1 00:00:00,000 --> 00:00:03,040 All right, so we're diving into data de-biasing 2 00:00:03,040 --> 00:00:07,080 with data models, D3M, improving subgroup robustness 3 00:00:07,080 --> 00:00:08,400 via data selection. 4 00:00:08,400 --> 00:00:09,240 Okay. 5 00:00:09,240 --> 00:00:14,080 It's about making AI models fairer, more accurate, 6 00:00:14,080 --> 00:00:15,720 especially when you're dealing with different groups 7 00:00:15,720 --> 00:00:16,760 within the data. 8 00:00:16,760 --> 00:00:17,600 Yeah. 9 00:00:17,600 --> 00:00:20,200 You know, like how an AI model might be amazing 10 00:00:20,200 --> 00:00:22,880 at recognizing faces of one age group 11 00:00:22,880 --> 00:00:24,160 but not so good with another. 12 00:00:24,160 --> 00:00:25,120 Yeah. 13 00:00:25,120 --> 00:00:27,000 This paper tackles that kind of problem. 14 00:00:27,000 --> 00:00:28,920 What's really fascinating here is this idea 15 00:00:28,920 --> 00:00:30,720 of worse group error. 16 00:00:30,720 --> 00:00:34,400 It's like a way to measure how much an AI model's performance 17 00:00:34,400 --> 00:00:37,960 just drops when it encounters a specific subgroup of data 18 00:00:37,960 --> 00:00:40,360 it might not have seen a lot of during its training. 19 00:00:40,360 --> 00:00:42,240 So it's like the model's Achilles heel, 20 00:00:42,240 --> 00:00:44,120 those groups where it just doesn't do as well. 21 00:00:44,120 --> 00:00:44,960 Exactly. 22 00:00:44,960 --> 00:00:47,200 For instance, imagine you are training an AI 23 00:00:47,200 --> 00:00:50,960 to classify images of faces as like old or young. 24 00:00:50,960 --> 00:00:51,800 Okay. 25 00:00:51,800 --> 00:00:54,840 But your training data has way more old men and young women. 26 00:00:54,840 --> 00:00:56,880 The model might do amazingly on those 27 00:00:56,880 --> 00:00:59,120 but then struggle with old women or young men. 28 00:00:59,120 --> 00:00:59,960 Got it. 29 00:00:59,960 --> 00:01:01,680 That's worst group error in action. 30 00:01:01,680 --> 00:01:02,600 Okay, that makes sense. 31 00:01:02,600 --> 00:01:05,400 So then how does this D3M approach come in? 32 00:01:05,400 --> 00:01:06,760 How does that tackle that? 33 00:01:06,760 --> 00:01:10,080 So D3M actually identifies and removes 34 00:01:10,080 --> 00:01:13,560 those specific training examples that are causing the model 35 00:01:13,560 --> 00:01:15,320 to make the bias predictions. 36 00:01:15,320 --> 00:01:17,200 So we're talking about actually deleting data. 37 00:01:17,200 --> 00:01:18,040 Yeah. 38 00:01:18,040 --> 00:01:20,120 Think of it like a detective sniffing out the bad apples 39 00:01:20,120 --> 00:01:21,280 in the data barrel. 40 00:01:21,280 --> 00:01:22,120 Okay. 41 00:01:22,120 --> 00:01:23,960 But doesn't that seem a bit risky? 42 00:01:23,960 --> 00:01:25,880 Like we could lose some valuable information. 43 00:01:25,880 --> 00:01:27,480 Yeah, it's a valid concern. 44 00:01:27,480 --> 00:01:28,320 Yeah. 45 00:01:28,320 --> 00:01:30,800 But D3M is more strategic than just randomly 46 00:01:30,800 --> 00:01:31,960 chucking out data. 47 00:01:31,960 --> 00:01:32,800 Okay. 48 00:01:32,800 --> 00:01:34,560 It uses something called data models. 49 00:01:34,560 --> 00:01:35,400 Data models. 50 00:01:35,400 --> 00:01:37,680 Think of them like AI fortune tellers. 51 00:01:37,680 --> 00:01:38,520 Okay. 52 00:01:38,520 --> 00:01:40,680 They predict how the model will behave based 53 00:01:40,680 --> 00:01:42,240 on the training data that you give it. 54 00:01:42,240 --> 00:01:44,360 Okay, so we've got these AI fortune tellers, 55 00:01:44,360 --> 00:01:46,080 these data models guiding us. 56 00:01:46,080 --> 00:01:46,920 Yeah. 57 00:01:46,920 --> 00:01:48,600 What are they specifically looking for in the data? 58 00:01:48,600 --> 00:01:52,200 They use this thing called a group alignment score. 59 00:01:52,200 --> 00:01:53,040 Okay. 60 00:01:53,040 --> 00:01:56,680 AI basically quantifies how much each training example 61 00:01:56,680 --> 00:01:58,840 is contributing to the model's bias. 62 00:01:58,840 --> 00:02:01,640 So a high AI, that data point is probably causing 63 00:02:01,640 --> 00:02:03,480 some trouble skewing those results. 64 00:02:03,480 --> 00:02:05,520 So if we remove those high AI, 65 00:02:05,520 --> 00:02:07,280 we can create a more balanced data set. 66 00:02:07,280 --> 00:02:08,120 Exactly. 67 00:02:08,120 --> 00:02:09,760 And therefore a more fair model. 68 00:02:09,760 --> 00:02:10,600 Exactly. 69 00:02:10,600 --> 00:02:13,400 And the papers show some very impressive results. 70 00:02:13,400 --> 00:02:14,240 Yeah. 71 00:02:14,240 --> 00:02:17,800 They tested D3M on four data sets. 72 00:02:17,800 --> 00:02:18,640 Okay. 73 00:02:18,640 --> 00:02:20,720 Celeb age, those faces that we talked about. 74 00:02:20,720 --> 00:02:21,560 Right. 75 00:02:21,560 --> 00:02:22,400 Celeb a blonde. 76 00:02:22,400 --> 00:02:23,240 Okay. 77 00:02:23,240 --> 00:02:25,120 Water birds, which is, you know, all about birds. 78 00:02:25,120 --> 00:02:25,960 All right. 79 00:02:25,960 --> 00:02:29,280 And even multi-NLI, which is a data set with text. 80 00:02:29,280 --> 00:02:30,120 Okay. 81 00:02:30,120 --> 00:02:31,160 And how did it perform? 