WEBVTT 00:00:00.000 --> 00:00:02.319 You know, when you picture how a machine learns, 00:00:02.580 --> 00:00:05.889 you... You probably imagine this massive cavernous 00:00:05.889 --> 00:00:08.310 library. I look totally. Just millions of books, 00:00:08.429 --> 00:00:11.929 billions of images, endless spreadsheets of data 00:00:11.929 --> 00:00:14.050 just being shoveled into a computer until it 00:00:14.050 --> 00:00:16.949 finally figures out how to recognize a cat or 00:00:16.949 --> 00:00:19.649 translate a sentence or diagnose a disease. Yeah, 00:00:19.649 --> 00:00:21.390 we tend to treat artificial intelligence like 00:00:21.390 --> 00:00:24.370 a speed reader that never sleeps. Right. Assuming 00:00:24.370 --> 00:00:26.230 the only way to teach it is by force feeding 00:00:26.230 --> 00:00:30.289 it absolutely everything we have. But welcome 00:00:30.289 --> 00:00:33.270 to today's deep dive. because if you're joining 00:00:33.270 --> 00:00:35.909 us, you are in for a pretty fascinating shift 00:00:35.909 --> 00:00:38.070 in perspective. Yeah, it really turns that whole 00:00:38.070 --> 00:00:40.649 idea upside down. It does. Today, our mission 00:00:40.649 --> 00:00:43.329 is to explore a comprehensive Wikipedia overview 00:00:43.329 --> 00:00:46.149 on a concept called active learning. We're going 00:00:46.149 --> 00:00:49.590 to look at how we can train AI, not by force 00:00:49.590 --> 00:00:52.350 feeding it massive amounts of data, but by teaching 00:00:52.350 --> 00:00:55.429 the AI to actually ask humans the right questions. 00:00:55.590 --> 00:00:58.009 Which is, I mean, it's a complete reversal of 00:00:58.009 --> 00:00:59.770 the brute force approach to education we're used 00:00:59.770 --> 00:01:02.310 to. Okay, let's unpack this. the fundamental 00:01:02.310 --> 00:01:04.530 problem we're looking at here is the snagging 00:01:04.530 --> 00:01:07.189 cost of human labor. Right, because for a long 00:01:07.189 --> 00:01:10.010 time, standard supervised learning relied entirely 00:01:10.010 --> 00:01:13.250 on the assumption that volume is the key to accuracy. 00:01:13.549 --> 00:01:15.590 Like just gather up all the answers up front, 00:01:15.689 --> 00:01:18.189 hand them to the algorithm, and just hope it 00:01:18.189 --> 00:01:20.709 eventually spots the underlying patterns. Exactly. 00:01:21.209 --> 00:01:24.090 But as the source material points out, that method 00:01:24.090 --> 00:01:27.870 hits a massive wall in the real world. We live 00:01:27.870 --> 00:01:30.750 in a world with oceans of raw, unlabeled data. 00:01:30.920 --> 00:01:34.120 It is literally everywhere. But paying human 00:01:34.120 --> 00:01:37.140 beings to sit there and manually label that data? 00:01:37.319 --> 00:01:39.060 Oh, it's a nightmare. Yeah, whether you're paying 00:01:39.060 --> 00:01:42.159 doctors to read thousands of x -rays or linguists 00:01:42.159 --> 00:01:44.560 to tag parts of speech in a million sentences, 00:01:44.920 --> 00:01:47.939 it's incredibly expensive, and it is agonizingly 00:01:47.939 --> 00:01:50.159 slow. So the question becomes, instead of paying 00:01:50.159 --> 00:01:52.900 experts to label a million random examples, what 00:01:52.900 --> 00:01:55.280 if we could teach the AI to look at the data 00:01:55.280 --> 00:01:57.900 and figure out the exact specific questions it 00:01:57.900 --> 00:02:00.480 needs to ask to learn the fastest? And that's 00:02:00.480 --> 00:02:02.120 the core definition of active learning, right? 00:02:02.140 --> 00:02:04.459 Yeah. It is an iterative supervised learning 00:02:04.459 --> 00:02:07.099 process. But instead of just passively receiving 00:02:07.099 --> 00:02:10.300 data, the algorithm interactively queries a human 00:02:10.300 --> 00:02:12.780 expert, which the literature calls the teacher 00:02:12.780 --> 00:02:16.039 or the Oracle Oracle. I love that. Yeah, it sounds 00:02:16.039 --> 00:02:19.020 very mystical. But the crucial detail here is 00:02:19.020 --> 00:02:21.580 that the algorithm isn't just asking for random 00:02:21.580 --> 00:02:24.919 labels. It is mathematically targeting the specific 00:02:24.919 --> 00:02:26.860 data points that will be the most informative 00:02:26.860 --> 00:02:28.639 to its current state of understanding, which 00:02:28.639 --> 00:02:32.120 means it learns faster, vastly faster. By choosing 00:02:32.120 --> 00:02:35.219 its own examples, the algorithm can learn a complex 00:02:35.219 --> 00:02:38.180 concept with a tiny fraction of the data points 00:02:38.180 --> 00:02:40.780 it would need in normal supervised learning. 