.799 --> 00:19:16.319 cooperative negotiation. It really does. It stops 00:19:16.319 --> 00:19:18.799 the negative transfer because the system literally 00:19:18.799 --> 00:19:21.099 won't allow a mathematical move that actively 00:19:21.099 --> 00:19:24.000 destroys another task's progress. It finds the 00:19:24.000 --> 00:19:26.720 path of most mutual benefit. It is quite literally 00:19:26.720 --> 00:19:29.279 diplomacy translated into calculus. Well, we 00:19:29.279 --> 00:19:31.160 have covered a massive amount of ground today. 00:19:31.240 --> 00:19:33.450 We certainly have. To wrap this deep dive up, 00:19:33.549 --> 00:19:35.990 I want to bring all these abstract, soaring concepts 00:19:35.990 --> 00:19:38.109 back down to earth for you. Because this isn't 00:19:38.109 --> 00:19:40.150 just theory sitting in a lab somewhere. Not at 00:19:40.150 --> 00:19:44.220 all. the textless, sweeping, real -world applications 00:19:44.220 --> 00:19:47.079 for this technology. They are using multi -task 00:19:47.079 --> 00:19:49.559 optimization to accelerate complex engineering 00:19:49.559 --> 00:19:52.599 designs. They are pushing it into cloud computing, 00:19:52.900 --> 00:19:55.779 aiming for these massive on -demand optimization 00:19:55.779 --> 00:19:58.859 services that can cater to multiple different 00:19:58.859 --> 00:20:01.619 customers simultaneously without cross -contamination. 00:20:01.859 --> 00:20:04.059 And it's making major breakthroughs in industrial 00:20:04.059 --> 00:20:06.740 manufacturing and chemistry, too. Oh, chemistry. 00:20:07.000 --> 00:20:09.140 Yeah, they're optimizing chemical reactions by 00:20:09.140 --> 00:20:11.789 looking at the shape variables across multiple 00:20:11.789 --> 00:20:13.930 different experiments at the same time. Okay. 00:20:14.250 --> 00:20:16.450 So it's finding the underlying physics of the 00:20:16.450 --> 00:20:19.410 reactions faster than human trial and error ever 00:20:19.410 --> 00:20:21.589 could. That's amazing. And for the technically 00:20:21.589 --> 00:20:23.769 curious listener out there who wants to see how 00:20:23.769 --> 00:20:26.589 this actually runs, the source even mentions 00:20:26.589 --> 00:20:30.269 a specific MATLAB software package called MALSAR. 00:20:30.410 --> 00:20:32.650 Yes, MALSAR. Which stands for Multitask Learning 00:20:32.650 --> 00:20:35.759 Via Structural Regularization. It implements 00:20:35.759 --> 00:20:38.680 a whole suite of these algorithms, from clustered 00:20:38.680 --> 00:20:41.740 multitask learning to robust low -rank learning. 00:20:42.059 --> 00:20:43.940 So the tools to play with this are out there 00:20:43.940 --> 00:20:46.039 right now. They are. And I think there is a profound 00:20:46.039 --> 00:20:48.700 takeaway here that extends far beyond the code. 00:20:48.819 --> 00:20:50.980 Oh, definitely. We've seen today that artificial 00:20:50.980 --> 00:20:52.960 intelligence achieves its highest potential, 00:20:53.200 --> 00:20:55.759 its deepest understanding of the world, not by 00:20:55.759 --> 00:20:58.380 hyper -focusing on a single isolated problem 00:20:58.380 --> 00:21:01.339 in a vacuum. Right. It achieves brilliance by 00:21:01.339 --> 00:21:05.039 finding the hidden harmony. in conflicting, unrelated, 00:21:05.480 --> 00:21:08.240 and continuously evolving tasks through cooperative 00:21:08.240 --> 00:21:11.799 games. It's beautiful, really. It is. So as you 00:21:11.799 --> 00:21:14.160 go about your week, it's worth asking yourself. 00:21:15.259 --> 00:21:17.920 In an era of constant human information overload, 00:21:18.640 --> 00:21:20.740 are we trying too hard to single task our way 00:21:20.740 --> 00:21:23.420 through life? That's a great question. What auxiliary 00:21:23.420 --> 00:21:25.660 tasks or unrelated hobbies could you be picking 00:21:25.660 --> 00:21:27.940 up right now that might secretly help you screen 00:21:27.940 --> 00:21:28.740 out your own noise?