82 00:02:31,160 --> 00:02:32,800 It performed really well. 83 00:02:32,800 --> 00:02:33,640 Yeah. 84 00:02:33,640 --> 00:02:37,760 D3M consistently outperformed other debicing methods, 85 00:02:37,760 --> 00:02:39,320 sometimes even beating approaches 86 00:02:39,320 --> 00:02:41,080 that had more information to work with. 87 00:02:41,080 --> 00:02:42,080 Hmm, that's interesting. 88 00:02:42,080 --> 00:02:43,040 It's like winning a race 89 00:02:43,040 --> 00:02:45,240 with one hand tied behind your back. 90 00:02:45,240 --> 00:02:48,240 So it's really exceeding expectations. 91 00:02:48,240 --> 00:02:49,080 Yeah. 92 00:02:49,080 --> 00:02:49,960 Okay, but let's be realistic. 93 00:02:49,960 --> 00:02:50,800 Yeah. 94 00:02:50,800 --> 00:02:53,120 So, this data comes perfectly labeled with groups 95 00:02:53,120 --> 00:02:55,000 like old man or young woman. 96 00:02:55,000 --> 00:02:55,840 Right. 97 00:02:55,840 --> 00:02:57,800 What happens when we don't have those clear-cut labels? 98 00:02:57,800 --> 00:02:59,760 That's where IUTO D3M comes in. 99 00:02:59,760 --> 00:03:00,600 Okay. 100 00:03:00,600 --> 00:03:03,360 It's like D3M's even more resourceful cousin, 101 00:03:03,360 --> 00:03:06,320 designed for when you don't have those handy group labels. 102 00:03:06,320 --> 00:03:07,960 It uses this technique called 103 00:03:07,960 --> 00:03:11,200 principal component analysis or PCA for short, 104 00:03:11,200 --> 00:03:13,440 to find those hidden patterns in the data 105 00:03:13,440 --> 00:03:15,080 that might be causing the bias. 106 00:03:15,080 --> 00:03:18,000 PCA, vaguely remember that from stats class. 107 00:03:18,000 --> 00:03:18,840 Yeah. 108 00:03:18,840 --> 00:03:19,880 It's like something of finding patterns. 109 00:03:19,880 --> 00:03:20,720 Exactly. 110 00:03:20,720 --> 00:03:24,200 Think of PCA as a super-powered pattern recognition machine. 111 00:03:24,200 --> 00:03:25,040 Okay. 112 00:03:25,040 --> 00:03:27,840 It can uncover these subtle clusters of data points 113 00:03:27,840 --> 00:03:30,560 that might be behaving differently from the rest. 114 00:03:30,560 --> 00:03:31,400 Right. 115 00:03:31,400 --> 00:03:33,720 So it's hinting at those hidden biases. 116 00:03:33,720 --> 00:03:37,840 So, AUTO D3M uses PCA to create its own 117 00:03:37,840 --> 00:03:41,000 like temporary labels based on these hidden patterns. 118 00:03:41,000 --> 00:03:41,840 It does, yeah. 119 00:03:41,840 --> 00:03:43,440 And once it has those pseudo labels, 120 00:03:43,440 --> 00:03:45,800 it can then apply the same D3M magic. 121 00:03:45,800 --> 00:03:46,640 Okay. 122 00:03:46,640 --> 00:03:47,480 Removing those troublemakers 123 00:03:47,480 --> 00:03:49,720 to create a more fair data set. 124 00:03:49,720 --> 00:03:51,680 That sounds almost too good to be true. 125 00:03:51,680 --> 00:03:52,520 It's a... 126 00:03:52,520 --> 00:03:54,760 Even when we're dealing with messy, unlabeled data, 127 00:03:54,760 --> 00:03:56,480 we can still make progress towards 128 00:03:56,480 --> 00:03:57,640 debiasing these models. 129 00:03:57,640 --> 00:03:58,480 That's right. 130 00:03:58,480 --> 00:04:00,520 And the researchers found that AUTO D3M 131 00:04:00,520 --> 00:04:02,160 performed surprisingly well, 132 00:04:02,160 --> 00:04:05,520 often achieving results that are very comparable to D3M, 133 00:04:05,520 --> 00:04:07,800 even without those predefined group labels. 134 00:04:07,800 --> 00:04:09,320 That's really fascinating stuff, 135 00:04:09,320 --> 00:04:12,400 but I have to admit, it's all getting a bit abstract for me. 136 00:04:12,400 --> 00:04:13,400 Okay. 137 00:04:13,400 --> 00:04:15,080 Do you have any concrete examples 138 00:04:15,080 --> 00:04:16,720 of what this looks like in practice? 139 00:04:16,720 --> 00:04:18,000 Absolutely. 140 00:04:18,000 --> 00:04:20,840 Let's look a bit closer at the qualitative findings 141 00:04:20,840 --> 00:04:23,200 from the CelebAge data set. 142 00:04:23,200 --> 00:04:26,240 They found that D3M often flagged images 143 00:04:26,240 --> 00:04:29,160 of young people with gray hair 144 00:04:29,160 --> 00:04:31,240 as potential troublemakers. 145 00:04:31,240 --> 00:04:32,080 Okay, that makes sense. 146 00:04:32,080 --> 00:04:32,920 Yeah. 147 00:04:32,920 --> 00:04:34,640 I could see how that would throw off an AI 148 00:04:34,640 --> 00:04:37,480 that's been trained to associate young 149 00:04:37,480 --> 00:04:38,640 with a certain hair color. 150 00:04:38,640 --> 00:04:39,480 Right. 151 00:04:39,480 --> 00:04:41,040 But did removing those images 152 00:04:41,040 --> 00:04:42,440 actually make a difference? 153 00:04:42,440 --> 00:04:43,560 It did. 154 00:04:43,560 --> 00:04:46,200 Removing those potentially mislabeled examples 155 00:04:46,200 --> 00:04:49,440 led to significant improvements in the model's accuracy, 156 00:04:49,440 --> 00:04:50,920 particularly for those groups 157 00:04:50,920 --> 00:04:53,000 that were initially performing poorly. 158 00:04:53,000 --> 00:04:55,920 So sometimes the bias isn't in the data itself, 159 00:04:55,920 --> 00:04:57,280 but how we've labeled it. 160 00:04:57,280 --> 00:04:58,160 You got it. 161 00:04:58,160 --> 00:04:59,520 To further test their methods, 162 00:04:59,520 --> 00:05:02,000 they decided to tackle a real behemoth. 163 00:05:02,000 --> 00:05:02,840 ImageNet. 164 00:05:02,840 --> 00:05:03,680 ImageNet. 