diff --git "a/MLwG/Lecture 1 - Deep Learning Foundations\357\274\232 review of basic DL models and optimization solvers [6Nctj1GCFVo].en.vtt" "b/MLwG/Lecture 1 - Deep Learning Foundations\357\274\232 review of basic DL models and optimization solvers [6Nctj1GCFVo].en.vtt" new file mode 100644 index 0000000000000000000000000000000000000000..3a0804c81fa531e80342025d6d9c5671767b8436 --- /dev/null +++ "b/MLwG/Lecture 1 - Deep Learning Foundations\357\274\232 review of basic DL models and optimization solvers [6Nctj1GCFVo].en.vtt" @@ -0,0 +1,13600 @@ +WEBVTT +Kind: captions +Language: en + +00:00:08.639 --> 00:00:11.190 align:start position:0% + +you<00:00:08.800> guys<00:00:08.960> hear<00:00:09.200> me<00:00:09.360> well + +00:00:11.190 --> 00:00:11.200 align:start position:0% +you guys hear me well + + +00:00:11.200 --> 00:00:14.709 align:start position:0% +you guys hear me well +yes<00:00:12.000> yes<00:00:12.559> yes<00:00:13.120> perfect + +00:00:14.709 --> 00:00:14.719 align:start position:0% +yes yes yes perfect + + +00:00:14.719 --> 00:00:17.670 align:start position:0% +yes yes yes perfect +all<00:00:14.799> right<00:00:15.120> so<00:00:15.519> uh<00:00:16.640> this<00:00:16.880> is<00:00:17.039> foundations<00:00:17.600> of + +00:00:17.670 --> 00:00:17.680 align:start position:0% +all right so uh this is foundations of + + +00:00:17.680 --> 00:00:19.429 align:start position:0% +all right so uh this is foundations of +deep<00:00:17.920> learning<00:00:18.320> course<00:00:18.640> my<00:00:18.800> name<00:00:18.960> is<00:00:19.119> joelle + +00:00:19.429 --> 00:00:19.439 align:start position:0% +deep learning course my name is joelle + + +00:00:19.439 --> 00:00:21.590 align:start position:0% +deep learning course my name is joelle +faizi<00:00:19.920> and<00:00:20.080> i'm<00:00:20.240> very<00:00:20.480> happy<00:00:20.880> to<00:00:21.199> be<00:00:21.359> your + +00:00:21.590 --> 00:00:21.600 align:start position:0% +faizi and i'm very happy to be your + + +00:00:21.600 --> 00:00:23.349 align:start position:0% +faizi and i'm very happy to be your +instructor<00:00:22.160> in<00:00:22.240> this<00:00:22.480> course + +00:00:23.349 --> 00:00:23.359 align:start position:0% +instructor in this course + + +00:00:23.359 --> 00:00:26.150 align:start position:0% +instructor in this course +i<00:00:23.600> hope<00:00:24.240> you<00:00:24.480> have<00:00:24.720> had<00:00:24.960> a<00:00:25.039> good<00:00:25.359> summer<00:00:25.920> and<00:00:26.080> i + +00:00:26.150 --> 00:00:26.160 align:start position:0% +i hope you have had a good summer and i + + +00:00:26.160 --> 00:00:28.310 align:start position:0% +i hope you have had a good summer and i +hope<00:00:26.400> you<00:00:26.640> and<00:00:26.800> your<00:00:26.960> family<00:00:27.279> are<00:00:27.439> doing<00:00:27.680> well + +00:00:28.310 --> 00:00:28.320 align:start position:0% +hope you and your family are doing well + + +00:00:28.320 --> 00:00:31.509 align:start position:0% +hope you and your family are doing well +in<00:00:28.480> these<00:00:29.119> complicated<00:00:29.760> situations + +00:00:31.509 --> 00:00:31.519 align:start position:0% +in these complicated situations + + +00:00:31.519 --> 00:00:34.389 align:start position:0% +in these complicated situations +uh<00:00:32.079> so<00:00:32.239> let's<00:00:32.640> start<00:00:32.960> so<00:00:33.360> today<00:00:33.760> i'm<00:00:34.000> going<00:00:34.160> to + +00:00:34.389 --> 00:00:34.399 align:start position:0% +uh so let's start so today i'm going to + + +00:00:34.399 --> 00:00:35.270 align:start position:0% +uh so let's start so today i'm going to +first + +00:00:35.270 --> 00:00:35.280 align:start position:0% +first + + +00:00:35.280 --> 00:00:37.590 align:start position:0% +first +briefly<00:00:35.840> talk<00:00:36.079> about<00:00:36.399> the<00:00:36.800> structure<00:00:37.280> of<00:00:37.360> the + +00:00:37.590 --> 00:00:37.600 align:start position:0% +briefly talk about the structure of the + + +00:00:37.600 --> 00:00:38.470 align:start position:0% +briefly talk about the structure of the +course + +00:00:38.470 --> 00:00:38.480 align:start position:0% +course + + +00:00:38.480 --> 00:00:42.470 align:start position:0% +course +and<00:00:38.879> also<00:00:39.760> um<00:00:40.640> some<00:00:41.360> top<00:00:41.680> topics<00:00:42.160> that<00:00:42.239> we<00:00:42.399> are + +00:00:42.470 --> 00:00:42.480 align:start position:0% +and also um some top topics that we are + + +00:00:42.480 --> 00:00:43.990 align:start position:0% +and also um some top topics that we are +going<00:00:42.640> to<00:00:42.840> cover + +00:00:43.990 --> 00:00:44.000 align:start position:0% +going to cover + + +00:00:44.000 --> 00:00:46.549 align:start position:0% +going to cover +and<00:00:44.079> then<00:00:44.320> we'll<00:00:44.800> basically<00:00:45.760> dive<00:00:46.079> into<00:00:46.320> some + +00:00:46.549 --> 00:00:46.559 align:start position:0% +and then we'll basically dive into some + + +00:00:46.559 --> 00:00:47.430 align:start position:0% +and then we'll basically dive into some +technical + +00:00:47.430 --> 00:00:47.440 align:start position:0% +technical + + +00:00:47.440 --> 00:00:49.190 align:start position:0% +technical +portions<00:00:47.840> of<00:00:48.000> this<00:00:48.160> lecture<00:00:48.640> in<00:00:48.719> terms<00:00:49.039> of + +00:00:49.190 --> 00:00:49.200 align:start position:0% +portions of this lecture in terms of + + +00:00:49.200 --> 00:00:51.830 align:start position:0% +portions of this lecture in terms of +understanding<00:00:50.160> basic<00:00:50.559> deep<00:00:50.800> learning + +00:00:51.830 --> 00:00:51.840 align:start position:0% +understanding basic deep learning + + +00:00:51.840 --> 00:00:54.790 align:start position:0% +understanding basic deep learning +models<00:00:52.239> and<00:00:52.399> optimization<00:00:52.960> problem<00:00:53.920> so<00:00:54.480> first + +00:00:54.790 --> 00:00:54.800 align:start position:0% +models and optimization problem so first + + +00:00:54.800 --> 00:00:55.270 align:start position:0% +models and optimization problem so first +of<00:00:54.960> all + +00:00:55.270 --> 00:00:55.280 align:start position:0% +of all + + +00:00:55.280 --> 00:00:58.950 align:start position:0% +of all +um<00:00:57.039> this<00:00:57.280> is<00:00:57.440> an<00:00:57.680> advanced + +00:00:58.950 --> 00:00:58.960 align:start position:0% +um this is an advanced + + +00:00:58.960 --> 00:01:01.189 align:start position:0% +um this is an advanced +deep<00:00:59.280> learning<00:00:59.680> course<00:01:00.000> so<00:01:00.399> you<00:01:00.559> need<00:01:00.800> to<00:01:00.960> have + +00:01:01.189 --> 00:01:01.199 align:start position:0% +deep learning course so you need to have + + +00:01:01.199 --> 00:01:01.990 align:start position:0% +deep learning course so you need to have +some + +00:01:01.990 --> 00:01:02.000 align:start position:0% +some + + +00:01:02.000 --> 00:01:04.070 align:start position:0% +some +basic<00:01:02.480> machine<00:01:02.879> learning<00:01:03.280> and<00:01:03.440> deep<00:01:03.680> learning + +00:01:04.070 --> 00:01:04.080 align:start position:0% +basic machine learning and deep learning + + +00:01:04.080 --> 00:01:05.509 align:start position:0% +basic machine learning and deep learning +background<00:01:04.720> before + +00:01:05.509 --> 00:01:05.519 align:start position:0% +background before + + +00:01:05.519 --> 00:01:07.190 align:start position:0% +background before +taking<00:01:05.840> this<00:01:06.080> course<00:01:06.479> if<00:01:06.640> you<00:01:06.720> don't<00:01:06.960> have + +00:01:07.190 --> 00:01:07.200 align:start position:0% +taking this course if you don't have + + +00:01:07.200 --> 00:01:08.710 align:start position:0% +taking this course if you don't have +such<00:01:07.439> a<00:01:07.520> background<00:01:08.080> i + +00:01:08.710 --> 00:01:08.720 align:start position:0% +such a background i + + +00:01:08.720 --> 00:01:10.870 align:start position:0% +such a background i +encourage<00:01:09.200> you<00:01:09.360> to<00:01:09.680> first<00:01:10.000> take<00:01:10.240> some + +00:01:10.870 --> 00:01:10.880 align:start position:0% +encourage you to first take some + + +00:01:10.880 --> 00:01:12.710 align:start position:0% +encourage you to first take some +introductory<00:01:11.680> course<00:01:11.920> in<00:01:12.080> machine<00:01:12.320> learning + +00:01:12.710 --> 00:01:12.720 align:start position:0% +introductory course in machine learning + + +00:01:12.720 --> 00:01:13.510 align:start position:0% +introductory course in machine learning +and<00:01:12.799> deep<00:01:12.960> learning + +00:01:13.510 --> 00:01:13.520 align:start position:0% +and deep learning + + +00:01:13.520 --> 00:01:16.390 align:start position:0% +and deep learning +and<00:01:13.760> maybe<00:01:14.240> the<00:01:14.400> next<00:01:14.640> years<00:01:14.960> you<00:01:15.040> can<00:01:15.280> take + +00:01:16.390 --> 00:01:16.400 align:start position:0% +and maybe the next years you can take + + +00:01:16.400 --> 00:01:19.030 align:start position:0% +and maybe the next years you can take +advanced<00:01:16.880> courses<00:01:17.280> in<00:01:17.360> deep<00:01:17.600> learning + +00:01:19.030 --> 00:01:19.040 align:start position:0% +advanced courses in deep learning + + +00:01:19.040 --> 00:01:20.950 align:start position:0% +advanced courses in deep learning +so<00:01:19.280> topics<00:01:19.680> that<00:01:19.759> we<00:01:19.920> are<00:01:20.000> going<00:01:20.240> to<00:01:20.400> cover + +00:01:20.950 --> 00:01:20.960 align:start position:0% +so topics that we are going to cover + + +00:01:20.960 --> 00:01:23.270 align:start position:0% +so topics that we are going to cover +these<00:01:21.200> are<00:01:21.520> pretty<00:01:22.000> advanced<00:01:22.640> and<00:01:22.799> modern + +00:01:23.270 --> 00:01:23.280 align:start position:0% +these are pretty advanced and modern + + +00:01:23.280 --> 00:01:24.710 align:start position:0% +these are pretty advanced and modern +topics + +00:01:24.710 --> 00:01:24.720 align:start position:0% +topics + + +00:01:24.720 --> 00:01:27.109 align:start position:0% +topics +including<00:01:25.439> deep<00:01:25.680> learning<00:01:26.159> optimization + +00:01:27.109 --> 00:01:27.119 align:start position:0% +including deep learning optimization + + +00:01:27.119 --> 00:01:28.950 align:start position:0% +including deep learning optimization +deep<00:01:27.360> learning<00:01:27.680> generalization + +00:01:28.950 --> 00:01:28.960 align:start position:0% +deep learning generalization + + +00:01:28.960 --> 00:01:32.230 align:start position:0% +deep learning generalization +neural<00:01:29.360> tangent<00:01:29.759> kernels<00:01:30.720> lottery<00:01:31.119> tickets + +00:01:32.230 --> 00:01:32.240 align:start position:0% +neural tangent kernels lottery tickets + + +00:01:32.240 --> 00:01:34.469 align:start position:0% +neural tangent kernels lottery tickets +uh<00:01:32.479> deep<00:01:32.640> learning<00:01:32.960> robustness<00:01:34.000> in<00:01:34.079> terms<00:01:34.320> of + +00:01:34.469 --> 00:01:34.479 align:start position:0% +uh deep learning robustness in terms of + + +00:01:34.479 --> 00:01:37.030 align:start position:0% +uh deep learning robustness in terms of +adversarial<00:01:35.119> attacks<00:01:35.520> and<00:01:35.600> defenses + +00:01:37.030 --> 00:01:37.040 align:start position:0% +adversarial attacks and defenses + + +00:01:37.040 --> 00:01:39.109 align:start position:0% +adversarial attacks and defenses +then<00:01:37.200> we<00:01:37.360> are<00:01:37.520> going<00:01:37.680> to<00:01:37.840> look<00:01:38.079> into<00:01:38.400> relaxing + +00:01:39.109 --> 00:01:39.119 align:start position:0% +then we are going to look into relaxing + + +00:01:39.119 --> 00:01:40.069 align:start position:0% +then we are going to look into relaxing +the<00:01:39.280> iid + +00:01:40.069 --> 00:01:40.079 align:start position:0% +the iid + + +00:01:40.079 --> 00:01:43.109 align:start position:0% +the iid +assumption<00:01:41.600> in<00:01:41.680> terms<00:01:42.000> of<00:01:42.079> domain<00:01:42.479> adaptation + +00:01:43.109 --> 00:01:43.119 align:start position:0% +assumption in terms of domain adaptation + + +00:01:43.119 --> 00:01:45.590 align:start position:0% +assumption in terms of domain adaptation +and<00:01:43.280> out<00:01:43.520> of<00:01:43.759> domain<00:01:44.159> generalization + +00:01:45.590 --> 00:01:45.600 align:start position:0% +and out of domain generalization + + +00:01:45.600 --> 00:01:47.590 align:start position:0% +and out of domain generalization +we<00:01:45.759> will<00:01:46.079> spend<00:01:46.479> a<00:01:46.560> couple<00:01:46.799> of<00:01:46.960> lectures<00:01:47.360> on + +00:01:47.590 --> 00:01:47.600 align:start position:0% +we will spend a couple of lectures on + + +00:01:47.600 --> 00:01:49.030 align:start position:0% +we will spend a couple of lectures on +deep<00:01:47.840> generative<00:01:48.320> models + +00:01:49.030 --> 00:01:49.040 align:start position:0% +deep generative models + + +00:01:49.040 --> 00:01:52.950 align:start position:0% +deep generative models +vaes<00:01:49.840> gans<00:01:50.799> flow<00:01:51.119> based<00:01:51.439> models + +00:01:52.950 --> 00:01:52.960 align:start position:0% +vaes gans flow based models + + +00:01:52.960 --> 00:01:55.350 align:start position:0% +vaes gans flow based models +and<00:01:53.040> their<00:01:53.360> mean<00:01:53.520> max<00:01:53.840> optimization<00:01:54.479> problems + +00:01:55.350 --> 00:01:55.360 align:start position:0% +and their mean max optimization problems + + +00:01:55.360 --> 00:01:56.709 align:start position:0% +and their mean max optimization problems +we'll<00:01:55.680> spend<00:01:55.920> the<00:01:56.159> lecture<00:01:56.479> on + +00:01:56.709 --> 00:01:56.719 align:start position:0% +we'll spend the lecture on + + +00:01:56.719 --> 00:01:58.950 align:start position:0% +we'll spend the lecture on +self-supervised<00:01:57.680> deep<00:01:57.920> learning + +00:01:58.950 --> 00:01:58.960 align:start position:0% +self-supervised deep learning + + +00:01:58.960 --> 00:02:01.109 align:start position:0% +self-supervised deep learning +and<00:01:59.119> a<00:01:59.200> few<00:01:59.439> lecture<00:01:59.920> on<00:02:00.079> deep<00:02:00.320> reinforcement + +00:02:01.109 --> 00:02:01.119 align:start position:0% +and a few lecture on deep reinforcement + + +00:02:01.119 --> 00:02:02.630 align:start position:0% +and a few lecture on deep reinforcement +learning + +00:02:02.630 --> 00:02:02.640 align:start position:0% +learning + + +00:02:02.640 --> 00:02:06.389 align:start position:0% +learning +and<00:02:04.079> we'll<00:02:04.320> have<00:02:04.479> one<00:02:04.719> lecture<00:02:05.119> on<00:02:05.520> attentions + +00:02:06.389 --> 00:02:06.399 align:start position:0% +and we'll have one lecture on attentions + + +00:02:06.399 --> 00:02:09.830 align:start position:0% +and we'll have one lecture on attentions +and<00:02:06.719> transformers<00:02:07.520> and<00:02:07.759> lstms<00:02:09.280> vision<00:02:09.599> deep + +00:02:09.830 --> 00:02:09.840 align:start position:0% +and transformers and lstms vision deep + + +00:02:09.840 --> 00:02:10.630 align:start position:0% +and transformers and lstms vision deep +learning + +00:02:10.630 --> 00:02:10.640 align:start position:0% +learning + + +00:02:10.640 --> 00:02:15.030 align:start position:0% +learning +explainable<00:02:12.080> deep<00:02:12.319> learning<00:02:13.440> fairness + +00:02:15.030 --> 00:02:15.040 align:start position:0% +explainable deep learning fairness + + +00:02:15.040 --> 00:02:17.270 align:start position:0% +explainable deep learning fairness +graph<00:02:15.360> neural<00:02:15.680> networks<00:02:16.480> meta<00:02:16.800> learning + +00:02:17.270 --> 00:02:17.280 align:start position:0% +graph neural networks meta learning + + +00:02:17.280 --> 00:02:19.670 align:start position:0% +graph neural networks meta learning +federated<00:02:17.840> learning<00:02:18.319> privacy<00:02:18.800> analytics + +00:02:19.670 --> 00:02:19.680 align:start position:0% +federated learning privacy analytics + + +00:02:19.680 --> 00:02:22.550 align:start position:0% +federated learning privacy analytics +so<00:02:19.840> as<00:02:20.000> you<00:02:20.080> can<00:02:20.239> see<00:02:20.959> this<00:02:21.200> is<00:02:21.360> a<00:02:21.520> very<00:02:22.000> long + +00:02:22.550 --> 00:02:22.560 align:start position:0% +so as you can see this is a very long + + +00:02:22.560 --> 00:02:23.750 align:start position:0% +so as you can see this is a very long +list<00:02:22.959> of + +00:02:23.750 --> 00:02:23.760 align:start position:0% +list of + + +00:02:23.760 --> 00:02:25.910 align:start position:0% +list of +topics<00:02:24.319> these<00:02:24.560> topics<00:02:25.120> are<00:02:25.280> very<00:02:25.520> important + +00:02:25.910 --> 00:02:25.920 align:start position:0% +topics these topics are very important + + +00:02:25.920 --> 00:02:27.030 align:start position:0% +topics these topics are very important +that<00:02:26.160> modern + +00:02:27.030 --> 00:02:27.040 align:start position:0% +that modern + + +00:02:27.040 --> 00:02:30.229 align:start position:0% +that modern +but<00:02:27.280> in<00:02:27.440> order<00:02:27.760> to<00:02:28.560> be<00:02:28.720> able<00:02:28.959> to<00:02:29.599> cover<00:02:30.000> these + +00:02:30.229 --> 00:02:30.239 align:start position:0% +but in order to be able to cover these + + +00:02:30.239 --> 00:02:31.110 align:start position:0% +but in order to be able to cover these +topics<00:02:30.720> i + +00:02:31.110 --> 00:02:31.120 align:start position:0% +topics i + + +00:02:31.120 --> 00:02:34.470 align:start position:0% +topics i +definitely<00:02:31.760> need<00:02:31.920> your<00:02:32.080> help<00:02:32.720> so<00:02:32.959> you<00:02:33.280> need<00:02:33.519> to + +00:02:34.470 --> 00:02:34.480 align:start position:0% +definitely need your help so you need to + + +00:02:34.480 --> 00:02:36.790 align:start