WEBVTT 00:00:00.000 --> 00:00:03.160 So imagine you've spent, I don't know, the last 00:00:03.160 --> 00:00:06.320 five years meticulously training a spam filter 00:00:06.320 --> 00:00:10.619 on your personal inbox. You've flagged the newsletters, 00:00:11.179 --> 00:00:13.880 the weird sales pitches, the phishing scams, 00:00:13.880 --> 00:00:17.600 and your AI is just a flawless digital bouncer 00:00:17.600 --> 00:00:21.079 at this point. It knows exactly what you personally 00:00:21.079 --> 00:00:23.679 consider junk. Yeah, it's highly tailored to 00:00:23.679 --> 00:00:26.640 you. Exactly. Now imagine taking that exact same 00:00:26.640 --> 00:00:29.679 perfectly trained AI and just plugging it into 00:00:29.679 --> 00:00:31.859 the inbox of an executive at a completely different 00:00:31.859 --> 00:00:35.560 company in a completely different industry. What 00:00:35.560 --> 00:00:37.920 actually happens? Well, I mean, it fails almost 00:00:37.920 --> 00:00:40.259 immediately. Really? Just right away? Yeah, pretty 00:00:40.259 --> 00:00:42.299 much. It starts letting in obvious corporate 00:00:42.299 --> 00:00:45.259 spam, and it's probably blocking critical client 00:00:45.259 --> 00:00:47.780 emails. The model is effectively blind in this 00:00:47.780 --> 00:00:49.700 new environment. Okay, let's unpack this for 00:00:49.700 --> 00:00:51.539 a second. Because to you, the listener, and to 00:00:51.539 --> 00:00:54.740 me, we think, why can't a smart AI just automatically 00:00:54.740 --> 00:00:56.579 work everywhere? Right, that's the expectation. 00:00:56.840 --> 00:00:59.859 Like, if it knows what spam is, why does changing 00:00:59.859 --> 00:01:02.810 the user matter so much? We expect these algorithms 00:01:02.810 --> 00:01:05.709 to be universally intelligent, but we're doing 00:01:05.709 --> 00:01:08.950 a deep dive today into a Wikipedia article on 00:01:08.950 --> 00:01:12.030 a field called domain adaptation. And it paints 00:01:12.030 --> 00:01:14.230 a very different picture. It really does. It 00:01:14.230 --> 00:01:16.430 basically suggests that machine learning models 00:01:16.430 --> 00:01:19.689 are incredibly literal. Like, they only know 00:01:19.689 --> 00:01:21.870 the exact reality they were raised in. Yeah, 00:01:21.930 --> 00:01:23.450 that's a great way to put it. And to understand 00:01:23.450 --> 00:01:25.989 why they're so literal, we kind of have to look 00:01:25.989 --> 00:01:29.170 at how these models are formally structured at 00:01:29.170 --> 00:01:32.040 a mathematical level. OK, lay it on us. So in 00:01:32.040 --> 00:01:33.620 standard supervised machine learning, you have 00:01:33.620 --> 00:01:35.620 two main components. You have an input space, 00:01:35.700 --> 00:01:38.700 which we represent as X, and a label space represented 00:01:38.700 --> 00:01:41.500 as Y. So just tying that back to our spam filter, 00:01:41.579 --> 00:01:44.459 X is the raw text of the incoming email, and 00:01:44.459 --> 00:01:47.140 Y is simply the label, right, like spam or not 00:01:47.140 --> 00:01:49.920 spam. Exactly. And the entire objective of the 00:01:49.920 --> 00:01:52.379 algorithm is to learn a mathematical model, which 00:01:52.379 --> 00:01:55.480 we call a hypothesis, or H. A hypothesis. Got 00:01:55.480 --> 00:01:58.120 it. Right. And this hypothesis has one job. It 00:01:58.120 --> 00:02:01.049 looks at an example from input space X. and it 00:02:01.049 --> 00:02:04.010 attaches the correct label from space Y. So H 00:02:04.010 --> 00:02:06.890 is basically the internal rulebook the AI develops 00:02:06.890 --> 00:02:10.289 to map the text to the right label. Precisely. 00:02:10.650 --> 00:02:13.189 The model writes that rulebook by studying a 00:02:13.189 --> 00:02:16.770 training sample, but this points us directly 00:02:16.770 --> 00:02:19.629 to the fatal flaw of standard machine learning. 00:02:20.050 --> 00:02:23.030 Which is what? Well, standard supervised learning 00:02:23.030 --> 00:02:26.370 operates on this massive assumption. It assumes 00:02:26.370 --> 00:02:28.389 that all the examples it learns from are drawn 00:02:28.389 --> 00:02:31.300 from a specific static source distribution. We 00:02:31.300 --> 00:02:34.039 call that D sub S. D sub S. Okay. Yeah. And the 00:02:34.039 --> 00:02:36.719 algorithm essentially hardwires itself to the 00:02:36.719 --> 00:02:39.000 belief that every single piece of data it encounters 00:02:39.000 --> 00:02:41.340 in the future will be drawn perfectly from that 00:02:41.340 --> 00:02:44.039 exact same D sub S distribution. Oh, wow. So 00:02:44.039 --> 00:02:45.500 it just assumes the universe will never change. 