165 00:05:03,680 --> 00:05:04,520 That's the one that everybody uses 166 00:05:04,520 --> 00:05:06,440 to train their image recognition models. 167 00:05:06,440 --> 00:05:07,280 Exactly. 168 00:05:07,280 --> 00:05:10,000 Now, ImageNet doesn't come with any predefined bias labels. 169 00:05:10,000 --> 00:05:10,840 Okay. 170 00:05:10,840 --> 00:05:13,640 So it was a perfect testing ground for AUTO D3M. 171 00:05:13,640 --> 00:05:15,080 They let it loose on ImageNet 172 00:05:15,080 --> 00:05:17,560 to see what hidden biases it could uncover. 173 00:05:17,560 --> 00:05:19,400 And did it find anything interesting? 174 00:05:19,400 --> 00:05:20,320 It did. 175 00:05:20,320 --> 00:05:23,720 It turns out that the model might be a bit too reliant 176 00:05:23,720 --> 00:05:27,880 on the color red when it classifies images of red wolves. 177 00:05:27,880 --> 00:05:30,400 And it seems to associate TengeFish 178 00:05:30,400 --> 00:05:33,120 with images that also have humans in them. 179 00:05:33,120 --> 00:05:33,960 Oh, interesting. 180 00:05:33,960 --> 00:05:36,560 Even though those two things aren't really related. 181 00:05:36,560 --> 00:05:39,440 So even with a massive data set like ImageNet, 182 00:05:39,440 --> 00:05:41,320 there could be these hidden biases. 183 00:05:41,320 --> 00:05:42,160 Exactly. 184 00:05:42,160 --> 00:05:43,120 These quirks in the data 185 00:05:43,120 --> 00:05:44,880 that can lead to skewed results. 186 00:05:44,880 --> 00:05:45,720 Yeah. 187 00:05:45,720 --> 00:05:49,360 And the exciting part is that both D3M and AUTO D3M 188 00:05:49,360 --> 00:05:51,600 were able to improve the model's performance 189 00:05:51,600 --> 00:05:53,360 on those specific categories 190 00:05:53,360 --> 00:05:56,160 without negatively impacting the overall accuracy. 191 00:05:56,160 --> 00:05:57,920 Okay. This is all really cool. 192 00:05:57,920 --> 00:05:59,320 But before we get too carried away, 193 00:05:59,320 --> 00:06:00,640 I would love to understand a little bit more 194 00:06:00,640 --> 00:06:01,800 about those data models. 195 00:06:01,800 --> 00:06:02,640 Sure. 196 00:06:02,640 --> 00:06:04,360 They seem to be doing a lot of the heavy lifting here. 197 00:06:04,360 --> 00:06:05,200 Yeah. 198 00:06:05,200 --> 00:06:06,400 How exactly do they work? 199 00:06:06,400 --> 00:06:07,240 Oh, yeah. 200 00:06:07,240 --> 00:06:08,720 Especially that trach method you mentioned earlier. 201 00:06:08,720 --> 00:06:09,560 Yeah. 202 00:06:09,560 --> 00:06:11,040 Is it as complicated as it sounds? 203 00:06:11,040 --> 00:06:12,000 Not at all. 204 00:06:12,000 --> 00:06:15,760 It's all about understanding how changes in the training data 205 00:06:15,760 --> 00:06:17,720 affect the model's predictions. 206 00:06:17,720 --> 00:06:18,560 Okay. 207 00:06:18,560 --> 00:06:19,680 Imagine you have a recipe 208 00:06:19,680 --> 00:06:22,800 and you want to know which ingredients are essential 209 00:06:22,800 --> 00:06:24,400 for that final taste. 210 00:06:24,400 --> 00:06:26,920 Okay. So the recipe is our AI model. 211 00:06:26,920 --> 00:06:27,760 Yeah. 212 00:06:27,760 --> 00:06:30,440 And the ingredients are the data points. 213 00:06:30,440 --> 00:06:31,280 Precisely. 214 00:06:31,280 --> 00:06:32,120 Okay. 215 00:06:32,120 --> 00:06:34,440 TIRAC is like experimenting with the recipe, 216 00:06:34,440 --> 00:06:36,600 tweaking the ingredients one by one 217 00:06:36,600 --> 00:06:38,280 and seeing how the flavor changes. 218 00:06:38,280 --> 00:06:39,120 Right. 219 00:06:39,120 --> 00:06:41,760 But in this case, the flavor is the model's predictions. 220 00:06:41,760 --> 00:06:44,080 So by carefully tweaking the training data 221 00:06:44,080 --> 00:06:46,280 and seeing how the model reacts, 222 00:06:46,280 --> 00:06:48,960 TIRAC can pinpoint those crucial data points 223 00:06:48,960 --> 00:06:49,800 that have big influence. 224 00:06:49,800 --> 00:06:50,640 Exactly. 225 00:06:50,640 --> 00:06:53,600 And that's how it creates those coefficient vectors, 226 00:06:53,600 --> 00:06:54,960 those that you see in the paper. 227 00:06:54,960 --> 00:06:55,800 Right. 228 00:06:55,800 --> 00:06:58,600 They basically tell you how much each data point 229 00:06:58,600 --> 00:07:00,680 contributes to the model's prediction 230 00:07:00,680 --> 00:07:02,120 for a specific example. 231 00:07:02,120 --> 00:07:05,400 So it's like giving each data point a power level. 232 00:07:05,400 --> 00:07:06,240 Yeah. 233 00:07:06,240 --> 00:07:08,640 Based on how much it influences the model's decisions. 234 00:07:08,640 --> 00:07:10,160 That's a great way to think about it. 235 00:07:10,160 --> 00:07:12,360 And once you have those power levels, 236 00:07:12,360 --> 00:07:14,600 you can start to see which data points 237 00:07:14,600 --> 00:07:18,080 are driving the bias and strategically remove them. 238 00:07:18,080 --> 00:07:20,040 This is all starting to make a lot more sense. 239 00:07:20,040 --> 00:07:20,880 Good. 240 00:07:20,880 --> 00:07:23,120 It's about understanding that subtle dance 241 00:07:23,120 --> 00:07:25,920 between the data and the model's predictions. 242 00:07:25,920 --> 00:07:26,760 Yeah. 243 00:07:26,760 --> 00:07:29,080 But like any dance, I imagine there can be 244 00:07:29,080 --> 00:07:30,640 some missteps along the way. 245 00:07:30,640 --> 00:07:31,480 Of course. 