position:0% +definitely need your help so you need to +read<00:02:34.800> the<00:02:35.040> papers<00:02:35.519> that<00:02:35.680> we<00:02:35.840> are<00:02:36.000> posting<00:02:36.560> for + +00:02:36.790 --> 00:02:36.800 align:start position:0% +read the papers that we are posting for + + +00:02:36.800 --> 00:02:38.229 align:start position:0% +read the papers that we are posting for +each<00:02:37.120> lecture + +00:02:38.229 --> 00:02:38.239 align:start position:0% +each lecture + + +00:02:38.239 --> 00:02:40.949 align:start position:0% +each lecture +before<00:02:38.879> the<00:02:39.040> lecture<00:02:39.680> so<00:02:39.920> i<00:02:40.080> cannot<00:02:40.480> emphasize + +00:02:40.949 --> 00:02:40.959 align:start position:0% +before the lecture so i cannot emphasize + + +00:02:40.959 --> 00:02:41.670 align:start position:0% +before the lecture so i cannot emphasize +it<00:02:41.120> more + +00:02:41.670 --> 00:02:41.680 align:start position:0% +it more + + +00:02:41.680 --> 00:02:43.830 align:start position:0% +it more +that<00:02:42.000> it<00:02:42.160> is<00:02:42.319> very<00:02:42.560> important<00:02:42.959> that<00:02:43.120> you<00:02:43.599> read + +00:02:43.830 --> 00:02:43.840 align:start position:0% +that it is very important that you read + + +00:02:43.840 --> 00:02:46.390 align:start position:0% +that it is very important that you read +these<00:02:44.080> papers<00:02:44.560> before<00:02:44.959> coming<00:02:45.280> to<00:02:45.440> the<00:02:45.599> class + +00:02:46.390 --> 00:02:46.400 align:start position:0% +these papers before coming to the class + + +00:02:46.400 --> 00:02:48.550 align:start position:0% +these papers before coming to the class +and<00:02:46.560> some<00:02:46.720> of<00:02:46.879> these<00:02:47.040> papers<00:02:47.519> are<00:02:48.160> pretty + +00:02:48.550 --> 00:02:48.560 align:start position:0% +and some of these papers are pretty + + +00:02:48.560 --> 00:02:50.150 align:start position:0% +and some of these papers are pretty +technical<00:02:49.360> so + +00:02:50.150 --> 00:02:50.160 align:start position:0% +technical so + + +00:02:50.160 --> 00:02:53.430 align:start position:0% +technical so +uh<00:02:50.640> in<00:02:51.280> uh<00:02:51.599> in<00:02:51.760> order<00:02:52.080> to<00:02:52.319> be<00:02:52.480> able<00:02:52.640> to<00:02:52.879> digest + +00:02:53.430 --> 00:02:53.440 align:start position:0% +uh in uh in order to be able to digest + + +00:02:53.440 --> 00:02:54.630 align:start position:0% +uh in uh in order to be able to digest +these<00:02:53.680> papers<00:02:54.080> you<00:02:54.160> should + +00:02:54.630 --> 00:02:54.640 align:start position:0% +these papers you should + + +00:02:54.640 --> 00:02:56.470 align:start position:0% +these papers you should +start<00:02:54.959> reading<00:02:55.360> it<00:02:55.599> you<00:02:55.680> know<00:02:55.840> couple<00:02:56.160> of<00:02:56.239> days + +00:02:56.470 --> 00:02:56.480 align:start position:0% +start reading it you know couple of days + + +00:02:56.480 --> 00:02:58.710 align:start position:0% +start reading it you know couple of days +before<00:02:56.800> the<00:02:56.959> lecture<00:02:57.599> so<00:02:57.840> don't<00:02:58.080> wait<00:02:58.400> until + +00:02:58.710 --> 00:02:58.720 align:start position:0% +before the lecture so don't wait until + + +00:02:58.720 --> 00:03:00.790 align:start position:0% +before the lecture so don't wait until +the<00:02:58.879> last<00:02:59.200> hour<00:02:59.519> before<00:02:59.840> the<00:03:00.000> lecture<00:03:00.400> because + +00:03:00.790 --> 00:03:00.800 align:start position:0% +the last hour before the lecture because + + +00:03:00.800 --> 00:03:01.750 align:start position:0% +the last hour before the lecture because +in<00:03:00.879> that<00:03:01.120> case + +00:03:01.750 --> 00:03:01.760 align:start position:0% +in that case + + +00:03:01.760 --> 00:03:04.550 align:start position:0% +in that case +probably<00:03:02.239> it<00:03:02.319> will<00:03:02.480> be<00:03:02.720> hard<00:03:03.040> to<00:03:04.000> understand + +00:03:04.550 --> 00:03:04.560 align:start position:0% +probably it will be hard to understand + + +00:03:04.560 --> 00:03:05.670 align:start position:0% +probably it will be hard to understand +and<00:03:04.720> digest<00:03:05.280> the + +00:03:05.670 --> 00:03:05.680 align:start position:0% +and digest the + + +00:03:05.680 --> 00:03:09.110 align:start position:0% +and digest the +concepts<00:03:06.080> and<00:03:06.239> techniques<00:03:06.720> of<00:03:07.360> these<00:03:07.599> papers + +00:03:09.110 --> 00:03:09.120 align:start position:0% +concepts and techniques of these papers + + +00:03:09.120 --> 00:03:11.030 align:start position:0% +concepts and techniques of these papers +so<00:03:09.280> another<00:03:09.680> thing<00:03:09.920> is<00:03:10.000> that<00:03:10.159> these<00:03:10.400> papers + +00:03:11.030 --> 00:03:11.040 align:start position:0% +so another thing is that these papers + + +00:03:11.040 --> 00:03:12.229 align:start position:0% +so another thing is that these papers +and<00:03:11.360> topics + +00:03:12.229 --> 00:03:12.239 align:start position:0% +and topics + + +00:03:12.239 --> 00:03:15.110 align:start position:0% +and topics +it's<00:03:12.400> a<00:03:12.800> tentative<00:03:13.440> list<00:03:14.000> so<00:03:14.159> we<00:03:14.319> may<00:03:14.560> add<00:03:14.879> or + +00:03:15.110 --> 00:03:15.120 align:start position:0% +it's a tentative list so we may add or + + +00:03:15.120 --> 00:03:17.910 align:start position:0% +it's a tentative list so we may add or +remove<00:03:15.519> some<00:03:15.760> of<00:03:15.840> the<00:03:16.000> topics<00:03:16.560> along<00:03:16.879> the<00:03:17.040> way + +00:03:17.910 --> 00:03:17.920 align:start position:0% +remove some of the topics along the way + + +00:03:17.920 --> 00:03:19.350 align:start position:0% +remove some of the topics along the way +and<00:03:18.159> this<00:03:18.400> is<00:03:18.480> the<00:03:18.720> first + +00:03:19.350 --> 00:03:19.360 align:start position:0% +and this is the first + + +00:03:19.360 --> 00:03:21.430 align:start position:0% +and this is the first +basically<00:03:19.840> draft<00:03:20.239> of<00:03:20.400> this<00:03:20.640> course<00:03:20.959> so<00:03:21.200> we'll + +00:03:21.430 --> 00:03:21.440 align:start position:0% +basically draft of this course so we'll + + +00:03:21.440 --> 00:03:22.390 align:start position:0% +basically draft of this course so we'll +see + +00:03:22.390 --> 00:03:22.400 align:start position:0% +see + + +00:03:22.400 --> 00:03:25.509 align:start position:0% +see +how<00:03:22.720> things<00:03:23.120> turn<00:03:23.440> out + +00:03:25.509 --> 00:03:25.519 align:start position:0% +how things turn out + + +00:03:25.519 --> 00:03:27.750 align:start position:0% +how things turn out +in<00:03:25.599> terms<00:03:25.920> of<00:03:26.000> the<00:03:26.239> questions<00:03:26.879> i<00:03:27.040> don't<00:03:27.280> want + +00:03:27.750 --> 00:03:27.760 align:start position:0% +in terms of the questions i don't want + + +00:03:27.760 --> 00:03:30.070 align:start position:0% +in terms of the questions i don't want +this<00:03:28.239> course<00:03:28.560> to<00:03:28.799> be<00:03:29.040> a<00:03:29.280> monologue + +00:03:30.070 --> 00:03:30.080 align:start position:0% +this course to be a monologue + + +00:03:30.080 --> 00:03:32.630 align:start position:0% +this course to be a monologue +even<00:03:30.319> though<00:03:30.560> we<00:03:30.640> are<00:03:30.799> in<00:03:30.959> a<00:03:31.040> virtual<00:03:31.760> setup<00:03:32.319> uh + +00:03:32.630 --> 00:03:32.640 align:start position:0% +even though we are in a virtual setup uh + + +00:03:32.640 --> 00:03:34.309 align:start position:0% +even though we are in a virtual setup uh +probably<00:03:32.959> most<00:03:33.200> of<00:03:33.360> you<00:03:33.440> have<00:03:33.599> taken + +00:03:34.309 --> 00:03:34.319 align:start position:0% +probably most of you have taken + + +00:03:34.319 --> 00:03:37.190 align:start position:0% +probably most of you have taken +some<00:03:34.560> other<00:03:34.959> courses<00:03:35.360> with<00:03:35.519> me<00:03:35.680> i<00:03:35.840> like<00:03:36.080> to<00:03:36.799> get + +00:03:37.190 --> 00:03:37.200 align:start position:0% +some other courses with me i like to get + + +00:03:37.200 --> 00:03:38.390 align:start position:0% +some other courses with me i like to get +questions<00:03:37.920> and + +00:03:38.390 --> 00:03:38.400 align:start position:0% +questions and + + +00:03:38.400 --> 00:03:40.710 align:start position:0% +questions and +you<00:03:38.480> know<00:03:38.720> have<00:03:38.959> some<00:03:39.280> dialogue<00:03:40.239> within<00:03:40.560> the + +00:03:40.710 --> 00:03:40.720 align:start position:0% +you know have some dialogue within the + + +00:03:40.720 --> 00:03:41.509 align:start position:0% +you know have some dialogue within the +lectures + +00:03:41.509 --> 00:03:41.519 align:start position:0% +lectures + + +00:03:41.519 --> 