00:02:45.580 --> 00:02:48.000 Like if the AI was trained on my personal emails, 00:02:48.280 --> 00:02:50.400 it assumes the entire world communicates exactly 00:02:50.400 --> 00:02:52.719 like my friends and family do. Exactly. It assumes 00:02:52.719 --> 00:02:54.759 the future will be a perfect statistical mirror 00:02:54.759 --> 00:02:58.099 of the past. But, you know, The real world is 00:02:58.099 --> 00:03:00.240 messy and it's constantly shifting. And this 00:03:00.240 --> 00:03:03.000 is where domain adaptation comes into play. Domain 00:03:03.000 --> 00:03:05.960 adaptation introduces a second related distribution, 00:03:06.419 --> 00:03:10.000 the target domain, or D sub T. D sub T target 00:03:10.000 --> 00:03:12.120 domain. Right. So the engineering challenge is 00:03:12.120 --> 00:03:14.139 transferring knowledge from the source domain, 00:03:14.400 --> 00:03:17.860 D sub S, to the target domain, D sub T, while 00:03:17.860 --> 00:03:19.580 committing the least amount of error possible 00:03:19.580 --> 00:03:22.360 along the way. So we're essentially forcing the 00:03:22.360 --> 00:03:25.560 AI to step into a new reality. Yes, exactly. 00:03:25.879 --> 00:03:27.919 And to give this some context, within the broader 00:03:27.919 --> 00:03:30.780 field of AI, researchers Pan and Yang developed 00:03:30.780 --> 00:03:33.199 a taxonomy for transfer learning back in 2010. 00:03:33.680 --> 00:03:35.960 They classified domain adaptation specifically 00:03:35.960 --> 00:03:39.319 as transductive transfer learning. Transductive. 00:03:39.699 --> 00:03:42.319 If the core task like catching spam remains the 00:03:42.319 --> 00:03:45.000 same, how does transductive differ from just, 00:03:45.000 --> 00:03:47.139 you know, standard learning? Well, it means the 00:03:47.139 --> 00:03:49.439 objective hasn't changed, but the marginal distributions 00:03:49.439 --> 00:03:51.580 of the data have. Meaning what, exactly? In other 00:03:51.580 --> 00:03:53.659 words, the overall demographic makeup of the 00:03:53.659 --> 00:03:56.800 data, say the total volume of emails or the ratio 00:03:56.800 --> 00:03:59.139 of newsletters to direct messages that differs 00:03:59.139 --> 00:04:02.199 between the two domains. Oh, I see. Furthermore... 00:04:01.759 --> 00:04:04.319 In transductive transfer learning, we assume 00:04:04.319 --> 00:04:08.020 we have absolutely zero labeled data for the 00:04:08.020 --> 00:04:10.379 new target domain. We only have the raw inputs. 00:04:10.560 --> 00:04:13.460 Just the raw x space. Right. Which makes it distinctly 00:04:13.460 --> 00:04:15.400 different from inductive transfer learning, where 00:04:15.400 --> 00:04:18.459 you actually have labels for the new task, or 00:04:18.459 --> 00:04:20.500 unsupervised transfer learning, where you have 00:04:20.500 --> 00:04:23.360 no labels in... either domain. OK, so standard 00:04:23.360 --> 00:04:26.420 machine learning is basically like studying for 00:04:26.420 --> 00:04:29.079 a test using only past exams, banking on the 00:04:29.079 --> 00:04:30.920 teacher, never changing their testing style. 00:04:31.019 --> 00:04:33.540 Yeah, that's a good analogy. But domain adaptation 00:04:33.540 --> 00:04:36.019 is preparing for the teacher to throw a massive 00:04:36.019 --> 00:04:39.759 curveball. But I have to push back on the premise 00:04:39.759 --> 00:04:41.790 here for a second. Sure, go ahead. If the new 00:04:41.790 --> 00:04:44.149 executive's inbox, the target domain, is fundamentally 00:04:44.149 --> 00:04:46.750 different, why try to salvage the old model at 00:04:46.750 --> 00:04:50.129 all? Building a new AI from scratch for the new 00:04:50.129 --> 00:04:52.389 inbox just seems like a cleaner solution, doesn't 00:04:52.389 --> 00:04:55.449 it? You'd think so, but starting from scratch 00:04:55.449 --> 00:04:58.209 is prohibitively expensive. Really? Oh yeah. 00:04:58.629 --> 00:05:00.870 It requires a massive amount of human labor to 00:05:00.870 --> 00:05:03.529 create a new, perfectly labeled data set for 00:05:03.529 --> 00:05:06.870 every single new user or new hospital or new 00:05:06.870 --> 00:05:10.279 city an AI is deployed in. Ah, right, all that 00:05:10.279 --> 00:05:13.300 manual tagging. Exactly. If we can salvage the 00:05:13.300 --> 00:05:15.879 foundational knowledge from the source domain, 00:05:16.500 --> 00:05:19.180 the basic understanding of Inglis syntax, the 00:05:19.180 --> 00:05:21.360 underlying psychological structure of a scam, 00:05:22.000 --> 00:05:24.120 and mathematically adapt it to the target domain, 00:05:24.350 --> 00:05:27.209 we save immense amounts of time and computational 00:05:27.209 --> 00:05:31.170 power. So we're