246 00:07:31,480 --> 00:07:34,240 What are some of the limitations of this approach? 247 00:07:34,240 --> 00:07:36,320 One limitation is the reliance 248 00:07:36,320 --> 00:07:38,280 on those data models themselves. 249 00:07:38,280 --> 00:07:39,120 Okay. 250 00:07:39,120 --> 00:07:40,440 There are powerful tools. 251 00:07:40,440 --> 00:07:41,280 Right. 252 00:07:41,280 --> 00:07:42,440 But there's still approximations 253 00:07:42,440 --> 00:07:44,600 of how the AI actually works. 254 00:07:44,600 --> 00:07:48,280 So there's always a chance that they might miss something 255 00:07:48,280 --> 00:07:51,280 or misinterpret the model's behavior. 256 00:07:51,280 --> 00:07:54,760 So it's like relying on a map that's slightly outdated. 257 00:07:54,760 --> 00:07:55,720 That's a great analogy. 258 00:07:55,720 --> 00:07:57,200 It might get you most of the way there. 259 00:07:57,200 --> 00:07:58,040 Yeah. 260 00:07:58,040 --> 00:07:59,960 But you might encounter a few unexpected detours 261 00:07:59,960 --> 00:08:00,800 along the way. 262 00:08:00,800 --> 00:08:01,640 Yeah. 263 00:08:01,640 --> 00:08:02,480 Yeah. 264 00:08:02,480 --> 00:08:03,520 Another thing to consider is that even the most 265 00:08:03,520 --> 00:08:06,920 sophisticated de-biasing method will struggle 266 00:08:06,920 --> 00:08:09,320 if the data itself is fundamentally flawed 267 00:08:09,320 --> 00:08:11,240 or riddled with inaccuracies. 268 00:08:11,240 --> 00:08:13,360 It's like trying to bake a cake with rotten eggs. 269 00:08:13,360 --> 00:08:14,200 Exactly. 270 00:08:14,200 --> 00:08:15,840 You can try all the fancy techniques you want, 271 00:08:15,840 --> 00:08:18,480 but the final result is still gonna be off. 272 00:08:18,480 --> 00:08:19,320 Yeah. 273 00:08:19,320 --> 00:08:21,240 Garbage in, garbage out, as they say. 274 00:08:21,240 --> 00:08:22,400 Exactly. 275 00:08:22,400 --> 00:08:24,640 And while this research focused on image 276 00:08:24,640 --> 00:08:27,880 and text classification, it's not yet clear 277 00:08:27,880 --> 00:08:30,760 how well this approach would work in other areas of AI, 278 00:08:30,760 --> 00:08:33,640 like natural language processing or robotics. 279 00:08:33,640 --> 00:08:34,480 Okay. 280 00:08:34,480 --> 00:08:35,800 So there's definitely more research needed 281 00:08:35,800 --> 00:08:37,520 to explore those possibilities. 282 00:08:37,520 --> 00:08:38,920 So this is a great step forward. 283 00:08:38,920 --> 00:08:39,760 Yeah. 284 00:08:39,760 --> 00:08:42,600 But it's definitely not the final word on AI de-biasing. 285 00:08:42,600 --> 00:08:43,440 Right. 286 00:08:43,440 --> 00:08:47,280 It's a powerful tool in our ever-evolving AI toolkit. 287 00:08:47,280 --> 00:08:49,120 But we need to keep pushing the boundaries 288 00:08:49,120 --> 00:08:51,320 and exploring those new approaches. 289 00:08:51,320 --> 00:08:53,280 Well, this has certainly been a deep dive. 290 00:08:53,280 --> 00:08:55,120 But before we get too carried away 291 00:08:55,120 --> 00:08:57,320 with the possibilities and the limitations, 292 00:08:57,320 --> 00:08:59,280 maybe we should pause here and give our listeners 293 00:08:59,280 --> 00:09:01,160 a chance to digest all this information. 294 00:09:01,160 --> 00:09:02,000 Okay. 295 00:09:02,000 --> 00:09:03,840 We'll be back soon to wrap up our discussion 296 00:09:03,840 --> 00:09:05,280 of this fascinating research. 297 00:09:05,280 --> 00:09:06,480 Sounds good. 298 00:09:06,480 --> 00:09:08,000 Welcome back to our deep dive 299 00:09:08,000 --> 00:09:10,560 into the world of AI de-biasing. 300 00:09:10,560 --> 00:09:13,120 You know, I'm still thinking about those AI fortune tellers, 301 00:09:13,120 --> 00:09:14,280 those data models. 302 00:09:14,280 --> 00:09:17,600 It's just amazing how they can help us spot those biases 303 00:09:17,600 --> 00:09:18,680 hiding in the data. 304 00:09:18,680 --> 00:09:19,520 Right. 305 00:09:19,520 --> 00:09:21,680 But you know, I'm always looking for that, so what factor? 306 00:09:21,680 --> 00:09:22,520 Okay. 307 00:09:22,520 --> 00:09:24,160 Why should our listeners care about all of this? 308 00:09:24,160 --> 00:09:27,440 How does this actually impact the real world? 309 00:09:27,440 --> 00:09:29,080 That's a very crucial question. 310 00:09:29,080 --> 00:09:29,920 Yeah. 311 00:09:29,920 --> 00:09:33,400 This research has the potential to make AI systems 312 00:09:33,400 --> 00:09:37,480 fairer and more trustworthy across the board. 313 00:09:37,480 --> 00:09:38,320 Okay. 314 00:09:38,320 --> 00:09:41,120 So imagine like AI powered hiring tools 315 00:09:41,120 --> 00:09:44,680 that don't discriminate based on gender or race 316 00:09:44,680 --> 00:09:48,080 or medical diagnosis systems that work well for all patients 317 00:09:48,080 --> 00:09:49,480 regardless of their background. 318 00:09:49,480 --> 00:09:51,440 So it's about building AI that works for everyone. 319 00:09:51,440 --> 00:09:52,360 Exactly. 320 00:09:52,360 --> 00:09:55,600 And this research shows us that focusing on the data itself 321 00:09:55,600 --> 00:09:57,280 is key to achieving that. 322 00:09:57,280 --> 00:09:58,320 It's like building a house, 323 00:09:58,320 --> 00:10:00,360 you need that solid foundation they write. 