00:02:40.939 --> 00:02:44.039 Wow. It speeds up the development process so 00:02:44.039 --> 00:02:46.580 dramatically that, in certain highly complex 00:02:46.580 --> 00:02:49.479 scenarios, an active learning model can essentially 00:02:49.479 --> 00:02:52.639 stand in for needing a quantum or supercomputer. 00:02:52.919 --> 00:02:55.419 Wait, okay, that is a staggering claim. Just 00:02:55.419 --> 00:02:57.300 by asking the right questions, you bypass the 00:02:57.300 --> 00:02:59.240 need for a supercomputer? In specific cases, 00:02:59.379 --> 00:03:02.539 yeah. That's wild. But I guess if we want to 00:03:02.539 --> 00:03:05.319 understand how the AI asks for help, we first 00:03:05.319 --> 00:03:07.520 have to understand how it maps out its own ignorance, 00:03:07.620 --> 00:03:09.500 you know? We have to look at the mathematics 00:03:09.500 --> 00:03:12.300 of the unknown. Right, and let's ground this 00:03:12.300 --> 00:03:15.020 in a real -world scenario from the sources. Imagine 00:03:15.020 --> 00:03:18.039 you are a biological researcher trying to engineer 00:03:18.300 --> 00:03:22.159 a new highly specific protein. Okay. The total 00:03:22.159 --> 00:03:25.139 universe of all possible data under consideration 00:03:25.139 --> 00:03:28.159 like every protein you could possibly test is 00:03:28.159 --> 00:03:30.819 represented mathematically by the letter T. Okay 00:03:30.819 --> 00:03:33.719 so T is the massive pile of everything. Everything. 00:03:33.939 --> 00:03:35.879 It includes the handful of proteins you've already 00:03:35.879 --> 00:03:38.580 tested in the lab plus the millions of variations 00:03:38.580 --> 00:03:40.680 you haven't even looked at yet. Correct. And 00:03:40.680 --> 00:03:43.180 remember active learning moves in cycles or iterations. 00:03:43.240 --> 00:03:46.159 Right. During every single iteration which the 00:03:46.159 --> 00:03:49.479 math labels as I, the algorithm's job is to take 00:03:49.479 --> 00:03:52.340 that massive universe of data, T, and break it 00:03:52.340 --> 00:03:54.800 down into three distinct separate subsets. Okay, 00:03:54.840 --> 00:03:57.599 so slicing up the pie. Exactly. The first subset 00:03:57.599 --> 00:03:59.599 is the data where the label is already known. 00:03:59.699 --> 00:04:03.219 We call this T sub K, I. So in our protein lab, 00:04:03.319 --> 00:04:06.300 T sub K, I is the small stack of lab reports 00:04:06.300 --> 00:04:08.819 sitting on your desk. The AI already has the 00:04:08.819 --> 00:04:10.840 answers for those. It knows exactly what those 00:04:10.840 --> 00:04:13.300 specific proteins do. Yes. Then you have the 00:04:13.300 --> 00:04:16.759 second subset, which is T sub U. U for unknown. 00:04:17.180 --> 00:04:20.240 Right. This is the vast ocean of data where the 00:04:20.240 --> 00:04:22.839 label is unknown. The millions of proteins we 00:04:22.839 --> 00:04:25.240 haven't tested, the AI is totally clueless about 00:04:25.240 --> 00:04:27.660 these. Which brings us to the third and really 00:04:27.660 --> 00:04:32.439 the most critical subset, T sub C, I. And C stands 00:04:32.439 --> 00:04:35.920 for chosen. Yes. These are the highly specific 00:04:35.920 --> 00:04:38.500 targeted data points that the AI plucks out of 00:04:38.500 --> 00:04:40.879 the unknown pile, hands to the human oracle, 00:04:41.100 --> 00:04:43.100 and basically says, please test this one next. 00:04:43.699 --> 00:04:46.939 Man, defining that chosen subset, figuring out 00:04:46.939 --> 00:04:50.300 exactly what goes into T sub C, that feels like 00:04:50.300 --> 00:04:52.420 the entire magic trick of this whole field. Oh, 00:04:52.439 --> 00:04:54.339 it is the entire battlefield of active learning 00:04:54.339 --> 00:04:56.899 research. The algorithm's sole objective is to 00:04:56.899 --> 00:04:59.259 extract the absolute maximum amount of learning 00:04:59.259 --> 00:05:01.860 from the minimum number of human queries. I kind 00:05:01.860 --> 00:05:03.920 of like to think of it like a student studying 00:05:03.920 --> 00:05:07.459 for a grueling final exam. Okay, how so? Well... 00:05:06.910 --> 00:05:08.889 A standard machine learning model is the student 00:05:08.889 --> 00:05:11.149 who insists on rereading the entire textbook 00:05:11.149 --> 00:05:13.089 from page one, even the chapters they already 00:05:13.089 --> 00:05:15.610 know perfectly. Right. Highly inefficient. Yeah. 00:05:16.290 --> 00:05:18.290 An active learning model is the smart student. 