recycling the AI's core intelligence 00:05:31.170 --> 00:05:33.290 rather than just throwing it in the trash. That 00:05:33.290 --> 00:05:34.930 makes sense. Exactly. But since we aren't starting 00:05:34.930 --> 00:05:37.250 from scratch, we have to isolate exactly what 00:05:37.250 --> 00:05:39.089 broke during the transition from the source to 00:05:39.089 --> 00:05:41.149 the target, right? Like, what actually changes 00:05:41.149 --> 00:05:43.889 in the data to confuse the AI so badly? Well, 00:05:43.970 --> 00:05:46.209 the source material outlines three common types 00:05:46.209 --> 00:05:49.189 of distribution shifts. The first, and probably 00:05:49.189 --> 00:05:51.329 the most straightforward, is called covariate 00:05:51.329 --> 00:05:54.949 shift. And in covariate shift, the input distributions 00:05:54.949 --> 00:05:58.589 change, but the fundamental mathematical relationship 00:05:58.589 --> 00:06:01.410 between the inputs and the labels remains perfectly 00:06:01.410 --> 00:06:04.970 unchanged. applying that to our spam filter, 00:06:05.110 --> 00:06:07.449 the new executive receives totally different 00:06:07.449 --> 00:06:10.129 types of emails than I do. The input distribution, 00:06:10.389 --> 00:06:13.629 like the vocabulary, the senders, the frequency 00:06:13.629 --> 00:06:16.110 of messages, it's completely different. But the 00:06:16.110 --> 00:06:18.629 definition of the label spam hasn't actually 00:06:18.629 --> 00:06:21.850 changed. Like a phishing scam trying to steal 00:06:21.850 --> 00:06:24.470 a password is still a phishing scam, whether 00:06:24.470 --> 00:06:27.959 it's aimed at me. or a CEO. Exactly. The frequency 00:06:27.959 --> 00:06:30.779 and style of the features changed, but the underlying 00:06:30.779 --> 00:06:33.279 truth did not. OK, got it. What's the second 00:06:33.279 --> 00:06:35.740 shift? The second type is prior shift, also known 00:06:35.740 --> 00:06:38.779 as label shift. This occurs when the label distribution 00:06:38.779 --> 00:06:41.420 differs between the data sets, but the conditional 00:06:41.420 --> 00:06:43.720 distribution of features stays the same. Label 00:06:43.720 --> 00:06:45.699 shift. Give me an example of that. The Wikipedia 00:06:45.699 --> 00:06:48.040 article actually uses a highly illustrative example 00:06:48.040 --> 00:06:51.600 for this. Imagine an AI trained to classify hair 00:06:51.600 --> 00:06:54.439 color in images, and it's trained exclusively 00:06:54.439 --> 00:06:57.279 on a data dataset from Italy. That's our source 00:06:57.279 --> 00:06:59.319 domain. Okay, so we can assume a much higher 00:06:59.319 --> 00:07:02.180 proportion of darker hair colors in that source 00:07:02.180 --> 00:07:04.519 dataset. Correct. Now we take that exact same 00:07:04.519 --> 00:07:07.079 model and deploy it in Norway, our target domain. 00:07:07.160 --> 00:07:09.930 Oh wow, okay. Right. The proportion of blonde 00:07:09.930 --> 00:07:12.370 to black hair differs wildly between these two 00:07:12.370 --> 00:07:15.209 populations. So the prior probability of the 00:07:15.209 --> 00:07:17.449 overall label distribution has completely flipped. 00:07:17.709 --> 00:07:19.449 But the conditional distribution meaning the 00:07:19.449 --> 00:07:21.910 specific physical traits conditionally tied to 00:07:21.910 --> 00:07:24.610 the label blonde that stays the same. Exactly. 00:07:24.750 --> 00:07:26.970 Like a blonde person in Italy looks fundamentally 00:07:26.970 --> 00:07:29.769 like a blonde person in Norway. The physical 00:07:29.769 --> 00:07:32.410 features of blondness haven't changed, only how 00:07:32.410 --> 00:07:35.649 often the AI encounters them. Precisely. And 00:07:35.649 --> 00:07:38.269 if the AI deployed in Norway is made aware of 00:07:38.269 --> 00:07:41.350 this shift in population proportions, it can 00:07:41.350 --> 00:07:43.750 exploit that prior knowledge to dramatically 00:07:43.750 --> 00:07:46.129 improve its estimates. It just needs to adjust 00:07:46.129 --> 00:07:48.769 its internal baseline to expect more blonde hair. 00:07:48.959 --> 00:07:50.980 So it just updates its baseline assumptions, 00:07:51.060 --> 00:07:53.199 that makes sense. But that leads us to the third 00:07:53.199 --> 00:07:56.180 shift, which honestly seems inherently more destructive. 00:07:56.379 --> 00:07:58.680 Yeah, the third is concept shift, or conditional 00:07:58.680 --> 00:08:01.000 shift. And this is where it gets really tricky. 