00:03:44.229 align:start position:0% +lectures +so<00:03:41.840> you<00:03:42.080> have<00:03:42.239> two<00:03:42.480> options<00:03:43.040> one<00:03:43.360> is<00:03:43.599> to<00:03:44.000> use + +00:03:44.229 --> 00:03:44.239 align:start position:0% +so you have two options one is to use + + +00:03:44.239 --> 00:03:45.589 align:start position:0% +so you have two options one is to use +the<00:03:44.480> raise + +00:03:45.589 --> 00:03:45.599 align:start position:0% +the raise + + +00:03:45.599 --> 00:03:48.949 align:start position:0% +the raise +hand<00:03:45.920> bottom<00:03:46.480> in<00:03:46.720> the + +00:03:48.949 --> 00:03:48.959 align:start position:0% + + + +00:03:48.959 --> 00:03:51.110 align:start position:0% + +in<00:03:49.120> the<00:03:49.360> zoom<00:03:49.680> or<00:03:49.920> you<00:03:50.000> can<00:03:50.239> just<00:03:50.560> post<00:03:50.879> your + +00:03:51.110 --> 00:03:51.120 align:start position:0% +in the zoom or you can just post your + + +00:03:51.120 --> 00:03:51.990 align:start position:0% +in the zoom or you can just post your +questions + +00:03:51.990 --> 00:03:52.000 align:start position:0% +questions + + +00:03:52.000 --> 00:03:53.910 align:start position:0% +questions +in<00:03:52.080> the<00:03:52.239> chat<00:03:52.560> box<00:03:52.959> and<00:03:53.120> i'll<00:03:53.439> read<00:03:53.680> your + +00:03:53.910 --> 00:03:53.920 align:start position:0% +in the chat box and i'll read your + + +00:03:53.920 --> 00:03:55.589 align:start position:0% +in the chat box and i'll read your +questions<00:03:54.319> to<00:03:54.400> the<00:03:54.560> class + +00:03:55.589 --> 00:03:55.599 align:start position:0% +questions to the class + + +00:03:55.599 --> 00:03:58.789 align:start position:0% +questions to the class +and<00:03:56.480> try<00:03:56.720> to<00:03:56.879> answer<00:03:57.120> them + +00:03:58.789 --> 00:03:58.799 align:start position:0% +and try to answer them + + +00:03:58.799 --> 00:04:03.910 align:start position:0% +and try to answer them +any<00:03:59.280> questions<00:03:59.599> so<00:03:59.760> far + +00:04:03.910 --> 00:04:03.920 align:start position:0% + + + +00:04:03.920 --> 00:04:08.390 align:start position:0% + +okay<00:04:04.879> good<00:04:05.120> so<00:04:05.439> let's<00:04:05.680> look<00:04:05.920> at<00:04:06.000> course<00:04:06.400> info + +00:04:08.390 --> 00:04:08.400 align:start position:0% +okay good so let's look at course info + + +00:04:08.400 --> 00:04:10.869 align:start position:0% +okay good so let's look at course info +as<00:04:08.640> i<00:04:08.720> mentioned<00:04:09.040> i'm<00:04:09.200> the<00:04:09.280> course<00:04:09.599> instructor + +00:04:10.869 --> 00:04:10.879 align:start position:0% +as i mentioned i'm the course instructor + + +00:04:10.879 --> 00:04:13.270 align:start position:0% +as i mentioned i'm the course instructor +and<00:04:11.040> we'll<00:04:11.280> have<00:04:11.680> three<00:04:12.080> amazing<00:04:12.640> tas<00:04:13.120> for + +00:04:13.270 --> 00:04:13.280 align:start position:0% +and we'll have three amazing tas for + + +00:04:13.280 --> 00:04:14.070 align:start position:0% +and we'll have three amazing tas for +this<00:04:13.519> course + +00:04:14.070 --> 00:04:14.080 align:start position:0% +this course + + +00:04:14.080 --> 00:04:17.349 align:start position:0% +this course +aya<00:04:14.640> neha<00:04:15.120> and<00:04:15.280> mukhang<00:04:16.880> they'll + +00:04:17.349 --> 00:04:17.359 align:start position:0% +aya neha and mukhang they'll + + +00:04:17.359 --> 00:04:20.870 align:start position:0% +aya neha and mukhang they'll +help<00:04:17.680> me<00:04:18.000> and<00:04:19.280> you<00:04:19.440> know<00:04:19.680> you<00:04:20.000> in + +00:04:20.870 --> 00:04:20.880 align:start position:0% +help me and you know you in + + +00:04:20.880 --> 00:04:23.189 align:start position:0% +help me and you know you in +different<00:04:21.519> parts<00:04:21.840> of<00:04:21.919> this<00:04:22.160> course<00:04:22.960> the + +00:04:23.189 --> 00:04:23.199 align:start position:0% +different parts of this course the + + +00:04:23.199 --> 00:04:24.230 align:start position:0% +different parts of this course the +topics + +00:04:24.230 --> 00:04:24.240 align:start position:0% +topics + + +00:04:24.240 --> 00:04:26.230 align:start position:0% +topics +you<00:04:24.320> know<00:04:24.479> it's<00:04:24.639> basically<00:04:25.360> the<00:04:25.520> same<00:04:25.759> topics + +00:04:26.230 --> 00:04:26.240 align:start position:0% +you know it's basically the same topics + + +00:04:26.240 --> 00:04:28.390 align:start position:0% +you know it's basically the same topics +that<00:04:26.400> we<00:04:26.639> saw<00:04:26.960> earlier + +00:04:28.390 --> 00:04:28.400 align:start position:0% +that we saw earlier + + +00:04:28.400 --> 00:04:31.110 align:start position:0% +that we saw earlier +lectures<00:04:28.880> will<00:04:29.040> be<00:04:29.440> online<00:04:29.919> so<00:04:30.080> i'm<00:04:30.479> recording + +00:04:31.110 --> 00:04:31.120 align:start position:0% +lectures will be online so i'm recording + + +00:04:31.120 --> 00:04:32.469 align:start position:0% +lectures will be online so i'm recording +these<00:04:31.360> lectures + +00:04:32.469 --> 00:04:32.479 align:start position:0% +these lectures + + +00:04:32.479 --> 00:04:36.150 align:start position:0% +these lectures +and<00:04:33.199> after<00:04:33.759> each<00:04:34.880> class<00:04:35.199> i'm<00:04:35.440> going<00:04:35.680> to + +00:04:36.150 --> 00:04:36.160 align:start position:0% +and after each class i'm going to + + +00:04:36.160 --> 00:04:38.550 align:start position:0% +and after each class i'm going to +post<00:04:36.479> these<00:04:36.639> lectures<00:04:37.360> but<00:04:37.600> i<00:04:38.000> strongly + +00:04:38.550 --> 00:04:38.560 align:start position:0% +post these lectures but i strongly + + +00:04:38.560 --> 00:04:40.830 align:start position:0% +post these lectures but i strongly +encourage<00:04:39.120> all<00:04:39.360> of<00:04:39.520> you<00:04:39.759> to + +00:04:40.830 --> 00:04:40.840 align:start position:0% +encourage all of you to + + +00:04:40.840 --> 00:04:43.670 align:start position:0% +encourage all of you to +participate<00:04:42.080> in<00:04:42.320> these<00:04:42.560> virtual<00:04:43.120> lectures + +00:04:43.670 --> 00:04:43.680 align:start position:0% +participate in these virtual lectures + + +00:04:43.680 --> 00:04:45.590 align:start position:0% +participate in these virtual lectures +because<00:04:43.919> as<00:04:44.080> i<00:04:44.240> mentioned<00:04:44.639> i<00:04:44.800> like<00:04:44.960> to + +00:04:45.590 --> 00:04:45.600 align:start position:0% +because as i mentioned i like to + + +00:04:45.600 --> 00:04:48.629 align:start position:0% +because as i mentioned i like to +have<00:04:45.919> a<00:04:46.000> lot<00:04:46.240> of<00:04:46.479> q<00:04:46.720> and<00:04:46.960> a<00:04:47.280> type<00:04:47.520> conversations + +00:04:48.629 --> 00:04:48.639 align:start position:0% +have a lot of q and a type conversations + + +00:04:48.639 --> 00:04:52.390 align:start position:0% +have a lot of q and a type conversations +during<00:04:48.960> the<00:04:49.120> lectures<00:04:51.120> so<00:04:51.360> office<00:04:51.759> hours + +00:04:52.390 --> 00:04:52.400 align:start position:0% +during the lectures so office hours + + +00:04:52.400 --> 00:04:54.629 align:start position:0% +during the lectures so office hours +my<00:04:52.639> office<00:04:52.960> hour<00:04:53.199> will<00:04:53.360> be<00:04:53.919> right<00:04:54.160> after + +00:04:54.629 --> 00:04:54.639 align:start position:0% +my office hour will be right after + + +00:04:54.639 --> 00:04:56.710 align:start position:0% +my office hour will be right after +tuesday's<00:04:55.280> lectures + +00:04:56.710 --> 00:04:56.720 align:start position:0% +tuesday's lectures + + +00:04:56.720 --> 00:05:00.550 align:start position:0% +tuesday's lectures +starting<00:04:57.120> next<00:04:57.440> week<00:04:58.320> and<00:04:58.880> ta<00:04:59.280> office<00:04:59.600> hours + +00:05:00.550 --> 00:05:00.560 align:start position:0% +starting next week and ta office hours + + +00:05:00.560 --> 00:05:04.390 align:start position:0% +starting next week and ta office hours +will<00:05:00.800> announce<00:05:01.199> it<00:05:02.160> shortly + +00:05:04.390 --> 00:05:04.400 align:start position:0% +will announce it shortly + + +00:05:04.400 --> 00:05:06.310 align:start position:0% +will announce it shortly +in<00:05:04.479> terms<00:05:04.800> of<00:05:04.880> the<00:05:05.199> background<00:05:06.000> as<00:05:06.240> i + +00:05:06.310 --> 00:05:06.320 align:start position:0% +in terms of the background as i + + +00:05:06.320 --> 00:05:07.749 align:start position:0% +in terms of the background as i +mentioned<00:05:06.639> this<00:05:06.800> is<00:05:06.960> an<00:05:07.199> advanced + +00:05:07.749 --> 00:05:07.759 align:start position:0% +mentioned this is an advanced + + +00:05:07.759 --> 00:05:10.870 align:start position:0% +mentioned