00:05:18.850 --> 00:05:21.290 They take a practice test, figure out, oh, I'm 00:05:21.290 --> 00:05:23.689 parable at calculus but great at algebra, and 00:05:23.689 --> 00:05:26.550 then they only take their absolute hardest calculus 00:05:26.550 --> 00:05:28.870 questions to the professor during office hours. 00:05:29.410 --> 00:05:31.610 That's actually a great analogy. But here is 00:05:31.610 --> 00:05:33.589 what I struggle with, and I really want to push 00:05:33.589 --> 00:05:36.430 back on this a bit. Go for it. If the AI doesn't 00:05:36.430 --> 00:05:39.209 know the answers to the unknown data, like if 00:05:39.209 --> 00:05:42.209 it's completely clueless about millions of proteins, 00:05:43.050 --> 00:05:45.230 how on earth does it know which questions are 00:05:45.230 --> 00:05:48.050 actually worth asking? Isn't there a massive 00:05:48.050 --> 00:05:50.990 risk that the AI just gets overwhelmed by totally 00:05:50.990 --> 00:05:54.069 useless, confusing, or uninformative examples? 00:05:54.350 --> 00:05:56.329 You're hitting on the biggest nightmare for developers 00:05:56.329 --> 00:05:58.750 in this space. If the algorithm asks the wrong 00:05:58.750 --> 00:06:01.139 questions, the whole system collapses. Because 00:06:01.139 --> 00:06:03.639 it's just spinning its wheels. Exactly. It wastes 00:06:03.639 --> 00:06:05.800 the oracle's time, it spends money, and it learns 00:06:05.800 --> 00:06:08.860 absolutely nothing. Avoiding uninformative examples 00:06:08.860 --> 00:06:11.819 is the whole crux of the field. The machine needs 00:06:11.819 --> 00:06:14.420 a way to evaluate the value of its own ignorance. 00:06:14.680 --> 00:06:16.439 Because if you're deploying this at scale, it's 00:06:16.439 --> 00:06:18.319 not just bothering one professor during office 00:06:18.319 --> 00:06:21.529 hours. Not at all. When you scale this up to 00:06:21.529 --> 00:06:24.129 large real -world projects where an AI needs 00:06:24.129 --> 00:06:26.550 to define its chosen subset at an industrial 00:06:26.550 --> 00:06:29.689 level It often plugs into human crowdsourcing 00:06:29.689 --> 00:06:32.310 frameworks like Amazon Mechanical Turk, right? 00:06:32.449 --> 00:06:35.149 We are talking about platforms where you might 00:06:35.149 --> 00:06:38.279 have hundreds or thousands of human beings in 00:06:38.279 --> 00:06:40.259 the active learning loop, sitting at their computers, 00:06:40.720 --> 00:06:43.660 and you are paying them per label. Right, actual 00:06:43.660 --> 00:06:46.740 cash. Yes. If your algorithm is feeding them 00:06:46.740 --> 00:06:49.439 useless, redundant questions, you are literally 00:06:49.439 --> 00:06:51.819 burning through your research budget for zero 00:06:51.819 --> 00:06:54.779 gain. So the AI is hunting for the perfect questions 00:06:54.779 --> 00:06:57.240 to ask us, because every question costs money. 00:06:57.399 --> 00:06:59.779 And according to the source material, there are 00:06:59.779 --> 00:07:02.500 three specific scenarios or environments where 00:07:02.500 --> 00:07:04.939 it conducts this hunt. Three ways it can ask 00:07:04.939 --> 00:07:06.639 a question. Let's look at the first one. Pool 00:07:06.639 --> 00:07:08.839 -based sampling. Pool -based sampling is currently 00:07:08.839 --> 00:07:11.600 the most well -known method out there. In this 00:07:11.600 --> 00:07:14.120 environment, the AI gets to evaluate the entire 00:07:14.120 --> 00:07:16.620 pool of unlabeled data at once before it makes 00:07:16.620 --> 00:07:18.519 any decisions. So it looks at the whole data 00:07:18.519 --> 00:07:21.160 set, say all million untested proteins, and it 00:07:21.160 --> 00:07:24.279 assigns a confidence score to every single one. 00:07:24.519 --> 00:07:27.360 Yes. So it's assigning a mathematical measurement 00:07:27.360 --> 00:07:30.199 of how well it thinks it understands that specific 00:07:30.199 --> 00:07:32.779 data point. Yeah. even without having the actual 00:07:32.779 --> 00:07:36.220 label. So it scores everything, sorts that massive 00:07:36.220 --> 00:07:39.100 list, selects the instances where its confidence 00:07:39.100 --> 00:07:42.120 is the absolute lowest, and then sends those 00:07:42.120 --> 00:07:45.220 specific, highly ambiguous points to the human 00:07:45.220 --> 00:07:47.879 teacher. Exactly. It sounds incredibly thorough, 00:07:48.180 --> 00:07:50.600 like it surveys the entire landscape before making 00:07:50.600 --> 00:07:53.360 a move. But evaluating a million things at once? 00:07:53.699 --> 00:07:56.699 That has to take a toll. Oh, it does. The theoretical 00:07:56.699 --> 00:07:59.620 limit to pool -based sampling is computer memory. 