00:08:01.480 --> 00:08:03.420 This is where the relationship between the features 00:08:03.420 --> 00:08:06.000 and the labels completely changes, even if the 00:08:06.000 --> 00:08:08.259 input distribution looks identical. Wait, wait. 00:08:08.420 --> 00:08:11.459 So concept shift means the literal rules of reality 00:08:11.459 --> 00:08:15.560 changed for the AI. The mapping of x to y is 00:08:15.560 --> 00:08:18.819 entirely rewritten. Yes. Think of medical diagnosis. 00:08:19.579 --> 00:08:21.980 You might have the exact same symptoms. Let's 00:08:21.980 --> 00:08:24.240 say a high fever and a specific type of rash. 00:08:24.560 --> 00:08:26.899 Those are your inputs. OK. But depending on the 00:08:26.899 --> 00:08:28.980 population or the geographic region you're in 00:08:28.980 --> 00:08:31.500 your domain, those exact same symptoms might 00:08:31.500 --> 00:08:33.840 indicate entirely different diseases. Oh, wow. 00:08:34.340 --> 00:08:36.840 The same symptom means a different disease. How 00:08:36.840 --> 00:08:39.379 does a model not just completely self -destruct 00:08:39.379 --> 00:08:41.799 when the facts change like that? I mean, if A... 00:08:41.909 --> 00:08:44.710 no longer equals B, everything it learned is 00:08:44.710 --> 00:08:46.450 fundamentally flawed. Well, it is the hardest 00:08:46.450 --> 00:08:50.250 shift to solve. And honestly, often, models do 00:08:50.250 --> 00:08:52.350 fail catastrophically if they aren't aggressively 00:08:52.350 --> 00:08:54.570 corrected. I can imagine. It requires the model 00:08:54.570 --> 00:08:57.889 to actively unlearn a deeply ingrained correlation 00:08:57.889 --> 00:09:00.629 and map those same inputs to an entirely new 00:09:00.629 --> 00:09:03.190 output based on the new environment. If it doesn't, 00:09:03.289 --> 00:09:05.450 it will confidently make the wrong diagnosis 00:09:05.450 --> 00:09:08.980 every single time. Man. Because concept ships 00:09:08.980 --> 00:09:12.159 fundamentally break the model's logic, engineers 00:09:12.159 --> 00:09:14.559 can't just feed it more data, right? They have 00:09:14.559 --> 00:09:17.460 to change how the model learns. Which I guess 00:09:17.460 --> 00:09:21.059 brings us to the data scenarios. How engineers 00:09:21.059 --> 00:09:23.620 perform this corrective surgery depends entirely 00:09:23.620 --> 00:09:25.460 on what resources they have in the target domain, 00:09:25.559 --> 00:09:27.799 doesn't it? It does. The data scenarios dictate 00:09:27.799 --> 00:09:29.759 the mathematical approach, and they fall into 00:09:29.759 --> 00:09:33.899 three distinct categories. The first is unsupervised 00:09:33.899 --> 00:09:37.370 domain adaptation. Unsupervised, meaning no labels. 00:09:37.649 --> 00:09:40.169 Exactly. This means we have an abundance of data 00:09:40.169 --> 00:09:42.509 from the target domain, but absolutely none of 00:09:42.509 --> 00:09:44.570 it is labeled. So we have a decade's worth of 00:09:44.570 --> 00:09:46.789 the new executives' emails, but nobody has gone 00:09:46.789 --> 00:09:49.809 through and manually categorized them. The AI 00:09:49.809 --> 00:09:51.850 has to look at the unlabeled patterns in the 00:09:51.850 --> 00:09:54.029 target domain and mathematically compare them 00:09:54.029 --> 00:09:55.809 with the labeled patterns it remembers from the 00:09:55.809 --> 00:09:57.789 source domain and just try to bridge the gap 00:09:57.789 --> 00:10:00.070 on its own. That's it. Now the second scenario 00:10:00.070 --> 00:10:02.269 is semi -supervised. This is where the vast majority 00:10:02.269 --> 00:10:05.120 of the target data is unlabeled. have a tiny 00:10:05.120 --> 00:10:08.320 handful of labeled examples. Like maybe the executive 00:10:08.320 --> 00:10:11.700 took five minutes to flag a few really obvious 00:10:11.700 --> 00:10:14.899 spam emails. Perfect example. Those few labels 00:10:14.899 --> 00:10:17.779 act as mathematical anchors to help align the 00:10:17.779 --> 00:10:20.399 distributions of the two domains. Got it. And 00:10:20.399 --> 00:10:22.700 the third. Finally, there's supervised domain 00:10:22.700 --> 00:10:24.720 adaptation where all the available data from 00:10:24.720 --> 00:10:27.179 the target domain is already labeled. Okay, but 00:10:27.179 --> 00:10:29.559 for you listening... If you're managing data 00:10:29.559 --> 00:10:32.620 for your own projects, unsupervised is clearly 00:10:32.620 --> 00:10:34.740 the holy grail here, right? Because it means 00:10:34.740 --> 00:10:38.299 you skip the incredibly tedious, expensive human 00:10:38.299 --> 00:10:41.139 labor of relabeling everything. Absolutely. And 00:10:41.139 --> 00:10:43.500 if you have fully supervised data, haven't you 00:10:43.500 --> 00:10:46.460 just bypassed the core problem of domain adaptation 00:10:46.460 --> 00:10:49.080 entirely? Well, you've certainly reduced the 00:10:49.080 --> 00:10:52.259 complexity. As the source notes, in a fully supervised 00:10:52.259 --> 00:10:55.259 scenario, domain adaptation essentially reduces 00:10:55.259 --> 00:10:57.899 to a refinement of the original source predictor. 