324 00:10:00,360 --> 00:10:02,080 If the data is biased, 325 00:10:02,080 --> 00:10:05,520 the AI that's built on top of it will inherit those flaws. 326 00:10:05,520 --> 00:10:06,360 Makes sense. 327 00:10:06,360 --> 00:10:08,520 So this isn't just some like abstract problem 328 00:10:08,520 --> 00:10:10,440 for researchers to puzzle over. 329 00:10:10,440 --> 00:10:13,560 It's about building AI that's ethical and responsible 330 00:10:13,560 --> 00:10:14,400 in the real world. 331 00:10:14,400 --> 00:10:15,320 Yeah, absolutely. 332 00:10:15,320 --> 00:10:17,200 But let's get back to the paper for a second. 333 00:10:17,200 --> 00:10:21,160 They talk about this group alignment score, AI. 334 00:10:21,160 --> 00:10:22,760 Can you break that down a little bit more? 335 00:10:22,760 --> 00:10:23,720 Absolutely. 336 00:10:23,720 --> 00:10:27,360 Remember, we're trying to pinpoint those training examples 337 00:10:27,360 --> 00:10:30,640 that are causing the model to make the biased predictions. 338 00:10:30,640 --> 00:10:33,840 The group alignment score is a way to actually measure 339 00:10:33,840 --> 00:10:36,120 how much each training example 340 00:10:36,120 --> 00:10:38,280 is contributing to that problem. 341 00:10:38,280 --> 00:10:39,720 So it's like a bias meter? 342 00:10:39,720 --> 00:10:41,280 That's a great way to put it. 343 00:10:41,280 --> 00:10:43,440 They use this clever formula 344 00:10:43,440 --> 00:10:45,720 that considers the model's performance 345 00:10:45,720 --> 00:10:47,560 on different subgroups. 346 00:10:47,560 --> 00:10:51,440 So a high AI means that data point 347 00:10:51,440 --> 00:10:54,240 is likely making that bias worse 348 00:10:54,240 --> 00:10:57,800 while a low AI suggests it's not a big contributor. 349 00:10:57,800 --> 00:10:59,280 So then we just get rid of the data points 350 00:10:59,280 --> 00:11:00,360 with the highest score. 351 00:11:00,360 --> 00:11:02,320 It's not quite as simple as that. 352 00:11:02,320 --> 00:11:04,120 There's always those trade-offs to consider 353 00:11:04,120 --> 00:11:06,040 when you're removing data, right? 354 00:11:06,040 --> 00:11:09,560 You have to be careful not to throw out valuable information 355 00:11:09,560 --> 00:11:11,120 along with the bias. 356 00:11:11,120 --> 00:11:12,640 That's a really good point. 357 00:11:12,640 --> 00:11:14,800 It's a delicate balance. 358 00:11:14,800 --> 00:11:18,600 But if we do it right, we can create a more balanced data set. 359 00:11:18,600 --> 00:11:21,320 It's exactly without sacrificing too much information. 360 00:11:21,320 --> 00:11:22,040 Exactly. 361 00:11:22,040 --> 00:11:25,320 And the paper shows that this approach can actually 362 00:11:25,320 --> 00:11:28,600 significantly improve the model's accuracy 363 00:11:28,600 --> 00:11:31,040 on those worse performing subgroups 364 00:11:31,040 --> 00:11:33,120 without hurting its overall performance. 365 00:11:33,120 --> 00:11:35,640 So it's a win-win, a fairer model 366 00:11:35,640 --> 00:11:38,120 without compromising accuracy. 367 00:11:38,120 --> 00:11:39,920 But how do you actually calculate 368 00:11:39,920 --> 00:11:42,440 this group alignment score? 369 00:11:42,440 --> 00:11:44,360 Seems like a lot of complex math. 370 00:11:44,360 --> 00:11:45,720 There is some math involved, 371 00:11:45,720 --> 00:11:48,640 but the basic idea is pretty straightforward. 372 00:11:48,640 --> 00:11:51,680 They look at how the model's predictions change 373 00:11:51,680 --> 00:11:55,320 when you remove a specific training example. 374 00:11:55,320 --> 00:11:58,080 If removing an example makes the model more accurate 375 00:11:58,080 --> 00:12:00,360 for those worse performing subgroups, 376 00:12:00,360 --> 00:12:03,480 then that example gets a high AI score. 377 00:12:03,480 --> 00:12:05,760 So it's like a what-if scenario. 378 00:12:05,760 --> 00:12:06,680 What if we remove this? 379 00:12:06,680 --> 00:12:08,440 Does it make it better or worse? 380 00:12:08,440 --> 00:12:09,280 Exactly. 381 00:12:09,280 --> 00:12:11,160 And if removing an example doesn't really change 382 00:12:11,160 --> 00:12:14,120 the model's predictions, it gets a low AI score, 383 00:12:14,120 --> 00:12:15,680 suggesting it's not a big problem. 384 00:12:15,680 --> 00:12:16,520 I see. 385 00:12:16,520 --> 00:12:19,440 So it's about measuring the influence of each data point 386 00:12:19,440 --> 00:12:22,000 on the model's overall behavior. 387 00:12:22,000 --> 00:12:24,040 But remember how we talked about those situations 388 00:12:24,040 --> 00:12:26,440 where you don't have clear group labels? 389 00:12:26,440 --> 00:12:28,640 That's where AUTO D3M comes in. 390 00:12:28,640 --> 00:12:29,640 Right. 391 00:12:29,640 --> 00:12:31,680 AUTO D3M is actually designed 392 00:12:31,680 --> 00:12:33,400 for those real world situations 393 00:12:33,400 --> 00:12:35,560 where you might not know the specific subgroups 394 00:12:35,560 --> 00:12:37,280 that are being affected by bias. 395 00:12:37,280 --> 00:12:38,280 So how does it work? 396 00:12:38,280 --> 00:12:42,040 Does it just like randomly pick data points to remove? 