00:08:00.769 --> 00:08:02.910 Evaluating and holding confidence scores for 00:08:02.910 --> 00:08:05.529 billions of data points simultaneously requires 00:08:05.529 --> 00:08:08.050 massive computational overhead. I can imagine. 00:08:08.250 --> 00:08:10.370 Think about how your own laptop slows down when 00:08:10.370 --> 00:08:13.470 you have too many tabs open. Now imagine an AI 00:08:13.470 --> 00:08:15.750 trying to hold millions of complex mathematical 00:08:15.750 --> 00:08:17.730 evaluations in its active memory. Yeah, that 00:08:17.730 --> 00:08:19.629 would fry my computer. And even if you have the 00:08:19.629 --> 00:08:21.329 world's biggest computer, you still have to deal 00:08:21.329 --> 00:08:23.009 with the human on the other end. And that's the 00:08:23.009 --> 00:08:26.250 practical limit, human fatigue. The oracle is 00:08:26.250 --> 00:08:29.470 a person. They get tired. their attention wanes, 00:08:29.509 --> 00:08:32.070 and they require a paycheck. No matter how many 00:08:32.070 --> 00:08:34.870 brilliant, ambiguous points the machine finds 00:08:34.870 --> 00:08:38.090 in its massive pool, the human can only process 00:08:38.090 --> 00:08:40.289 a limited number before they start making mistakes. 00:08:40.669 --> 00:08:43.710 Okay, so if looking at the whole pool crashes 00:08:43.710 --> 00:08:46.330 your computer's memory, how do we fix that? Do 00:08:46.330 --> 00:08:49.789 we just, like, force the AI to put blinders on 00:08:49.789 --> 00:08:52.210 and look at data one piece at a time? That is 00:08:52.210 --> 00:08:54.370 exactly what the second scenario does. It's called 00:08:54.370 --> 00:08:57.590 stream -based selective sampling. Right, so if 00:08:57.590 --> 00:09:00.090 the pool -based method is a guy standing on a 00:09:00.090 --> 00:09:02.570 balcony looking at a massive crowd of data all 00:09:02.570 --> 00:09:05.490 at once, the stream -based method is a guy standing 00:09:05.490 --> 00:09:08.110 at a turnstile checking tickets one by one. That's 00:09:08.110 --> 00:09:10.330 a great way to put it. The AI looks at the data 00:09:10.330 --> 00:09:12.509 sequentially. For every single item that passes 00:09:12.509 --> 00:09:14.909 through, it makes a split -second decision. Do 00:09:14.909 --> 00:09:17.009 I label this myself, or do I ask the teacher? 00:09:17.259 --> 00:09:20.139 But wait, if the AI is making these rapid -fire 00:09:20.139 --> 00:09:22.519 decisions one by one, especially early on when 00:09:22.519 --> 00:09:24.820 it hasn't learned much yet, it has no idea what's 00:09:24.820 --> 00:09:26.720 coming next in the stream. Isn't it just flying 00:09:26.720 --> 00:09:29.820 blind? Yes. That is the major concession you 00:09:29.820 --> 00:09:32.139 make with stream -based sampling. Because it 00:09:32.139 --> 00:09:33.860 only sees a data point right in front of its 00:09:33.860 --> 00:09:36.700 nose, it doesn't know if a vastly more informative 00:09:36.700 --> 00:09:38.940 data point is coming 10 steps down the line. 00:09:38.940 --> 00:09:42.559 Oh, man. It lacks a global view of the data set's 00:09:42.559 --> 00:09:45.399 distribution. So it might panic and ask the human 00:09:45.399 --> 00:09:47.919 to label a slightly confusing data point now, 00:09:48.440 --> 00:09:51.039 completely unaware that a perfectly clear, incredibly 00:09:51.039 --> 00:09:52.960 educational data point was about to come through 00:09:52.960 --> 00:09:56.990 the turnstile a minute later. Exactly. And because 00:09:56.990 --> 00:09:59.309 it can't efficiently capitalize on the structure 00:09:59.309 --> 00:10:01.649 of the data, the algorithm tends to be a bit 00:10:01.649 --> 00:10:05.370 more needy. Needy AI. Yeah. It asks more questions 00:10:05.370 --> 00:10:07.830 overall, meaning the human teacher ends up spending 00:10:07.830 --> 00:10:10.669 more effort supplying labels compared to the 00:10:10.669 --> 00:10:13.029 pool -based approach. You save computer memory, 00:10:13.330 --> 00:10:16.190 but you might waste more human time. OK, so pool 00:10:16.190 --> 00:10:18.629 -based looks at everything but eats up memory. 00:10:18.919 --> 00:10:21.039 Stream -based looks at one thing at a time, but 00:10:21.039 --> 00:10:23.139 wastes human effort because it lacks context. 00:10:23.240 --> 00:10:25.940 But what if you are in a situation where you 00:10:25.940 --> 00:10:28.700 don't even have a pool or a stream? What if your 00:10:28.700 --> 00:10:31.399 starting data is almost non -existent? Then you 00:10:31.399 --> 00:10:33.419 need a different approach entirely. Here's where 00:10:33.419 --> 00:10:36.200 it gets really interesting. The third scenario. 