00:10:58.039 --> 00:11:00.240 Just a touch up. Yeah, it's akin to taking that 00:11:00.240 --> 00:11:02.399 bittily hair color network and fine -tuning it 00:11:02.399 --> 00:11:04.840 with a fully labeled Norway data set. It's an 00:11:04.840 --> 00:11:07.240 incremental update rather than a massive leap 00:11:07.240 --> 00:11:09.860 across an unknown statistical chasm. Okay, so 00:11:09.860 --> 00:11:12.240 we understand the shifts. We know what kind of 00:11:12.240 --> 00:11:14.480 data we might be working with. Now let's get 00:11:14.480 --> 00:11:16.980 into the engine room. Let's do it. What are the 00:11:16.980 --> 00:11:19.620 actual mathematical mechanisms the algorithmic 00:11:19.620 --> 00:11:22.559 master plans use to execute this transfer of 00:11:22.559 --> 00:11:25.320 knowledge? The source material lists four main 00:11:25.320 --> 00:11:27.860 principles and they represent totally different 00:11:27.860 --> 00:11:30.500 philosophies of problem solving. Let's start 00:11:30.500 --> 00:11:33.799 with reweighting algorithms. Okay, so the objective 00:11:33.799 --> 00:11:36.299 of a reweighting algorithm is to take the labeled 00:11:36.299 --> 00:11:39.320 samples from the source domain and manually adjust 00:11:39.320 --> 00:11:42.100 their statistical significance in the loss function. 00:11:42.350 --> 00:11:44.690 So that the source distribution mathematically 00:11:44.690 --> 00:11:47.649 mirrors the target distribution. Exactly. Okay, 00:11:47.730 --> 00:11:49.470 I want to break down how that actually works 00:11:49.470 --> 00:11:52.149 in practice. Are we artificially duplicating 00:11:52.149 --> 00:11:55.070 data, or are we changing how the AI perceives 00:11:55.070 --> 00:11:57.529 the data it already has? We're changing how it 00:11:57.529 --> 00:12:00.309 perceives the data by applying a technique called 00:12:00.309 --> 00:12:02.929 important sampling. Important sampling? Right. 00:12:03.289 --> 00:12:05.590 During training, every time an AI makes a mistake, 00:12:05.730 --> 00:12:08.090 it receives a mathematical penalty, which updates 00:12:08.090 --> 00:12:11.169 its rulebook. In reweighting, we multiply the 00:12:11.169 --> 00:12:13.809 penalty for certain types of source data to force 00:12:13.809 --> 00:12:16.570 the AI to care about them more. Oh, I see. Let's 00:12:16.570 --> 00:12:18.970 say you're adapting that spam filter to the executive 00:12:18.970 --> 00:12:21.990 and the unlabeled target data suggests they receive 00:12:21.990 --> 00:12:24.509 a massive amount of invoices. Right, and the 00:12:24.509 --> 00:12:27.210 AI looks back at its source data, my inbox, and 00:12:27.210 --> 00:12:30.330 realizes I only received like one or two fake 00:12:30.330 --> 00:12:33.250 invoices a year. It originally learned to mostly 00:12:33.250 --> 00:12:35.909 ignore them because they were statistically insignificant. 00:12:36.409 --> 00:12:39.440 But under a reweighting algorithm, the engineers 00:12:39.440 --> 00:12:42.620 intervene. They tell the algorithm to heavily 00:12:42.620 --> 00:12:45.379 multiply the penalty for getting those specific 00:12:45.379 --> 00:12:48.299 invoice emails wrong during its retraining phase. 00:12:50.259 --> 00:12:53.059 By doing so, they force the old source data to 00:12:53.059 --> 00:12:55.500 simulate the shape and priorities of the new 00:12:55.500 --> 00:12:58.220 reality. It's like telling the AI, remember those 00:12:58.220 --> 00:13:00.580 rare examples you barely paid attention to back 00:13:00.580 --> 00:13:02.899 home? Pretend those are now the most important 00:13:02.899 --> 00:13:05.090 things in the world. That makes total sense. 00:13:05.350 --> 00:13:07.070 It's very effective. So the second principle 00:13:07.070 --> 00:13:10.049 is iterative algorithms. The text describes this 00:13:10.049 --> 00:13:13.389 as a form of auto labeling, right? Yes. Often 00:13:13.389 --> 00:13:15.970 referred to as pseudo labeling. You take the 00:13:15.970 --> 00:13:18.149 model you trained on the source data and you 00:13:18.149 --> 00:13:21.230 let it loose on the unlabeled target data. It 00:13:21.230 --> 00:13:23.269 makes its best predictions and automatically 00:13:23.269 --> 00:13:26.370 applies a label to the target examples it feels 00:13:26.370 --> 00:13:29.450 most confident about. It takes its own test and 00:13:29.450 --> 00:13:32.909 grades it. But wait, isn't that a massive vulnerability? 