397 00:12:42,040 --> 00:12:42,880 No, not at all. 398 00:12:42,880 --> 00:12:43,720 Okay. 399 00:12:43,720 --> 00:12:46,400 AUTO D3M uses that technique we discussed earlier, 400 00:12:46,400 --> 00:12:48,840 principle component analysis or PCA. 401 00:12:48,840 --> 00:12:52,120 It uses PCA to actually find those hidden patterns 402 00:12:52,120 --> 00:12:54,240 in the data that might end up fate bias. 403 00:12:54,240 --> 00:12:56,400 Okay, so PCA is like that friend 404 00:12:56,400 --> 00:12:59,000 who always spots the hidden patterns in everything. 405 00:12:59,000 --> 00:13:00,200 That's a great analogy. 406 00:13:00,200 --> 00:13:02,160 Like the one who can solve those connected dots 407 00:13:02,160 --> 00:13:03,360 puzzles in a flash. 408 00:13:03,360 --> 00:13:04,200 Right. 409 00:13:04,200 --> 00:13:06,880 PCA helps us see that bigger picture in the data. 410 00:13:06,880 --> 00:13:07,720 Okay. 411 00:13:07,720 --> 00:13:10,720 Revealing those subtle clusters of data points 412 00:13:10,720 --> 00:13:12,880 that might be behaving differently from the rest. 413 00:13:12,880 --> 00:13:16,080 And those clusters often correspond to hidden subgroups. 414 00:13:16,080 --> 00:13:16,920 Exactly. 415 00:13:16,920 --> 00:13:18,240 That might be experiencing bias, 416 00:13:18,240 --> 00:13:20,200 even if we don't have labels for them. 417 00:13:20,200 --> 00:13:21,040 Exactly. 418 00:13:21,040 --> 00:13:23,800 AUTO D3M uses this information from PCA 419 00:13:23,800 --> 00:13:27,240 to actually create pseudo labels for those hidden groups. 420 00:13:27,240 --> 00:13:28,920 So we're essentially creating labels 421 00:13:28,920 --> 00:13:30,480 where none existed before. 422 00:13:30,480 --> 00:13:31,320 Precisely. 423 00:13:31,320 --> 00:13:32,960 Based on the way the data is structured. 424 00:13:32,960 --> 00:13:33,800 Yeah. 425 00:13:33,800 --> 00:13:35,640 And once you have those pseudo labels, 426 00:13:35,640 --> 00:13:38,200 you can then apply the same D3M technique, 427 00:13:38,200 --> 00:13:40,200 removing the most influential data points 428 00:13:40,200 --> 00:13:41,880 to create a more balanced data set. 429 00:13:41,880 --> 00:13:44,120 It's like giving a voice to those hidden groups. 430 00:13:44,120 --> 00:13:46,040 That's a powerful way to think about it. 431 00:13:46,040 --> 00:13:46,880 Yeah. 432 00:13:46,880 --> 00:13:47,960 And the researchers actually found 433 00:13:47,960 --> 00:13:50,200 that even with these pseudo labels, 434 00:13:50,200 --> 00:13:55,200 AUTO D3M was surprisingly effective at reducing bias. 435 00:13:55,360 --> 00:13:56,280 That's really encouraging. 436 00:13:56,280 --> 00:13:57,120 Yeah. 437 00:13:57,120 --> 00:13:58,520 It seems like we're making real progress 438 00:13:58,520 --> 00:14:01,800 in tackling this issue of bias in AI. 439 00:14:01,800 --> 00:14:02,680 Right. 440 00:14:02,680 --> 00:14:04,440 But you mentioned earlier that the researchers 441 00:14:04,440 --> 00:14:08,320 also look at some specific examples of data points 442 00:14:08,320 --> 00:14:10,400 that were being flagged by their methods. 443 00:14:10,400 --> 00:14:11,440 Can we talk about those? 444 00:14:11,440 --> 00:14:12,560 Absolutely. 445 00:14:12,560 --> 00:14:15,120 One of the most interesting findings was that D3M 446 00:14:15,120 --> 00:14:18,480 often flagged examples that had some sort of labeling error. 447 00:14:18,480 --> 00:14:21,080 For instance, in the Celeba Age data set, 448 00:14:21,080 --> 00:14:23,360 images of young people with gray hair 449 00:14:23,360 --> 00:14:26,560 were often identified as potentially problematic. 450 00:14:26,560 --> 00:14:27,400 Right. 451 00:14:27,400 --> 00:14:28,240 That makes sense. 452 00:14:28,240 --> 00:14:29,080 Yeah. 453 00:14:29,080 --> 00:14:31,240 If the models primarily learned to associate young 454 00:14:31,240 --> 00:14:33,240 with a certain hair color, 455 00:14:33,240 --> 00:14:34,760 seeing a young person with gray hair 456 00:14:34,760 --> 00:14:36,320 is going to throw it off. 457 00:14:36,320 --> 00:14:37,160 Right. 458 00:14:37,160 --> 00:14:39,920 And when they removed these mislabeled examples, 459 00:14:39,920 --> 00:14:42,160 they saw a significant improvement 460 00:14:42,160 --> 00:14:43,720 in the model's performance, 461 00:14:43,720 --> 00:14:45,080 particularly for those groups 462 00:14:45,080 --> 00:14:46,440 that were initially struggling. 463 00:14:46,440 --> 00:14:49,120 So sometimes the bias isn't in the data itself, 464 00:14:49,120 --> 00:14:50,560 but in the way it's been labeled. 465 00:14:50,560 --> 00:14:51,400 You got it. 466 00:14:51,400 --> 00:14:53,200 It's like that saying garbage in, garbage out. 467 00:14:53,200 --> 00:14:54,440 Exactly. 468 00:14:54,440 --> 00:14:57,440 We tend to focus so much on the algorithms and the models, 469 00:14:57,440 --> 00:15:00,480 but the data itself is the foundation. 470 00:15:00,480 --> 00:15:02,720 If that foundation is flawed, 471 00:15:02,720 --> 00:15:04,240 the whole structure is compromised. 472 00:15:04,240 --> 00:15:05,800 It's like building a house on sand. 473 00:15:05,800 --> 00:15:06,640 Yeah. 474 00:15:06,640 --> 00:15:07,480 No matter how strong the walls are, 475 00:15:07,480 --> 00:15:08,320 it's going to collapse eventually. 476 00:15:08,320 --> 00:15:09,160 Exactly. 477 00:15:09,160 --> 00:15:10,520 So data quality over quantity. 478 00:15:10,520 --> 00:15:11,360 Absolutely. 479 00:15:11,360 --> 00:15:13,520 This research really underscores that point. 