this is an advanced +course<00:05:08.400> so<00:05:08.639> you<00:05:08.880> need<00:05:09.120> to<00:05:09.520> have<00:05:10.240> a<00:05:10.400> strong + +00:05:10.870 --> 00:05:10.880 align:start position:0% +course so you need to have a strong + + +00:05:10.880 --> 00:05:12.870 align:start position:0% +course so you need to have a strong +background<00:05:11.520> in<00:05:11.680> linear<00:05:12.080> algebra + +00:05:12.870 --> 00:05:12.880 align:start position:0% +background in linear algebra + + +00:05:12.880 --> 00:05:15.909 align:start position:0% +background in linear algebra +probability<00:05:13.600> and<00:05:13.759> optimization<00:05:15.199> and<00:05:15.520> also + +00:05:15.909 --> 00:05:15.919 align:start position:0% +probability and optimization and also + + +00:05:15.919 --> 00:05:17.909 align:start position:0% +probability and optimization and also +you<00:05:16.080> need<00:05:16.240> to<00:05:16.400> have<00:05:16.560> some<00:05:16.880> basic<00:05:17.360> background + +00:05:17.909 --> 00:05:17.919 align:start position:0% +you need to have some basic background + + +00:05:17.919 --> 00:05:19.350 align:start position:0% +you need to have some basic background +in<00:05:18.080> machine<00:05:18.479> learning + +00:05:19.350 --> 00:05:19.360 align:start position:0% +in machine learning + + +00:05:19.360 --> 00:05:22.830 align:start position:0% +in machine learning +and<00:05:19.759> deep<00:05:20.080> learning<00:05:21.520> otherwise<00:05:22.080> it<00:05:22.240> will<00:05:22.400> be + +00:05:22.830 --> 00:05:22.840 align:start position:0% +and deep learning otherwise it will be + + +00:05:22.840 --> 00:05:26.070 align:start position:0% +and deep learning otherwise it will be +hard<00:05:24.400> for<00:05:24.639> you<00:05:24.800> to<00:05:25.039> follow<00:05:25.440> some<00:05:25.600> of<00:05:25.759> the + +00:05:26.070 --> 00:05:26.080 align:start position:0% +hard for you to follow some of the + + +00:05:26.080 --> 00:05:28.310 align:start position:0% +hard for you to follow some of the +lectures<00:05:26.479> that<00:05:26.639> you<00:05:26.880> have + +00:05:28.310 --> 00:05:28.320 align:start position:0% +lectures that you have + + +00:05:28.320 --> 00:05:30.790 align:start position:0% +lectures that you have +and<00:05:29.039> obviously<00:05:29.520> you<00:05:29.680> need<00:05:29.840> to<00:05:30.000> have<00:05:30.320> some + +00:05:30.790 --> 00:05:30.800 align:start position:0% +and obviously you need to have some + + +00:05:30.800 --> 00:05:33.990 align:start position:0% +and obviously you need to have some +programming<00:05:31.360> backgrounds<00:05:31.919> in<00:05:32.000> python + +00:05:33.990 --> 00:05:34.000 align:start position:0% +programming backgrounds in python + + +00:05:34.000 --> 00:05:37.189 align:start position:0% +programming backgrounds in python +so<00:05:34.240> the<00:05:34.400> course<00:05:34.720> is<00:05:35.199> going<00:05:35.440> to<00:05:35.600> be<00:05:36.560> heavily + +00:05:37.189 --> 00:05:37.199 align:start position:0% +so the course is going to be heavily + + +00:05:37.199 --> 00:05:40.310 align:start position:0% +so the course is going to be heavily +project<00:05:37.759> oriented<00:05:39.039> so<00:05:39.199> we<00:05:39.440> are<00:05:39.520> going<00:05:39.840> to + +00:05:40.310 --> 00:05:40.320 align:start position:0% +project oriented so we are going to + + +00:05:40.320 --> 00:05:43.909 align:start position:0% +project oriented so we are going to +have<00:05:40.720> a<00:05:41.120> mini<00:05:41.600> conference<00:05:42.560> in<00:05:42.639> this<00:05:42.960> course + +00:05:43.909 --> 00:05:43.919 align:start position:0% +have a mini conference in this course + + +00:05:43.919 --> 00:05:46.950 align:start position:0% +have a mini conference in this course +so<00:05:44.160> how<00:05:45.199> does<00:05:45.440> it<00:05:45.600> work<00:05:46.000> so + +00:05:46.950 --> 00:05:46.960 align:start position:0% +so how does it work so + + +00:05:46.960 --> 00:05:48.950 align:start position:0% +so how does it work so +students<00:05:47.280> they<00:05:47.440> are<00:05:47.520> going<00:05:47.680> to<00:05:47.840> form<00:05:48.160> groups + +00:05:48.950 --> 00:05:48.960 align:start position:0% +students they are going to form groups + + +00:05:48.960 --> 00:05:50.710 align:start position:0% +students they are going to form groups +and<00:05:49.600> similar<00:05:50.080> to<00:05:50.240> some + +00:05:50.710 --> 00:05:50.720 align:start position:0% +and similar to some + + +00:05:50.720 --> 00:05:52.550 align:start position:0% +and similar to some +other<00:05:51.039> machine<00:05:51.360> learning<00:05:51.759> conferences<00:05:52.320> will + +00:05:52.550 --> 00:05:52.560 align:start position:0% +other machine learning conferences will + + +00:05:52.560 --> 00:05:53.909 align:start position:0% +other machine learning conferences will +have<00:05:52.880> a<00:05:53.039> double + +00:05:53.909 --> 00:05:53.919 align:start position:0% +have a double + + +00:05:53.919 --> 00:05:59.110 align:start position:0% +have a double +blind<00:05:54.960> uh<00:05:55.759> submission<00:05:56.479> and<00:05:56.639> review<00:05:57.120> process + +00:05:59.110 --> 00:05:59.120 align:start position:0% +blind uh submission and review process + + +00:05:59.120 --> 00:06:04.070 align:start position:0% +blind uh submission and review process +so<00:05:59.919> basically<00:06:02.000> each<00:06:02.400> student + +00:06:04.070 --> 00:06:04.080 align:start position:0% +so basically each student + + +00:06:04.080 --> 00:06:06.710 align:start position:0% +so basically each student +also<00:06:04.479> will<00:06:05.120> act<00:06:05.440> as<00:06:05.600> a<00:06:05.759> reviewer<00:06:06.319> so<00:06:06.479> you<00:06:06.639> will + +00:06:06.710 --> 00:06:06.720 align:start position:0% +also will act as a reviewer so you will + + +00:06:06.720 --> 00:06:08.469 align:start position:0% +also will act as a reviewer so you will +review<00:06:07.120> three<00:06:07.440> submissions<00:06:08.240> in<00:06:08.400> a + +00:06:08.469 --> 00:06:08.479 align:start position:0% +review three submissions in a + + +00:06:08.479 --> 00:06:10.309 align:start position:0% +review three submissions in a +double-blind<00:06:09.280> format + +00:06:10.309 --> 00:06:10.319 align:start position:0% +double-blind format + + +00:06:10.319 --> 00:06:12.150 align:start position:0% +double-blind format +and<00:06:10.639> everything<00:06:11.199> you<00:06:11.280> know<00:06:11.440> should<00:06:11.680> be<00:06:11.919> in<00:06:12.080> a + +00:06:12.150 --> 00:06:12.160 align:start position:0% +and everything you know should be in a + + +00:06:12.160 --> 00:06:13.749 align:start position:0% +and everything you know should be in a +double-blind<00:06:12.800> format + +00:06:13.749 --> 00:06:13.759 align:start position:0% +double-blind format + + +00:06:13.759 --> 00:06:15.670 align:start position:0% +double-blind format +otherwise<00:06:14.240> there<00:06:14.400> will<00:06:14.560> be<00:06:14.720> a<00:06:14.800> heavy<00:06:15.120> penalty + +00:06:15.670 --> 00:06:15.680 align:start position:0% +otherwise there will be a heavy penalty + + +00:06:15.680 --> 00:06:16.870 align:start position:0% +otherwise there will be a heavy penalty +on<00:06:15.759> that + +00:06:16.870 --> 00:06:16.880 align:start position:0% +on that + + +00:06:16.880 --> 00:06:19.510 align:start position:0% +on that +and<00:06:17.199> each<00:06:17.440> submission<00:06:18.240> can<00:06:18.560> have<00:06:18.960> up<00:06:19.120> to<00:06:19.280> four + +00:06:19.510 --> 00:06:19.520 align:start position:0% +and each submission can have up to four + + +00:06:19.520 --> 00:06:21.029 align:start position:0% +and each submission can have up to four +students<00:06:19.919> so<00:06:20.080> we<00:06:20.319> encourage + +00:06:21.029 --> 00:06:21.039 align:start position:0% +students so we encourage + + +00:06:21.039 --> 00:06:24.230 align:start position:0% +students so we encourage +to<00:06:21.840> uh<00:06:22.160> for<00:06:22.319> you<00:06:22.560> to<00:06:22.880> form<00:06:23.280> teams + +00:06:24.230 --> 00:06:24.240 align:start position:0% +to uh for you to form teams + + +00:06:24.240 --> 00:06:27.270 align:start position:0% +to uh for you to form teams +uh<00:06:25.039> and<00:06:25.520> collaborate<00:06:26.080> with<00:06:26.319> other<00:06:26.639> students + +00:06:27.270 --> 00:06:27.280 align:start position:0% +uh and collaborate with other students + + +00:06:27.280 --> 00:06:29.350 align:start position:0% +uh and collaborate with other students +i<00:06:27.520> understand<00:06:28.160> it<00:06:28.400> is<00:06:28.720> a<00:06:28.880> little<00:06:29.120> bit + +00:06:29.350 --> 00:06:29.360 align:start position:0% +i understand it is a little bit + + +00:06:29.360 --> 00:06:31.350 align:start position:0% +i understand it is a little bit +difficult<00:06:30.080> especially<00:06:30.639> since<00:06:30.880> we<00:06:30.960> are<00:06:31.120> in<00:06:31.280> a + +00:06:31.350 --> 