00:13:33.289 --> 00:13:35.769 How so? Well, if the model is already confused 00:13:35.769 --> 00:13:38.009 by the new domain and it auto -labels something 00:13:38.009 --> 00:13:40.529 incorrectly, it's just going to use that incorrect 00:13:40.529 --> 00:13:42.990 data to train itself further. It would just create 00:13:42.990 --> 00:13:45.049 a feedback loop of its own mistakes. That is 00:13:45.049 --> 00:13:47.470 exactly the danger. That phenomenon is called 00:13:47.470 --> 00:13:50.230 confirmation bias or error amplification, and 00:13:50.230 --> 00:13:52.490 it is the primary risk of iterative approaches. 00:13:52.730 --> 00:13:56.029 So how do they fix that? To mitigate this, engineers 00:13:56.029 --> 00:13:59.769 employ strict confidence thresholds. The AI doesn't 00:13:59.769 --> 00:14:02.809 auto -label everything. It evaluates the mathematical 00:14:02.809 --> 00:14:05.769 certainty of its predictions and might only apply 00:14:05.769 --> 00:14:09.470 pseudo -labels to the top 1 % of data it is most 00:14:09.470 --> 00:14:11.600 overwhelmingly confident about. Oh, I get it. 00:14:11.600 --> 00:14:13.919 So it only anchors itself to the things it absolutely 00:14:13.919 --> 00:14:16.299 knows for sure and then retrains a new model 00:14:16.299 --> 00:14:19.159 based on those anchors. Exactly. It repeats that 00:14:19.159 --> 00:14:22.759 process iteratively. Train a model, label the 00:14:22.759 --> 00:14:25.500 highest confidence target examples, combine those 00:14:25.500 --> 00:14:27.779 with the source data, and train a new model. 00:14:27.980 --> 00:14:30.480 It slowly bootstraps its way toward understanding 00:14:30.480 --> 00:14:32.740 the target domain, expanding its confidence with 00:14:32.740 --> 00:14:34.960 each generation. Yeah, it builds its own stepping 00:14:34.960 --> 00:14:38.269 stones, essentially. Fascinating. Now, the third 00:14:38.269 --> 00:14:40.809 principle seems to take an entirely different, 00:14:41.049 --> 00:14:44.549 almost combative approach. It's called the search 00:14:44.549 --> 00:14:47.529 of a common representation space. Yes. And this 00:14:47.529 --> 00:14:50.190 is where the engineering becomes highly sophisticated. 00:14:50.529 --> 00:14:53.230 The goal is to mathematically map the source 00:14:53.230 --> 00:14:56.370 domain and the target domain into a shared space 00:14:56.370 --> 00:14:58.870 where their distributions overlap so perfectly 00:14:58.870 --> 00:15:01.129 that they are indistinguishable from one another. 00:15:01.330 --> 00:15:02.850 Wait, you're trying to erase the differences 00:15:02.850 --> 00:15:06.200 between Italy and Norway? How do you force two 00:15:06.200 --> 00:15:08.720 distinct data sets to become indistinguishable? 00:15:08.759 --> 00:15:11.740 By utilizing adversarial machine learning. Adversarial? 00:15:11.759 --> 00:15:13.840 Yeah. You actually set up two neural networks 00:15:13.840 --> 00:15:16.460 and force them to compete in minimax games. Like 00:15:16.460 --> 00:15:18.399 they're playing against each other. Exactly. 00:15:19.179 --> 00:15:22.220 The first network is a feature extractor. Its 00:15:22.220 --> 00:15:24.940 job is to look at the data and pull out the relevant 00:15:24.940 --> 00:15:28.220 information to perform the primary task, like 00:15:28.220 --> 00:15:30.919 classifying hair color. Okay. The second network 00:15:30.919 --> 00:15:33.820 acts as a domain discriminator. Its sole purpose 00:15:33.820 --> 00:15:36.039 is to look at the features extracted by the first 00:15:36.039 --> 00:15:38.960 network and guess. Did this image come from Italy 00:15:38.960 --> 00:15:41.500 or did it come from Norway? This sounds like 00:15:41.500 --> 00:15:44.139 a police sketch artist forced to draw a suspect 00:15:44.139 --> 00:15:47.039 for a detective. That is an excellent way to 00:15:47.039 --> 00:15:48.840 conceptualize it. Like if the sketch artist, 00:15:48.860 --> 00:15:51.000 the feature extractor draws the suspect standing 00:15:51.000 --> 00:15:53.299 in front of the Coliseum or wearing a heavy Norwegian 00:15:53.299 --> 00:15:56.960 winter coat, the detect, the discriminator immediately 00:15:56.960 --> 00:15:59.740 knows where the image came from. Right. And in 00:15:59.740 --> 00:16:02.539 this adversarial setup, the sketch artist is 00:16:02.539 --> 00:16:05.519 actively mathematically penalized through something 00:16:05.519 --> 00:16:08.100 called a gradient reversal layer. If the detective 00:16:08.100 --> 00:16:10.659 successfully guesses the origin. Whoa. So the 00:16:10.659 --> 00:16:13.580 sketch artist is forced to stop drawing the backgrounds, 00:16:13.720 --> 00:16:15.940 the clothing, the lighting. everything that gives 00:16:15.940 --> 00:16:18.360 away the domain, they're forced to draw only 00:16:18.360 --> 00:16:21.500 the face. They extract only the core universal 00:16:21.500 --> 00:16:24.620 features of hair color that are completely independent 00:16:24.620 --> 00:16:28.440 of geography. Exactly. By heavily penalizing 00:16:28.440 --> 00:16:31.379 the network for retaining domain -specific information, 00:16:32.080 --> 00:16:34.100 you force it to become blind to the environment. 