00:10:36.899 --> 00:10:39.179 Membership query synthesis. This is a radical 00:10:39.179 --> 00:10:41.500 departure from the other two. In this scenario, 00:10:41.740 --> 00:10:44.519 the AI doesn't just look at existing data. It 00:10:44.519 --> 00:10:47.860 literally plays Frankenstein. It does. It uses 00:10:47.860 --> 00:10:50.759 generative AI techniques to synthesize entirely 00:10:50.759 --> 00:10:53.620 artificial data from scratch, and then asks the 00:10:53.620 --> 00:10:56.600 human to label its creation. The example from 00:10:56.600 --> 00:11:00.419 the sources is wild. The animal one. Yes. If 00:11:00.419 --> 00:11:02.620 the AI is trying to learn the visual boundary 00:11:02.620 --> 00:11:05.059 between humans and animals, it might generate 00:11:05.059 --> 00:11:08.159 a bizarre clipped synthetic image of a leg, a 00:11:08.159 --> 00:11:10.220 leg that doesn't actually exist anywhere in the 00:11:10.220 --> 00:11:12.340 real world, and present it to the human oracle 00:11:12.340 --> 00:11:16.080 asking, is this appendage human or animal? Theoretically, 00:11:16.220 --> 00:11:19.059 it's a brilliant concept. The AI can probe the 00:11:19.059 --> 00:11:21.879 exact boundaries of its own knowledge by manifesting 00:11:21.879 --> 00:11:23.899 the precise data point that would clear up its 00:11:23.899 --> 00:11:26.360 confusion. But the constraints of reality make 00:11:26.360 --> 00:11:29.379 this exceptionally difficult. The massive challenge 00:11:29.379 --> 00:11:32.139 of query synthesis is that the synthetic data 00:11:32.139 --> 00:11:35.419 must strictly obey the natural laws and constraints 00:11:35.419 --> 00:11:37.990 of the real world. Because the AI doesn't actually 00:11:37.990 --> 00:11:40.129 understand what a leg is, it just understands 00:11:40.129 --> 00:11:42.570 pixel values and mathematical distributions. 00:11:42.830 --> 00:11:44.529 Precisely. And the source material provides a 00:11:44.529 --> 00:11:47.389 fantastic medical example of where this generative 00:11:47.389 --> 00:11:51.309 process can fail completely. Imagine an AI generating 00:11:51.309 --> 00:11:54.529 a synthetic blood test to learn about diagnosing 00:11:54.529 --> 00:11:57.750 diseases. OK. In a real white blood cell differential 00:11:57.750 --> 00:12:00.210 test, a doctor looks at different components 00:12:00.210 --> 00:12:03.129 of white blood cells. Because these components 00:12:03.129 --> 00:12:05.970 are percentages of a whole, they most perfectly 00:12:05.970 --> 00:12:09.389 add up to exactly 100%. Right. Basic math. A 00:12:09.389 --> 00:12:11.950 human doctor knows this instinctively. But if 00:12:11.950 --> 00:12:14.629 the AI lacks that strict mathematical constraint, 00:12:15.070 --> 00:12:17.230 it might generate a synthetic blood test where 00:12:17.230 --> 00:12:20.370 the white blood cells add up to 110%. It's generating 00:12:20.370 --> 00:12:22.850 an impossibility. It's handing a doctor a patient 00:12:22.850 --> 00:12:25.330 profile that violates the laws of mathematics. 00:12:25.669 --> 00:12:27.889 And it gets even more complicated when variables 00:12:27.889 --> 00:12:30.710 are biologically dependent on one another. The 00:12:30.710 --> 00:12:33.929 source mentions two specific enzymes, ALT and 00:12:33.929 --> 00:12:36.490 AST. which doctors use to measure liver function. 00:12:36.549 --> 00:12:39.429 Okay, I've heard of those. There are strict physiological 00:12:39.429 --> 00:12:41.570 rules about how these enzymes behave together. 00:12:42.129 --> 00:12:44.950 If a patient is chronically ill, those enzymes 00:12:44.950 --> 00:12:49.009 elevate in specific ratios. So if the AI hallucinated 00:12:49.009 --> 00:12:51.809 a chronically ill patient profile where the AST 00:12:51.809 --> 00:12:54.409 enzyme is perfectly normal, sitting at the lower 00:12:54.409 --> 00:12:57.169 limit of 8 units per liter, but the ALT enzyme 00:12:57.169 --> 00:13:00.279 is wildly above normal, That's not just a weird 00:13:00.279 --> 00:13:03.019 patient. No, it's impossible. That is a physiological 00:13:03.019 --> 00:13:05.500 contradiction. It's like generating a synthetic 00:13:05.500 --> 00:13:07.620 house where the second floor is floating 10 feet 00:13:07.620 --> 00:13:10.139 above the first floor without any walls. The 00:13:10.139 --> 00:13:12.759 individual pieces exist, but the physics connecting 00:13:12.759 --> 00:13:16.179 them are fundamentally broken. Exactly. The AI's 00:13:16.179 --> 00:13:18.340 mathematical imagination has to be tightly bound 00:13:18.340 --> 00:13:21.919 by biological and physical reality. If it isn't, 00:13:22.100 --> 00:13:24.340 the query synthesis method just wastes the human 00:13:24.340 --> 00:13:26.580 expert time by asking them to diagnose nonsense. 