480 00:15:13,520 --> 00:15:16,400 But they didn't stop at the Celeba Age data set. 481 00:15:16,400 --> 00:15:17,240 Okay. 482 00:15:17,240 --> 00:15:19,640 They also applied their methods to ImageNet. 483 00:15:19,640 --> 00:15:21,560 Oh yeah, ImageNet, the king of image data sets. 484 00:15:21,560 --> 00:15:22,480 Exactly. 485 00:15:22,480 --> 00:15:26,480 Now ImageNet doesn't come with any predefined bias labels. 486 00:15:26,480 --> 00:15:29,760 So it was a perfect testing ground for their methods. 487 00:15:29,760 --> 00:15:32,320 And they discovered some interesting things. 488 00:15:32,320 --> 00:15:35,600 For example, they found that the model was overly reliant 489 00:15:35,600 --> 00:15:39,240 on the color red when classifying red wolf. 490 00:15:39,240 --> 00:15:41,040 So it's like the model was making assumptions 491 00:15:41,040 --> 00:15:42,720 based on superficial features. 492 00:15:42,720 --> 00:15:43,560 Exactly. 493 00:15:43,560 --> 00:15:45,400 Instead of actually understanding what a red wolf is. 494 00:15:45,400 --> 00:15:48,040 It was picking up on a spurious correlation 495 00:15:48,040 --> 00:15:49,800 that wasn't really indicative 496 00:15:49,800 --> 00:15:52,200 of the true characteristics of a red wolf. 497 00:15:52,200 --> 00:15:53,040 Right. 498 00:15:53,040 --> 00:15:55,080 And they also found that the model was associating 499 00:15:55,080 --> 00:15:59,680 tinch fish with images that also contained humans. 500 00:15:59,680 --> 00:16:00,520 Oh wow. 501 00:16:00,520 --> 00:16:02,400 Even though the presence of a human 502 00:16:02,400 --> 00:16:05,840 is completely irrelevant to whether or not a tinch is present. 503 00:16:05,840 --> 00:16:07,400 So it's like the model was learning 504 00:16:07,400 --> 00:16:10,200 these weird random patterns in the data. 505 00:16:10,200 --> 00:16:11,040 Precisely. 506 00:16:11,040 --> 00:16:14,640 That didn't actually reflect any real world relationship. 507 00:16:14,640 --> 00:16:15,480 Yeah. 508 00:16:15,480 --> 00:16:17,200 And these kinds of spurious correlations 509 00:16:17,200 --> 00:16:19,720 can really skew the model's predictions, 510 00:16:19,720 --> 00:16:21,520 especially when we're using these models 511 00:16:21,520 --> 00:16:23,960 to make decisions that impact people's lives. 512 00:16:23,960 --> 00:16:25,440 That's a little unnerving to be honest. 513 00:16:25,440 --> 00:16:26,280 It is. 514 00:16:26,280 --> 00:16:30,360 But thankfully D3M and IUTO D3M were up to the challenge. 515 00:16:30,360 --> 00:16:31,200 They were. 516 00:16:31,200 --> 00:16:33,400 By applying their methods to ImageNet, 517 00:16:33,400 --> 00:16:36,720 they were able to significantly improve the model's accuracy 518 00:16:36,720 --> 00:16:38,280 on those specific categories 519 00:16:38,280 --> 00:16:41,640 all without negatively impacting the overall performance. 520 00:16:41,640 --> 00:16:42,480 Okay. 521 00:16:42,480 --> 00:16:44,800 So it's a promising sign that these techniques 522 00:16:44,800 --> 00:16:47,920 can be applied to a variety of data sets and tasks. 523 00:16:47,920 --> 00:16:49,800 That is good news. 524 00:16:49,800 --> 00:16:51,400 But like you said earlier, 525 00:16:51,400 --> 00:16:54,600 we need to be cautious about assuming one solution fits all. 526 00:16:54,600 --> 00:16:55,440 Right. 527 00:16:55,440 --> 00:16:58,000 What are some of the potential pitfalls or challenges 528 00:16:58,000 --> 00:17:00,440 we need to consider as we move forward? 529 00:17:00,440 --> 00:17:02,040 So where do we go from here? 530 00:17:02,040 --> 00:17:03,280 Yeah, that's what I'm thinking. 531 00:17:03,280 --> 00:17:05,640 We've explored the problem, the solution, 532 00:17:05,640 --> 00:17:07,640 even some real world examples. 533 00:17:07,640 --> 00:17:09,800 But I'm curious about like what's next. 534 00:17:09,800 --> 00:17:10,920 That's the exciting part. 535 00:17:10,920 --> 00:17:13,880 This research opens up so many possibilities. 536 00:17:13,880 --> 00:17:17,840 And there are tons of like fascinating directions to explore. 537 00:17:17,840 --> 00:17:19,400 Give us a glimpse into the future. 538 00:17:19,400 --> 00:17:20,240 Okay. 539 00:17:20,240 --> 00:17:21,320 What are some of the most, you know, 540 00:17:21,320 --> 00:17:24,000 promising avenues for research? 541 00:17:24,000 --> 00:17:25,920 One area that's right for exploration 542 00:17:25,920 --> 00:17:30,280 is applying these techniques to other types of AI models 543 00:17:30,280 --> 00:17:33,040 beyond just image and text classification. 544 00:17:33,040 --> 00:17:33,880 Okay. 545 00:17:33,880 --> 00:17:38,680 So could you use D3M to de-bias a recommendation system, 546 00:17:38,680 --> 00:17:39,880 for instance, 547 00:17:39,880 --> 00:17:41,080 to make sure it's not, you know, 548 00:17:41,080 --> 00:17:43,880 unfairly promoting certain products or services? 549 00:17:43,880 --> 00:17:44,720 That's a great point. 550 00:17:44,720 --> 00:17:47,640 So instead of just tweaking like images or text, 551 00:17:47,640 --> 00:17:50,760 we could use these methods to make sure algorithms are fair 552 00:17:50,760 --> 00:17:52,360 across all sorts of applications. 553 00:17:52,360 --> 00:17:53,200 Exactly. 554 00:17:53,200 --> 00:17:56,600 Think about the potential impact on fields like healthcare, 555 00:17:56,600 --> 00:17:59,400 finance, even criminal justice. 