00:16:34.590 --> 00:16:37.090 It strips away all the local shortcuts it used 00:16:37.090 --> 00:16:39.850 to rely on, leaving only the fundamental truth 00:16:39.850 --> 00:16:42.570 of the concept. That's brilliant. By destroying 00:16:42.570 --> 00:16:45.350 its ability to differentiate the domains, you 00:16:45.350 --> 00:16:47.889 force it to be universally smart rather than 00:16:47.889 --> 00:16:51.480 locally smart. You force the AI to focus on the 00:16:51.480 --> 00:16:54.059 signal by actively punishing it for seeing the 00:16:54.059 --> 00:16:56.039 noise. I love that. That's very elegant. Which 00:16:56.039 --> 00:16:58.480 brings us to the fourth algorithmic principle, 00:16:58.700 --> 00:17:00.700 and this one relies on a completely different 00:17:00.700 --> 00:17:02.840 branch of mathematics. Right. The hierarchical 00:17:02.840 --> 00:17:05.000 Bayesian model. Right. This approach involves 00:17:05.000 --> 00:17:07.599 constructing a Bayesian factorization model for 00:17:07.599 --> 00:17:10.460 counts represented mathematically as p of n. 00:17:10.599 --> 00:17:13.140 OK, let's translate p of n and factorization 00:17:13.140 --> 00:17:15.480 models into something tangible. How does this 00:17:15.480 --> 00:17:17.839 actually solve the domain problem? Well, let's 00:17:17.839 --> 00:17:21.660 say our data involves counting the frequency 00:17:21.660 --> 00:17:25.220 of specific words in an email that's rn. The 00:17:25.220 --> 00:17:27.400 model seeks to understand the probability, the 00:17:27.400 --> 00:17:30.940 p, of those word counts. To do this across different 00:17:30.940 --> 00:17:34.079 domains, a hierarchical Bayesian model breaks 00:17:34.079 --> 00:17:37.140 the data down into latent representation. Latent 00:17:37.140 --> 00:17:39.859 representations, meaning? They're hidden underlying 00:17:39.859 --> 00:17:42.579 variables that explain the data we observe. So 00:17:42.579 --> 00:17:44.339 instead of just looking at the surface level 00:17:44.339 --> 00:17:48.279 words, the AI tries to deduce the hidden topics 00:17:48.279 --> 00:17:50.839 generating those words. Correct. But the brilliance 00:17:50.839 --> 00:17:53.220 of the hierarchical approach is how it structures 00:17:53.220 --> 00:17:56.099 those hidden variables. The model mathematically 00:17:56.099 --> 00:17:59.619 derives latent factors that are entirely specific 00:17:59.619 --> 00:18:02.680 to a single domain, but it also derives global 00:18:02.680 --> 00:18:05.430 latent. factors that are shared across all domains. 00:18:05.569 --> 00:18:07.730 Oh, I see. It creates a filing system for its 00:18:07.730 --> 00:18:09.670 own knowledge. Basically, yeah. It essentially 00:18:09.670 --> 00:18:12.289 says, I've analyzed the data, and this particular 00:18:12.289 --> 00:18:14.970 rule about language only applies to the CEO's 00:18:14.970 --> 00:18:17.089 inbox. I'll put that in the domain -specific 00:18:17.089 --> 00:18:19.690 folder, but this other rule about how phishing 00:18:19.690 --> 00:18:22.039 links are formatted applies everywhere. I'll 00:18:22.039 --> 00:18:23.839 put that in the global folder. That's a great 00:18:23.839 --> 00:18:26.640 analogy. It mathematically separates the universal 00:18:26.640 --> 00:18:29.900 truths from the local quirks. It builds a hierarchy 00:18:29.900 --> 00:18:32.460 of knowledge, allowing for a specialized adaptation 00:18:32.460 --> 00:18:35.079 while maintaining a shared foundational understanding. 00:18:35.309 --> 00:18:37.589 What's really remarkable is that this isn't just 00:18:37.589 --> 00:18:39.509 theoretical math sitting in an academic paper 00:18:39.509 --> 00:18:42.910 from 2010. This is active, deployable engineering. 00:18:43.069 --> 00:18:45.369 It really is. The source material emphasizes 00:18:45.369 --> 00:18:48.390 that if you, listening to this, are dealing with 00:18:48.390 --> 00:18:50.569 a covariate shift in your own data right now, 00:18:50.769 --> 00:18:53.369 you don't have to invent a gradient reversal 00:18:53.369 --> 00:18:55.750 layer from scratch. Definitely not. The theory 00:18:55.750 --> 00:18:58.390 has been beautifully translated into accessible 00:18:58.390 --> 00:19:01.410 software packages over the past decades. If you 00:19:01.410 --> 00:19:04.230 work in the Python ecosystem, which dominates 00:19:04.230 --> 00:19:07.190 this space, there are comprehensive tools available. 