00:13:26.840 --> 00:13:29.759 Okay, so the AI has its scenario. It's either 00:13:29.759 --> 00:13:32.539 evaluating a massive pool, watching a sequential 00:13:32.539 --> 00:13:35.440 stream, or synthesizing its own data from scratch. 00:13:35.840 --> 00:13:37.759 But that brings us to the core engine of the 00:13:37.759 --> 00:13:40.100 whole operation. The strategy. Right. What is 00:13:40.100 --> 00:13:43.200 the actual internal mathematical logic the AI 00:13:43.200 --> 00:13:45.679 uses to point at a piece of data and say, ah! 00:13:45.950 --> 00:13:48.149 This is the one. This is the data point that 00:13:48.149 --> 00:13:50.409 will teach me the most. You're talking about 00:13:50.409 --> 00:13:52.529 query strategies. Yeah. This is the internal 00:13:52.529 --> 00:13:55.129 brain of active learning. The literature organizes 00:13:55.129 --> 00:13:57.929 these into a few fascinating mathematical categories. 00:13:58.129 --> 00:14:00.710 And looking at the sources, some of these strategies 00:14:00.710 --> 00:14:03.289 sound incredibly intuitive, like expected error 00:14:03.289 --> 00:14:06.830 reduction. But how does an AI know a piece of 00:14:06.830 --> 00:14:09.590 data will reduce its errors if it doesn't even 00:14:09.590 --> 00:14:12.250 know what the data is yet? How does that simulation 00:14:12.250 --> 00:14:14.870 actually run? Well, it's a game of probabilities. 00:14:15.159 --> 00:14:18.320 The AI looks at an unlabeled data point and calculates 00:14:18.320 --> 00:14:21.139 the probability of all possible answers. It basically 00:14:21.139 --> 00:14:23.679 says to itself, OK, there is a 70 % chance the 00:14:23.679 --> 00:14:26.120 human will label this a cat and a 30 % chance 00:14:26.120 --> 00:14:28.799 they will label it a dog. OK. If they say cat, 00:14:29.059 --> 00:14:31.259 I calculate that my overall model accuracy will 00:14:31.259 --> 00:14:34.159 improve by 5%. If they say dog, my accuracy improves 00:14:34.159 --> 00:14:37.600 by 2%. It mathematically multiplies those probabilities 00:14:37.600 --> 00:14:40.220 to get an expected average drop in its error 00:14:40.220 --> 00:14:42.720 rate. Oh. Yeah, it runs that complex probability 00:14:42.720 --> 00:14:46.919 simulation for every single unlabeled point and 00:14:46.919 --> 00:14:49.019 asks for the one with the highest expected payoff. 00:14:49.159 --> 00:14:52.039 That is mind blowing. It's mapping out all the 00:14:52.039 --> 00:14:54.159 alternate futures of what the human might say 00:14:54.159 --> 00:14:57.379 and choosing the future with the best educational 00:14:57.379 --> 00:15:00.289 return on investment. And the sources also mentioned 00:15:00.289 --> 00:15:02.009 variance reduction, which is similar, right? 00:15:02.210 --> 00:15:05.330 Yes. But instead of focusing on error, variance 00:15:05.330 --> 00:15:07.330 reduction seeks out the data point that will 00:15:07.330 --> 00:15:09.809 minimize the mathematical spread or variance 00:15:09.809 --> 00:15:13.149 of the AI's own uncertainty. It wants to tighten 00:15:13.149 --> 00:15:15.250 its confidence intervals across the board. OK. 00:15:15.549 --> 00:15:17.929 But my absolute favorite strategy mentioned in 00:15:17.929 --> 00:15:20.009 the deep dive sources is something called query 00:15:20.009 --> 00:15:22.809 by committee. Oh, the committee approach is highly 00:15:22.809 --> 00:15:26.620 effective. Yeah. I picture it like a hung jury 00:15:26.620 --> 00:15:29.080 in a courtroom. Instead of just training one 00:15:29.080 --> 00:15:31.740 single AI model, you train a variety of different 00:15:31.740 --> 00:15:33.759 models on the limited data you already have. 00:15:33.960 --> 00:15:36.139 They become your committee. Then you feed them 00:15:36.139 --> 00:15:38.500 an unlabeled piece of data and ask them all to 00:15:38.500 --> 00:15:40.700 vote on what they think it is. If all the models 00:15:40.700 --> 00:15:42.759 in the committee agree, if they all look at an 00:15:42.759 --> 00:15:45.000 image and say, That's definitely a stop sign. 00:15:45.379 --> 00:15:47.759 Great. You don't need to ask the human. But if 00:15:47.759 --> 00:15:49.860 the committee is completely split. Right. If 00:15:49.860 --> 00:15:52.320 half the models scream, it's a stop sign, and 00:15:52.320 --> 00:15:54.299 the other half scream, it's a speed limit sign, 00:15:54.960 --> 00:15:57.620 that massive internal disagreement triggers an 00:15:57.620 --> 00:16:00.340 alarm. That specific data point is