556 00:17:59,400 --> 00:18:02,920 Imagine a world where algorithms are actually actively working 557 00:18:02,920 --> 00:18:05,440 to promote fairness and equity 558 00:18:05,440 --> 00:18:08,080 rather than perpetuating existing biases. 559 00:18:08,080 --> 00:18:09,200 That's a powerful vision. 560 00:18:09,200 --> 00:18:10,040 Yeah. 561 00:18:10,040 --> 00:18:11,800 But I'm guessing there are some hurdles to overcome 562 00:18:11,800 --> 00:18:12,720 before we get there. 563 00:18:12,720 --> 00:18:14,920 Of course, one challenge is scalability. 564 00:18:14,920 --> 00:18:15,760 Okay. 565 00:18:15,760 --> 00:18:17,360 These methods work well on, you know, 566 00:18:17,360 --> 00:18:19,000 relatively small data sets. 567 00:18:19,000 --> 00:18:22,400 But how do you apply them to the massive data sets 568 00:18:22,400 --> 00:18:25,320 used to train today's most powerful AI models? 569 00:18:25,320 --> 00:18:27,640 It's like trying to bake a cake the size of a house. 570 00:18:27,640 --> 00:18:28,480 Yeah. 571 00:18:28,480 --> 00:18:30,000 Using a recipe designed for a single cupcake. 572 00:18:30,000 --> 00:18:31,520 That's a perfect analogy. 573 00:18:31,520 --> 00:18:32,360 Right. 574 00:18:32,360 --> 00:18:35,520 We need to develop more efficient and scalable methods 575 00:18:35,520 --> 00:18:39,800 for analyzing and debiasing these huge data sets. 576 00:18:39,800 --> 00:18:41,600 So a bit of a computational mountain to climb. 577 00:18:41,600 --> 00:18:42,440 Exactly. 578 00:18:42,440 --> 00:18:45,920 And another challenge is the need for more robust 579 00:18:45,920 --> 00:18:47,440 and reliable data models. 580 00:18:47,440 --> 00:18:48,280 Right. 581 00:18:48,280 --> 00:18:51,720 As we discussed earlier, these models are still approximations 582 00:18:51,720 --> 00:18:54,120 of how the AI actually works. 583 00:18:54,120 --> 00:18:54,960 Yeah. 584 00:18:54,960 --> 00:18:57,600 There's always that chance of like misinterpretation 585 00:18:57,600 --> 00:18:58,440 or error. 586 00:18:58,440 --> 00:18:59,280 Right. 587 00:18:59,280 --> 00:19:00,960 We need to make sure our AI fortune tellers 588 00:19:00,960 --> 00:19:02,760 are giving us accurate predictions. 589 00:19:02,760 --> 00:19:03,600 Precisely. 590 00:19:03,600 --> 00:19:06,600 We need to invest in research that improves the accuracy 591 00:19:06,600 --> 00:19:08,560 and reliability of these data models 592 00:19:08,560 --> 00:19:11,720 so we can be more confident in that debiasing process. 593 00:19:11,720 --> 00:19:13,560 So it's about sharpening our tools 594 00:19:13,560 --> 00:19:15,560 as we explore new applications. 595 00:19:15,560 --> 00:19:16,400 Exactly. 596 00:19:16,400 --> 00:19:19,440 But are there any like broader societal considerations 597 00:19:19,440 --> 00:19:21,960 we need to keep in mind as this research progresses? 598 00:19:21,960 --> 00:19:23,000 Absolutely. 599 00:19:23,000 --> 00:19:26,000 As AI becomes more integrated into our lives, 600 00:19:26,000 --> 00:19:29,560 it's crucial to have open and transparent conversations 601 00:19:29,560 --> 00:19:32,720 about the ethical implications of these technologies. 602 00:19:32,720 --> 00:19:35,360 It's not just about building AI that works. 603 00:19:35,360 --> 00:19:38,080 It's about building AI that aligns with our values. 604 00:19:38,080 --> 00:19:39,240 Well said. 605 00:19:39,240 --> 00:19:43,960 We need to work with ethicists, social scientists, 606 00:19:43,960 --> 00:19:47,840 policymakers to ensure that AI is developed and deployed 607 00:19:47,840 --> 00:19:50,080 in a way that benefits all of humanity. 608 00:19:50,080 --> 00:19:52,520 So this research isn't just a technical achievement. 609 00:19:52,520 --> 00:19:53,360 Right. 610 00:19:53,360 --> 00:19:56,360 It's a starting point for a much larger conversation 611 00:19:56,360 --> 00:19:57,600 about the future of AI. 612 00:19:57,600 --> 00:20:00,600 It's a powerful reminder that we have a responsibility 613 00:20:00,600 --> 00:20:03,480 to use these amazing technologies wisely 614 00:20:03,480 --> 00:20:05,320 and to make sure they're used to create 615 00:20:05,320 --> 00:20:07,520 a more just and equitable world. 616 00:20:07,520 --> 00:20:08,480 That's a great note to end on. 617 00:20:08,480 --> 00:20:10,360 This has been a truly insightful deep dive 618 00:20:10,360 --> 00:20:12,640 into the world of AI debiasing. 619 00:20:12,640 --> 00:20:13,480 It has. 620 00:20:13,480 --> 00:20:15,840 We have explored the challenges, the solutions, 621 00:20:15,840 --> 00:20:18,400 and the incredible potential for a future 622 00:20:18,400 --> 00:20:20,280 where AI works for everyone. 623 00:20:20,280 --> 00:20:22,160 It's been a pleasure exploring this with you. 624 00:20:22,160 --> 00:20:25,480 We hope this deep dive has sparked your curiosity 625 00:20:25,480 --> 00:20:27,000 and encouraged you to think critically 626 00:20:27,000 --> 00:20:29,040 about the role of AI in our lives. 627 00:20:29,040 --> 00:20:31,440 And remember, the conversation doesn't stop here. 628 00:20:31,440 --> 00:20:33,920 Stay engaged, stay informed, 629 00:20:33,920 --> 00:20:35,840 and let's work together to build a future 630 00:20:35,840 --> 00:20:59,840 where AI empowers us all.