00:19:07.230 --> 00:19:09.950 Like what? Well, you have SCADA, which stands 00:19:09.950 --> 00:19:12.029 for site kit adaptation, and that's designed 00:19:12.029 --> 00:19:14.190 to integrate smoothly with standard machine learning 00:19:14.190 --> 00:19:17.650 workflows. There's also ADPT, the awesome domain 00:19:17.650 --> 00:19:20.829 adaptation Python toolbox, and TLib, the transfer 00:19:20.829 --> 00:19:22.910 learning library. Nice. And if you're working 00:19:22.910 --> 00:19:25.470 outside of Python? For engineers operating within 00:19:25.470 --> 00:19:28.009 MATLAB, there is the appropriately named Domain 00:19:28.009 --> 00:19:31.500 Adaptation Toolbox. These packages compile the 00:19:31.500 --> 00:19:33.859 reweighting algorithms, the adversarial networks, 00:19:34.420 --> 00:19:36.680 and the iterative autolabelers into ready -to 00:19:36.680 --> 00:19:39.240 -use functions. So the theoretical math has really 00:19:39.240 --> 00:19:41.380 just been commoditized into practical tools. 00:19:41.880 --> 00:19:44.019 Exactly. Well, let's synthesize the journey we've 00:19:44.019 --> 00:19:46.490 taken today. We started with the realization 00:19:46.490 --> 00:19:49.670 that an AI is fundamentally captive to the specific 00:19:49.670 --> 00:19:51.769 environment it was trained in. Right, it's literal. 00:19:51.990 --> 00:19:54.650 And we explored the engineering required to help 00:19:54.650 --> 00:19:57.710 a model step from a safe, familiar source domain 00:19:57.710 --> 00:20:01.829 D sub S into a wild, unpredictable target domain 00:20:01.829 --> 00:20:05.009 D sub T. We navigated the distinct types of statistical 00:20:05.009 --> 00:20:07.670 failures. covariate shifts where the input patterns 00:20:07.670 --> 00:20:10.390 change, prior shifts where the demographic proportions 00:20:10.390 --> 00:20:12.910 flip, and concept shifts where the foundational 00:20:12.910 --> 00:20:15.359 rules of reality are completely rewritten. And 00:20:15.359 --> 00:20:18.279 we broke down how engineers execute the fits 00:20:18.279 --> 00:20:21.700 by maliciously reweighting the AI's memories, 00:20:22.200 --> 00:20:24.460 by carefully allowing it to auto -label its future 00:20:24.460 --> 00:20:27.140 through pseudo -labels, by using adversarial 00:20:27.140 --> 00:20:30.059 networks to strip away environmental noise, and 00:20:30.059 --> 00:20:32.380 by building Bayesian hierarchies to separate 00:20:32.380 --> 00:20:35.900 local quirks from universal truths. It is a remarkable 00:20:35.900 --> 00:20:38.680 mathematical process of unlearning and relearning, 00:20:39.160 --> 00:20:42.039 which Honestly, if we revisit our earlier discussion, 00:20:42.259 --> 00:20:44.680 brings up a rather profound implication regarding 00:20:44.680 --> 00:20:46.940 concept shifts. Oh, right. The scenario where 00:20:46.940 --> 00:20:49.099 the exact same symptom points to an entirely 00:20:49.099 --> 00:20:51.059 different disease depending on the hospital the 00:20:51.059 --> 00:20:54.160 AI is deployed in. Yeah. If an AI model is constantly 00:20:54.160 --> 00:20:57.160 adapting to concept shifts, where the very relationship 00:20:57.160 --> 00:20:59.759 between inputs and outputs fundamentally changes 00:20:59.759 --> 00:21:02.619 to fit a new reality, at what point does the 00:21:02.619 --> 00:21:04.519 adapted model become an entirely new entity? 00:21:04.680 --> 00:21:08.039 Wow. It's the AI equivalent of the ship of Theseus. 00:21:08.400 --> 00:21:10.460 If you replace every plank of wood on a ship 00:21:10.460 --> 00:21:13.200 over time, is it still the same ship? Exactly. 00:21:13.359 --> 00:21:15.859 If the model has to continuously overwrite its 00:21:15.859 --> 00:21:18.519 foundational mapping of X to Y just to survive 00:21:18.519 --> 00:21:21.839 a new domain, are we actually transferring knowledge? 00:21:22.180 --> 00:21:25.259 Or are we slowly overwriting the original AI's 00:21:25.259 --> 00:21:27.579 reality until absolutely nothing of the source 00:21:27.579 --> 00:21:30.180 domain remains? If the same symptom now means 00:21:30.180 --> 00:21:32.539 a different disease, the model's original truth 00:21:32.539 --> 00:21:34.819 is just gone. Something for you to chew on the 00:21:34.819 --> 00:21:36.900 next time you hear about a quote unquote highly 00:21:36.900 --> 00:21:40.019 trained AI being deployed into a brand new environment. 00:21:40.480 --> 00:21:42.319 Whose reality is it actually experiencing?