the one you 00:16:00.340 --> 00:16:02.440 send to the human oracle, because it is clearly 00:16:02.440 --> 00:16:04.740 sitting right on the chaotic boundary of your 00:16:04.740 --> 00:16:07.039 model's understanding. What's fascinating here 00:16:07.039 --> 00:16:09.340 is how that boundary actually looks when we map 00:16:09.340 --> 00:16:11.799 it out mathematically. And that brings us to 00:16:11.799 --> 00:16:14.580 a specific geometrical query strategy called 00:16:14.580 --> 00:16:17.340 the minimum marginal hyperplane. OK, minimum 00:16:17.340 --> 00:16:19.200 marginal hyperplane. That sounds like a warp 00:16:19.200 --> 00:16:21.019 drive from a sci -fi movie. We definitely need 00:16:21.019 --> 00:16:23.000 to break that down. It sounds intimidating, I 00:16:23.000 --> 00:16:26.809 know, but it's quite elegant visually. To understand 00:16:26.809 --> 00:16:29.549 it, we first need to quickly define what a support 00:16:29.549 --> 00:16:33.929 vector machine, or SVM, is. An SVM is a classic 00:16:33.929 --> 00:16:36.809 classification algorithm. Its entire job is to 00:16:36.809 --> 00:16:39.269 look at data and draw a rigid boundary between 00:16:39.269 --> 00:16:42.120 different categories. Okay, so imagine a massive 00:16:42.120 --> 00:16:45.019 multi -dimensional room. On one side of the room, 00:16:45.120 --> 00:16:47.720 you have all your labeled known cats. On the 00:16:47.720 --> 00:16:49.940 other side, you have all your labeled known dogs. 00:16:50.399 --> 00:16:52.940 The SVM's job is to draw a line straight down 00:16:52.940 --> 00:16:55.120 the middle of the room to separate them perfectly. 00:16:55.559 --> 00:16:57.820 Exactly. And in a high dimensional mathematical 00:16:57.820 --> 00:17:00.379 space, that dividing line isn't just a simple 00:17:00.379 --> 00:17:03.240 line drawn on a piece of paper. It's a flat surface 00:17:03.240 --> 00:17:06.380 called a hyperplane. I am picturing a giant sheet 00:17:06.380 --> 00:17:08.839 of glass cutting the room completely in half. 00:17:09.160 --> 00:17:11.319 Cats on the left, dogs on the right. That's a 00:17:11.319 --> 00:17:14.180 perfect visual. I know. You introduced your unknown, 00:17:14.400 --> 00:17:16.539 unlabeled data points. They scatter all over 00:17:16.539 --> 00:17:19.180 the room. The algorithm calculates the physical 00:17:19.180 --> 00:17:22.119 distance from every single unlabeled point to 00:17:22.119 --> 00:17:24.440 that glass wall, the separating hyperplane. OK. 00:17:24.720 --> 00:17:26.700 That distance is mathematically referred to as 00:17:26.700 --> 00:17:29.900 the margin, or W. So if an unlabeled data point 00:17:29.900 --> 00:17:32.599 has a massive W, it means it is sitting far, 00:17:32.700 --> 00:17:35.099 far away in the back corner of the room, deep 00:17:35.099 --> 00:17:38.039 in cat territory. The AI is highly confident 00:17:38.039 --> 00:17:40.349 it's a cat. It doesn't need to bother the human 00:17:40.349 --> 00:17:43.710 oracle to ask. Precisely. But the minimum marginal 00:17:43.710 --> 00:17:46.430 hyperplane method does the exact opposite. It 00:17:46.430 --> 00:17:48.849 hunts specifically for the data points with the 00:17:48.849 --> 00:17:52.250 smallest w. These are the points whose noses 00:17:52.250 --> 00:17:53.910 are practically pressed right up against the 00:17:53.910 --> 00:17:56.470 glass. They are the ultimate fence setters. The 00:17:56.470 --> 00:17:59.839 SVM is agonizingly uncertain about them because 00:17:59.839 --> 00:18:02.019 they are resting right on the razor's edge of 00:18:02.019 --> 00:18:04.220 the two categories. So by sending those specific 00:18:04.220 --> 00:18:06.579 points, the ones with the smallest margin to 00:18:06.579 --> 00:18:09.640 the human oracle for labeling, the AI rapidly 00:18:09.640 --> 00:18:12.819 clarifies the exact position and angle of that 00:18:12.819 --> 00:18:15.299 dividing wall. It's just brilliant. It completely 00:18:15.299 --> 00:18:18.220 ignores the easy stuff and zeros in exclusively 00:18:18.220 --> 00:18:20.980 on the most difficult borderline cases. It is 00:18:20.980 --> 00:18:23.849 incredibly efficient. But with all these different 00:18:23.849 --> 00:18:26.049 strategies we've discussed, expected error reduction, 00:18:26.410 --> 00:18:28.670 variance simulations, query by committee, minimum 00:18:28.670 --> 00:18:30.750 marginal hyperplanes, it raises an important 00:18:30.750 --> 00:18:33.529 question. How does an engineer even know which 00:18:33.529 --> 00:18:36.109 one to pick? Yeah. If you are building a model 00:18:36.109 --> 00:18:39.509 to read complex satellite imagery, do you use 00:18:39.509 --> 00:18:41.569 a committee of models or do you use a hyperplane? 00:18:41.849 --> 00:18:44.230 I imagine historically that just required a lot 00:18:44.230 --> 00:18:46.150 of trial and error by the human engineer. It 00:18:46.150 --> 00:18:49.390 did. The engineer had to guess which heuristic 00:18:49.390 --> 00:18:52.319 would work best for their specific data. But 00:18:52.319 --> 00:18:54.480 the cutting edge of the field right now is stepping 00:18:54.480 --> 00:18:57.960 beyond manual selection. We're moving into meta 00:18:57.960 --> 00:18:59.779 -learning. Algorithms learning how to learn. 00:19:00.019 --> 00:19:02.619 Yes. We're now using machine learning algorithms 00:19:02.619 --> 00:19:06.240 to evaluate the data environment itself and autonomously 00:19:06.240 --> 00:19:10.039 learn the optimal active learning strategy. We're 00:19:10.039 --> 00:19:12.779 literally training AI to figure out the best 00:19:12.779 --> 00:19:16.160 possible way to ask humans for help. That's incredible. 00:19:16.359 --> 00:19:18.599 It is a recursive layer of optimization that 00:19:18.599 --> 00:19:21.019 is pushing the efficiency of these systems to 00:19:21.019 --> 00:19:23.220 unprecedented levels. So what does this all mean 00:19:23.220 --> 00:19:25.799 for you listening to this right now? When you 00:19:25.799 --> 00:19:27.900 step back from the hyperplanes and the probability 00:19:27.900 --> 00:19:30.359 formulas... What we've really explored today 00:19:30.359 --> 00:19:32.500 is a profound shift in our relationship with 00:19:32.500 --> 00:19:34.920 technology. Yeah, a huge shift. We started with 00:19:34.920 --> 00:19:38.079 the bottleneck of expensive human labor, staring 00:19:38.079 --> 00:19:42.240 down oceans of raw, unlabeled data. And the solution 00:19:42.240 --> 00:19:44.720 wasn't just building a faster, hungrier machine 00:19:44.720 --> 00:19:47.700 that consumes more power. The solution was transforming 00:19:47.700 --> 00:19:51.039 the AI into an inquisitive student, a student 00:19:51.039 --> 00:19:53.539 that mathematically calculates its own uncertainty, 00:19:54.039 --> 00:19:56.299 that can intelligently manage its own memory, 00:19:56.619 --> 00:19:59.160 that respects the constraints of bio reality, 00:19:59.500 --> 00:20:02.380 and that strategically queries human experts 00:20:02.380 --> 00:20:04.920 to build optimal models faster than we ever thought 00:20:04.920 --> 00:20:07.740 possible. Exactly. It means the future of technology 00:20:07.740 --> 00:20:10.460 isn't just about passively feeding data to machines. 00:20:10.559 --> 00:20:13.519 It's a dynamic, interactive dialogue between 00:20:13.519 --> 00:20:16.519 human expertise and machine efficiency. It is 00:20:16.519 --> 00:20:19.420 a powerful dialogue, but if we connect this to 00:20:19.420 --> 00:20:22.079 the bigger picture, it does leave us with a critical, 00:20:22.240 --> 00:20:24.579 lingering question, one that the field is only 00:20:24.579 --> 00:20:27.220 just beginning to grapple with. Ooh, what's that? 00:20:27.519 --> 00:20:30.099 If an active learning AI is explicitly programmed 00:20:30.099 --> 00:20:32.220 to hunt down only the most confusing, boundary 00:20:32.220 --> 00:20:34.720 -pushing, and ambiguous edge cases to learn from, 00:20:35.240 --> 00:20:37.200 what happens to its understanding of the mundane? 00:20:37.380 --> 00:20:40.240 If an AI is only ever trained by looking at the 00:20:40.240 --> 00:20:44.400 most bizarre, overlapping, complex examples at 00:20:44.400 --> 00:20:47.480 the very margins of our world, Could it eventually 00:20:47.480 --> 00:20:50.059 develop a fundamentally skewed perspective of 00:20:50.059 --> 00:20:52.640 what normal reality actually looks like? That 00:20:52.640 --> 00:20:55.799 is wild. It's like if a doctor only ever studied 00:20:55.799 --> 00:20:58.259 the rarest, most incredibly complicated diseases 00:20:58.259 --> 00:21:00.900 on Earth, and then suddenly couldn't recognize 00:21:00.900 --> 00:21:03.279 a common cold because they never spent any time 00:21:03.279 --> 00:21:05.559 looking at one. Exactly. The smartest student 00:21:05.559 --> 00:21:08.299 in the room, only asking the hardest questions, 00:21:08.759 --> 00:21:10.779 but maybe losing sight of the baseline entirely. 00:21:11.189 --> 00:21:13.670 That is a fascinating and slightly terrifying 00:21:13.670 --> 00:21:15.609 place to leave it. Thank you for joining us on 00:21:15.609 --> 00:21:17.690 this deep dive into the source material. We will 00:21:17.690 --> 00:21:18.369 catch you next time.