00:06:31.360 align:start position:0% +difficult especially since we are in a + + +00:06:31.360 --> 00:06:32.230 align:start position:0% +difficult especially since we are in a +virtual + +00:06:32.230 --> 00:06:32.240 align:start position:0% +virtual + + +00:06:32.240 --> 00:06:35.590 align:start position:0% +virtual +situation<00:06:33.360> but<00:06:33.600> uh<00:06:34.080> we<00:06:34.319> have<00:06:34.639> piazza + +00:06:35.590 --> 00:06:35.600 align:start position:0% +situation but uh we have piazza + + +00:06:35.600 --> 00:06:39.029 align:start position:0% +situation but uh we have piazza +that<00:06:36.319> you<00:06:36.560> can<00:06:37.039> post<00:06:37.440> your<00:06:38.240> uh + +00:06:39.029 --> 00:06:39.039 align:start position:0% +that you can post your uh + + +00:06:39.039 --> 00:06:41.430 align:start position:0% +that you can post your uh +problems<00:06:39.520> or<00:06:39.759> projects<00:06:40.479> and<00:06:40.639> then<00:06:41.039> see<00:06:41.199> if + +00:06:41.430 --> 00:06:41.440 align:start position:0% +problems or projects and then see if + + +00:06:41.440 --> 00:06:43.350 align:start position:0% +problems or projects and then see if +other<00:06:41.759> students<00:06:42.160> are<00:06:42.319> interested<00:06:42.800> in<00:06:42.960> joining + +00:06:43.350 --> 00:06:43.360 align:start position:0% +other students are interested in joining + + +00:06:43.360 --> 00:06:44.629 align:start position:0% +other students are interested in joining +your + +00:06:44.629 --> 00:06:44.639 align:start position:0% +your + + +00:06:44.639 --> 00:06:47.749 align:start position:0% +your +your<00:06:44.800> team<00:06:46.319> so<00:06:46.560> one<00:06:46.720> thing<00:06:46.880> to<00:06:47.039> emphasize<00:06:47.520> is + +00:06:47.749 --> 00:06:47.759 align:start position:0% +your team so one thing to emphasize is + + +00:06:47.759 --> 00:06:49.350 align:start position:0% +your team so one thing to emphasize is +all<00:06:47.919> deadlines<00:06:48.479> are<00:06:48.639> final + +00:06:49.350 --> 00:06:49.360 align:start position:0% +all deadlines are final + + +00:06:49.360 --> 00:06:51.110 align:start position:0% +all deadlines are final +and<00:06:49.520> we<00:06:49.759> won't<00:06:50.080> extend<00:06:50.479> it<00:06:50.639> because<00:06:50.960> the + +00:06:51.110 --> 00:06:51.120 align:start position:0% +and we won't extend it because the + + +00:06:51.120 --> 00:06:53.110 align:start position:0% +and we won't extend it because the +deadlines<00:06:51.599> are<00:06:51.759> cascaded<00:06:52.319> so<00:06:52.479> if<00:06:52.639> you<00:06:52.720> extend + +00:06:53.110 --> 00:06:53.120 align:start position:0% +deadlines are cascaded so if you extend + + +00:06:53.120 --> 00:06:53.909 align:start position:0% +deadlines are cascaded so if you extend +one<00:06:53.280> of<00:06:53.440> them + +00:06:53.909 --> 00:06:53.919 align:start position:0% +one of them + + +00:06:53.919 --> 00:06:55.990 align:start position:0% +one of them +it's<00:06:54.080> going<00:06:54.240> to<00:06:54.319> be<00:06:54.479> hard<00:06:54.800> to<00:06:55.039> adjust<00:06:55.599> every + +00:06:55.990 --> 00:06:56.000 align:start position:0% +it's going to be hard to adjust every + + +00:06:56.000 --> 00:06:57.189 align:start position:0% +it's going to be hard to adjust every +other<00:06:56.240> deadline + +00:06:57.189 --> 00:06:57.199 align:start position:0% +other deadline + + +00:06:57.199 --> 00:06:59.830 align:start position:0% +other deadline +so<00:06:57.360> here<00:06:57.599> are<00:06:57.680> the<00:06:57.919> important<00:06:58.400> dates<00:06:58.639> for<00:06:58.880> us + +00:06:59.830 --> 00:06:59.840 align:start position:0% +so here are the important dates for us + + +00:06:59.840 --> 00:07:00.230 align:start position:0% +so here are the important dates for us +the + +00:07:00.230 --> 00:07:00.240 align:start position:0% +the + + +00:07:00.240 --> 00:07:02.309 align:start position:0% +the +beginning<00:07:00.720> of<00:07:00.880> october<00:07:01.520> you<00:07:01.680> need<00:07:01.840> to<00:07:02.000> form + +00:07:02.309 --> 00:07:02.319 align:start position:0% +beginning of october you need to form + + +00:07:02.319 --> 00:07:03.430 align:start position:0% +beginning of october you need to form +your<00:07:02.560> team + +00:07:03.430 --> 00:07:03.440 align:start position:0% +your team + + +00:07:03.440 --> 00:07:05.909 align:start position:0% +your team +and<00:07:03.680> you<00:07:03.919> need<00:07:04.080> to<00:07:04.639> select<00:07:05.039> the<00:07:05.199> topic<00:07:05.759> the + +00:07:05.909 --> 00:07:05.919 align:start position:0% +and you need to select the topic the + + +00:07:05.919 --> 00:07:07.749 align:start position:0% +and you need to select the topic the +project + +00:07:07.749 --> 00:07:07.759 align:start position:0% +project + + +00:07:07.759 --> 00:07:09.350 align:start position:0% +project +in<00:07:07.919> a<00:07:08.000> week<00:07:08.240> we<00:07:08.400> are<00:07:08.560> going<00:07:08.720> to<00:07:08.880> give<00:07:09.039> you<00:07:09.199> a + +00:07:09.350 --> 00:07:09.360 align:start position:0% +in a week we are going to give you a + + +00:07:09.360 --> 00:07:11.830 align:start position:0% +in a week we are going to give you a +feedback<00:07:10.160> ups<00:07:10.479> or<00:07:10.639> dance + +00:07:11.830 --> 00:07:11.840 align:start position:0% +feedback ups or dance + + +00:07:11.840 --> 00:07:15.110 align:start position:0% +feedback ups or dance +and<00:07:12.800> the<00:07:12.960> first<00:07:13.280> week<00:07:13.520> of<00:07:13.840> november<00:07:14.800> there's + +00:07:15.110 --> 00:07:15.120 align:start position:0% +and the first week of november there's + + +00:07:15.120 --> 00:07:16.710 align:start position:0% +and the first week of november there's +an<00:07:15.280> abstract<00:07:15.599> submission<00:07:16.160> deadline<00:07:16.560> so + +00:07:16.710 --> 00:07:16.720 align:start position:0% +an abstract submission deadline so + + +00:07:16.720 --> 00:07:17.110 align:start position:0% +an abstract submission deadline so +you'll + +00:07:17.110 --> 00:07:17.120 align:start position:0% +you'll + + +00:07:17.120 --> 00:07:19.749 align:start position:0% +you'll +post<00:07:17.360> your<00:07:17.599> abstracts<00:07:18.560> and<00:07:18.880> within<00:07:19.199> a<00:07:19.280> week<00:07:19.520> of + +00:07:19.749 --> 00:07:19.759 align:start position:0% +post your abstracts and within a week of + + +00:07:19.759 --> 00:07:20.150 align:start position:0% +post your abstracts and within a week of +that + +00:07:20.150 --> 00:07:20.160 align:start position:0% +that + + +00:07:20.160 --> 00:07:23.110 align:start position:0% +that +you<00:07:20.319> will<00:07:20.560> have<00:07:20.880> your<00:07:21.280> full<00:07:21.520> paper<00:07:22.240> in<00:07:22.560> icml + +00:07:23.110 --> 00:07:23.120 align:start position:0% +you will have your full paper in icml + + +00:07:23.120 --> 00:07:24.390 align:start position:0% +you will have your full paper in icml +format + +00:07:24.390 --> 00:07:24.400 align:start position:0% +format + + +00:07:24.400 --> 00:07:27.189 align:start position:0% +format +and<00:07:24.560> then<00:07:25.120> the<00:07:25.280> review<00:07:25.919> period<00:07:26.479> will<00:07:26.800> start + +00:07:27.189 --> 00:07:27.199 align:start position:0% +and then the review period will start + + +00:07:27.199 --> 00:07:28.550 align:start position:0% +and then the review period will start +right<00:07:27.440> after<00:07:27.759> that + +00:07:28.550 --> 00:07:28.560 align:start position:0% +right after that + + +00:07:28.560 --> 00:07:30.629 align:start position:0% +right after that +you<00:07:28.720> will<00:07:28.800> be<00:07:28.960> assigned<00:07:29.360> to<00:07:29.520> three<00:07:29.840> papers<00:07:30.479> in + +00:07:30.629 --> 00:07:30.639 align:start position:0% +you will be assigned to three papers in + + +00:07:30.639 --> 00:07:31.589 align:start position:0% +you will be assigned to three papers in +groups + +00:07:31.589 --> 00:07:31.599 align:start position:0% +groups + + +00:07:31.599 --> 00:07:34.870 align:start position:0% +groups +and<00:07:32.080> you<00:07:32.319> review<00:07:32.800> these<00:07:33.440> papers + +00:07:34.870 --> 00:07:34.880 align:start position:0% +and you review these papers + + +00:07:34.880 --> 00:07:37.029 align:start position:0% +and you review these papers +and<00:07:35.520> after<00:07:35.919> that<00:07:36.080> we'll<00:07:36.240> have<00:07:36.400> a<00:07:36.479> rebuttal + +00:07:37.029 --> 00:07:37.039 align:start position:0% +and after that we'll have a rebuttal + + +00:07:37.039 --> 00:07:38.070 align:start position:0% +and after that we'll have a rebuttal +period<00:07:37.360> so<00:07:37.599> we'll + +00:07:38.070 --> 00:07:38.080 align:start position:0% +period so we'll + + +00:07:38.080 --> 00:07:41.110 align:start position:0% +period so we'll +reveal<00:07:38.479> these<00:07:38.800> uh<00:07:39.280> reviews<00:07:39.840> to<00:07:40.479> authors<00:07:40.960> of + +00:07:41.110 --> 00:07:41.120 align:start position:0% +reveal these uh reviews to authors of + + +00:07:41.120 --> 00:07:41.430 align:start position:0% +reveal these uh reviews to authors of +the + +00:07:41.430 --> 00:07:41.440 align:start position:0% +the + + +00:07:41.440 --> 00:07:44.790 align:start position:0% +the +papers<00:07:42.400> and<00:07:43.039> have<00:07:43.280> them<00:07:44.000> basically + +00:07:44.790 --> 00:07:44.800 align:start position:0% +papers and have them basically + + +00:07:44.800 --> 00:07:47.189 align:start position:0% +papers and have them basically +write<00:07:45.680> responses<00:07:46.240> to<00:07:46.400> the<00:07:46.560> comments<00:07:46.960> raised + +00:07:47.189 --> 00:07:47.199 align:start position:0% +write responses to the comments raised + + +00:07:47.199 --> 00:07:48.710 align:start position:0% +write responses to the comments raised +by<00:07:47.360> the<00:07:47.440> reviewers + +00:07:48.710 --> 00:07:48.720 align:start position:0% +by the reviewers + + +00:07:48.720 --> 00:07:50.469 align:start position:0% +by the reviewers +and<00:07:48.960> after<00:07:49.280> that<00:07:49.520> we'll<00:07:49.680> have<00:07:49.919> a<00:07:50.080> couple<00:07:50.319> of + +00:07:50.469 --> 00:07:50.479 align:start position:0% +and after that we'll have a couple of + + +00:07:50.479 --> 00:07:53.430 align:start position:0% +and after that we'll have a couple of +days<00:07:50.720> for<00:07:51.360> discussions<00:07:52.000> among<00:07:52.240> reviewers + +00:07:53.430 --> 00:07:53.440 align:start position:0% +days for discussions among reviewers + + +00:07:53.440 --> 00:07:57.430 align:start position:0% +days for discussions among reviewers +and<00:07:54.400> after<00:07:54.800> that<00:07:55.360> acs<00:07:56.319> which<00:07:56.479> will<00:07:56.639> be<00:07:56.879> my<00:07:57.039> tas + +00:07:57.430 --> 00:07:57.440 align:start position:0% +and after that acs which will be my tas + + +00:07:57.440 --> 00:07:58.710 align:start position:0% +and after that acs which will be my tas +and<00:07:57.599> myself<00:07:58.080> will + +00:07:58.710 --> 00:07:58.720 align:start position:0% +and myself will + + +00:07:58.720 --> 00:08:01.830 align:start position:0% +and myself will +finalize<00:07:59.680> the<00:08:00.080> scores<00:08:00.479> for<00:08:00.720> each<00:08:00.960> paper + +00:08:01.830 --> 00:08:01.840 align:start position:0% +finalize the scores for each paper + + +00:08:01.840 --> 00:08:06.150 align:start position:0% +finalize the scores for each paper +and<00:08:02.400> also<00:08:02.960> evaluate<00:08:03.520> reviewers + +00:08:06.150 --> 00:08:06.160 align:start position:0% + + + +00:08:06.160 --> 00:08:08.950 align:start position:0% + +so<00:08:06.319> we'll<00:08:06.560> have<00:08:06.720> another<00:08:07.680> part<00:08:08.560> lecture + +00:08:08.950 --> 00:08:08.960 align:start position:0% +so we'll have another part lecture + + +00:08:08.960 --> 00:08:10.790 align:start position:0% +so we'll have another part lecture +describing + +00:08:10.790 --> 00:08:10.800 align:start position:0% +describing + + +00:08:10.800 --> 00:08:13.350 align:start position:0% +describing +we<00:08:11.039> already<00:08:11.520> shared<00:08:11.759> the<00:08:12.000> google<00:08:12.800> sheet<00:08:13.120> for + +00:08:13.350 --> 00:08:13.360 align:start position:0% +we already shared the google sheet for + + +00:08:13.360 --> 00:08:15.589 align:start position:0% +we already shared the google sheet for +you<00:08:13.520> to<00:08:13.680> sign<00:08:13.919> up<00:08:14.080> to<00:08:14.240> scribe<00:08:14.560> lectures + +00:08:15.589 --> 00:08:15.599 align:start position:0% +you to sign up to scribe lectures + + +00:08:15.599 --> 00:08:18.629 align:start position:0% +you to sign up to scribe lectures +uh<00:08:16.160> i'm<00:08:16.319> gonna<00:08:16.960> basically<00:08:17.599> share<00:08:17.840> the<00:08:18.080> notes + +00:08:18.629 --> 00:08:18.639 align:start position:0% +uh i'm gonna basically share the notes + + +00:08:18.639 --> 00:08:21.270 align:start position:0% +uh i'm gonna basically share the notes +uh<00:08:19.680> these<00:08:19.919> describe<00:08:20.400> notes<00:08:20.720> after<00:08:21.039> each + +00:08:21.270 --> 00:08:21.280 align:start position:0% +uh these describe notes after each + + +00:08:21.280 --> 00:08:21.990 align:start position:0% +uh these describe notes after each +lecture + +00:08:21.990 --> 00:08:22.000 align:start position:0% +lecture + + +00:08:22.000 --> 00:08:24.230 align:start position:0% +lecture +so<00:08:22.240> in<00:08:22.400> addition<00:08:22.800> to<00:08:22.960> having<00:08:23.360> the<00:08:23.599> videos<00:08:24.080> for + +00:08:24.230 --> 00:08:24.240 align:start position:0% +so in addition to having the videos for + + +00:08:24.240 --> 00:08:25.110 align:start position:0% +so in addition to having the videos for +the<00:08:24.400> lectures + +00:08:25.110 --> 00:08:25.120 align:start position:0% +the lectures + + +00:08:25.120 --> 00:08:27.830 align:start position:0% +the lectures +you<00:08:25.280> will<00:08:25.520> have<00:08:25.759> access<00:08:26.240> to<00:08:26.639> some<00:08:27.440> scribe + +00:08:27.830 --> 00:08:27.840 align:start position:0% +you will have access to some scribe + + +00:08:27.840 --> 00:08:30.309 align:start position:0% +you will have access to some scribe +notes<00:08:28.160> from<00:08:28.400> the<00:08:28.639> lectures + +00:08:30.309 --> 00:08:30.319 align:start position:0% +notes from the lectures + + +00:08:30.319 --> 00:08:32.630 align:start position:0% +notes from the lectures +in<00:08:30.479> terms<00:08:30.720> of<00:08:30.879> the<00:08:31.199> exams<00:08:31.680> we'll<00:08:31.919> have<00:08:32.159> only + +00:08:32.630 --> 00:08:32.640 align:start position:0% +in terms of the exams we'll have only + + +00:08:32.640 --> 00:08:34.230 align:start position:0% +in terms of the exams we'll have only +one<00:08:33.039> exam + +00:08:34.230 --> 00:08:34.240 align:start position:0% +one exam + + +00:08:34.240 --> 00:08:36.949 align:start position:0% +one exam +and<00:08:34.479> that<00:08:34.719> is<00:08:34.880> in<00:08:34.959> fact<00:08:35.279> required<00:08:36.080> because + +00:08:36.949 --> 00:08:36.959 align:start position:0% +and that is in fact required because + + +00:08:36.959 --> 00:08:38.949 align:start position:0% +and that is in fact required because +this<00:08:37.200> is<00:08:37.360> a<00:08:37.440> qualifying<00:08:38.080> course<00:08:38.479> you'll<00:08:38.719> have + +00:08:38.949 --> 00:08:38.959 align:start position:0% +this is a qualifying course you'll have + + +00:08:38.959 --> 00:08:40.310 align:start position:0% +this is a qualifying course you'll have +one<00:08:39.200> final<00:08:39.599> exam + +00:08:40.310 --> 00:08:40.320 align:start position:0% +one final exam + + +00:08:40.320 --> 00:08:42.790 align:start position:0% +one final exam +it<00:08:40.479> is<00:08:40.640> take<00:08:40.959> home<00:08:41.279> and<00:08:41.519> it<00:08:41.680> will<00:08:41.839> be<00:08:42.080> on<00:08:42.479> third + +00:08:42.790 --> 00:08:42.800 align:start position:0% +it is take home and it will be on third + + +00:08:42.800 --> 00:08:44.710 align:start position:0% +it is take home and it will be on third +of<00:08:43.039> december + +00:08:44.710 --> 00:08:44.720 align:start position:0% +of december + + +00:08:44.720 --> 00:08:46.550 align:start position:0% +of december +so<00:08:44.880> in<00:08:44.959> terms<00:08:45.200> of<00:08:45.279> the<00:08:45.440> grading<00:08:45.920> we'll<00:08:46.080> have<00:08:46.399> 10 + +00:08:46.550 --> 00:08:46.560 align:start position:0% +so in terms of the grading we'll have 10 + + +00:08:46.560 --> 00:08:48.630 align:start position:0% +so in terms of the grading we'll have 10 +percent<00:08:46.959> for<00:08:47.120> the<00:08:47.680> lecture<00:08:48.000> scribe + +00:08:48.630 --> 00:08:48.640 align:start position:0% +percent for the lecture scribe + + +00:08:48.640 --> 00:08:51.910 align:start position:0% +percent for the lecture scribe +we<00:08:48.800> have<00:08:49.040> 25<00:08:49.839> for<00:08:50.000> the<00:08:50.160> final<00:08:51.200> 40 + +00:08:51.910 --> 00:08:51.920 align:start position:0% +we have 25 for the final 40 + + +00:08:51.920 --> 00:08:54.230 align:start position:0% +we have 25 for the final 40 +for<00:08:52.080> the<00:08:52.240> paper<00:08:52.560> score<00:08:52.959> and<00:08:53.200> 25<00:08:53.920> for<00:08:54.080> the + +00:08:54.230 --> 00:08:54.240 align:start position:0% +for the paper score and 25 for the + + +00:08:54.240 --> 00:08:56.150 align:start position:0% +for the paper score and 25 for the +review<00:08:54.560> score<00:08:54.959> as<00:08:55.120> you<00:08:55.279> can<00:08:55.360> see<00:08:55.519> it<00:08:55.680> is + +00:08:56.150 --> 00:08:56.160 align:start position:0% +review score as you can see it is + + +00:08:56.160 --> 00:08:59.190 align:start position:0% +review score as you can see it is +heavily<00:08:57.120> project<00:08:57.519> oriented<00:08:58.000> 65<00:08:58.880> of<00:08:59.040> your + +00:08:59.190 --> 00:08:59.200 align:start position:0% +heavily project oriented 65 of your + + +00:08:59.200 --> 00:09:00.870 align:start position:0% +heavily project oriented 65 of your +grades<00:08:59.519> will<00:08:59.760> basically<00:09:00.240> come<00:09:00.480> from<00:09:00.640> the + +00:09:00.870 --> 00:09:00.880 align:start position:0% +grades will basically come from the + + +00:09:00.880 --> 00:09:02.550 align:start position:0% +grades will basically come from the +project<00:09:01.279> and<00:09:01.440> also<00:09:01.600> the<00:09:01.760> reviews<00:09:02.160> that<00:09:02.320> you'll + +00:09:02.550 --> 00:09:02.560 align:start position:0% +project and also the reviews that you'll + + +00:09:02.560 --> 00:09:02.870 align:start position:0% +project and also the reviews that you'll +be + +00:09:02.870 --> 00:09:02.880 align:start position:0% +be + + +00:09:02.880 --> 00:09:07.190 align:start position:0% +be +doing<00:09:03.600> for<00:09:04.000> uh<00:09:04.560> for<00:09:04.959> other<00:09:05.360> uh<00:09:05.680> papers + +00:09:07.190 --> 00:09:07.200 align:start position:0% +doing for uh for other uh papers + + +00:09:07.200 --> 00:09:10.310 align:start position:0% +doing for uh for other uh papers +all<00:09:07.360> right<00:09:07.519> so<00:09:08.240> that<00:09:08.560> is<00:09:08.720> it<00:09:09.120> for<00:09:09.600> the<00:09:09.920> course + +00:09:10.310 --> 00:09:10.320 align:start position:0% +all right so that is it for the course + + +00:09:10.320 --> 00:09:10.870 align:start position:0% +all right so that is it for the course +uh + +00:09:10.870 --> 00:09:10.880 align:start position:0% +uh + + +00:09:10.880 --> 00:09:13.269 align:start position:0% +uh +introduction<00:09:11.920> um<00:09:12.399> if<00:09:12.560> you<00:09:12.640> have<00:09:12.880> any + +00:09:13.269 --> 00:09:13.279 align:start position:0% +introduction um if you have any + + +00:09:13.279 --> 00:09:14.550 align:start position:0% +introduction um if you have any +questions<00:09:13.680> i<00:09:13.760> think<00:09:13.920> it's<00:09:14.080> a<00:09:14.160> good<00:09:14.320> time + +00:09:14.550 --> 00:09:14.560 align:start position:0% +questions i think it's a good time + + +00:09:14.560 --> 00:09:16.150 align:start position:0% +questions i think it's a good time +either<00:09:14.880> to<00:09:15.040> raise<00:09:15.279> your<00:09:15.519> hand + +00:09:16.150 --> 00:09:16.160 align:start position:0% +either to raise your hand + + +00:09:16.160 --> 00:09:22.310 align:start position:0% +either to raise your hand +or<00:09:16.560> post<00:09:16.880> it<00:09:17.120> on<00:09:17.279> the<00:09:17.600> chat<00:09:20.839> box<00:09:21.920> all<00:09:22.000> right<00:09:22.160> so + +00:09:22.310 --> 00:09:22.320 align:start position:0% +or post it on the chat box all right so + + +00:09:22.320 --> 00:09:23.030 align:start position:0% +or post it on the chat box all right so +i<00:09:22.480> see<00:09:22.720> that + +00:09:23.030 --> 00:09:23.040 align:start position:0% +i see that + + +00:09:23.040 --> 00:09:25.190 align:start position:0% +i see that +you<00:09:23.200> know<00:09:23.519> there's<00:09:23.839> a<00:09:24.000> conflict<00:09:24.560> for<00:09:24.800> some<00:09:24.959> of + +00:09:25.190 --> 00:09:25.200 align:start position:0% +you know there's a conflict for some of + + +00:09:25.200 --> 00:09:26.710 align:start position:0% +you know there's a conflict for some of +you<00:09:25.519> for<00:09:25.839> office<00:09:26.240> hours + +00:09:26.710 --> 00:09:26.720 align:start position:0% +you for office hours + + +00:09:26.720 --> 00:09:29.750 align:start position:0% +you for office hours +um<00:09:27.360> you<00:09:27.519> know<00:09:28.480> it's<00:09:28.720> a<00:09:28.800> this<00:09:29.040> is<00:09:29.120> a<00:09:29.200> very<00:09:29.519> big + +00:09:29.750 --> 00:09:29.760 align:start position:0% +um you know it's a this is a very big + + +00:09:29.760 --> 00:09:31.430 align:start position:0% +um you know it's a this is a very big +class<00:09:30.080> so<00:09:30.240> it<00:09:30.399> is<00:09:30.560> very<00:09:30.800> hard + +00:09:31.430 --> 00:09:31.440 align:start position:0% +class so it is very hard + + +00:09:31.440 --> 00:09:33.910 align:start position:0% +class so it is very hard +to<00:09:31.600> find<00:09:31.920> a<00:09:32.000> time<00:09:32.320> that<00:09:32.560> works<00:09:32.800> for<00:09:33.120> everyone + +00:09:33.910 --> 00:09:33.920 align:start position:0% +to find a time that works for everyone + + +00:09:33.920 --> 00:09:35.509 align:start position:0% +to find a time that works for everyone +but<00:09:34.160> i<00:09:34.320> will<00:09:34.720> try<00:09:35.040> to + +00:09:35.509 --> 00:09:35.519 align:start position:0% +but i will try to + + +00:09:35.519 --> 00:09:38.790 align:start position:0% +but i will try to +some<00:09:35.760> weeks<00:09:36.080> i'll<00:09:36.320> try<00:09:36.560> to<00:09:37.360> change<00:09:37.760> the<00:09:38.399> timing + +00:09:38.790 --> 00:09:38.800 align:start position:0% +some weeks i'll try to change the timing + + +00:09:38.800 --> 00:09:40.550 align:start position:0% +some weeks i'll try to change the timing +for<00:09:39.040> my<00:09:39.279> office<00:09:39.600> hour<00:09:39.920> so + +00:09:40.550 --> 00:09:40.560 align:start position:0% +for my office hour so + + +00:09:40.560 --> 00:09:43.110 align:start position:0% +for my office hour so +uh<00:09:41.040> hopefully<00:09:41.360> with<00:09:41.600> high<00:09:41.760> probability<00:09:42.959> all + +00:09:43.110 --> 00:09:43.120 align:start position:0% +uh hopefully with high probability all + + +00:09:43.120 --> 00:09:45.110 align:start position:0% +uh hopefully with high probability all +of<00:09:43.279> you<00:09:43.440> will<00:09:43.600> be<00:09:43.760> able<00:09:43.920> to<00:09:44.160> attend + +00:09:45.110 --> 00:09:45.120 align:start position:0% +of you will be able to attend + + +00:09:45.120 --> 00:09:51.190 align:start position:0% +of you will be able to attend +in<00:09:45.279> some<00:09:45.519> of<00:09:45.680> my<00:09:45.839> office<00:09:48.839> hours + +00:09:51.190 --> 00:09:51.200 align:start position:0% +in some of my office hours + + +00:09:51.200 --> 00:09:52.550 align:start position:0% +in some of my office hours +all<00:09:51.279> right<00:09:51.440> so<00:09:51.600> there's<00:09:51.839> another<00:09:52.240> question + +00:09:52.550 --> 00:09:52.560 align:start position:0% +all right so there's another question + + +00:09:52.560 --> 00:09:55.110 align:start position:0% +all right so there's another question +what<00:09:52.800> is<00:09:52.959> the<00:09:53.200> typical<00:09:54.160> team<00:09:54.480> size<00:09:54.800> you're + +00:09:55.110 --> 00:09:55.120 align:start position:0% +what is the typical team size you're + + +00:09:55.120 --> 00:09:57.110 align:start position:0% +what is the typical team size you're +expecting<00:09:55.600> for<00:09:55.839> the<00:09:56.080> project + +00:09:57.110 --> 00:09:57.120 align:start position:0% +expecting for the project + + +00:09:57.120 --> 00:10:00.150 align:start position:0% +expecting for the project +uh<00:09:58.160> it<00:09:58.320> is<00:09:58.640> up<00:09:58.800> to<00:09:59.040> four<00:09:59.279> but<00:09:59.519> we<00:09:59.680> are<00:09:59.839> very + +00:10:00.150 --> 00:10:00.160 align:start position:0% +uh it is up to four but we are very + + +00:10:00.160 --> 00:10:02.310 align:start position:0% +uh it is up to four but we are very +flexible<00:10:00.800> to<00:10:00.959> be<00:10:01.120> honest<00:10:01.440> so<00:10:01.600> you<00:10:01.760> can<00:10:02.000> do + +00:10:02.310 --> 00:10:02.320 align:start position:0% +flexible to be honest so you can do + + +00:10:02.320 --> 00:10:04.949 align:start position:0% +flexible to be honest so you can do +a<00:10:02.480> project<00:10:02.880> by<00:10:03.040> yourself<00:10:03.680> you<00:10:03.760> know<00:10:04.079> two<00:10:04.640> with + +00:10:04.949 --> 00:10:04.959 align:start position:0% +a project by yourself you know two with + + +00:10:04.959 --> 00:10:05.509 align:start position:0% +a project by yourself you know two with +you<00:10:05.040> know + +00:10:05.509 --> 00:10:05.519 align:start position:0% +you know + + +00:10:05.519 --> 00:10:07.670 align:start position:0% +you know +one<00:10:05.760> another<00:10:06.160> student<00:10:06.560> or<00:10:06.720> some<00:10:07.279> you<00:10:07.360> know + +00:10:07.670 --> 00:10:07.680 align:start position:0% +one another student or some you know + + +00:10:07.680 --> 00:10:08.949 align:start position:0% +one another student or some you know +other<00:10:08.000> students + +00:10:08.949 --> 00:10:08.959 align:start position:0% +other students + + +00:10:08.959 --> 00:10:10.949 align:start position:0% +other students +uh<00:10:09.360> we<00:10:09.600> don't<00:10:09.839> want<00:10:10.160> you<00:10:10.320> know<00:10:10.640> the<00:10:10.720> only + +00:10:10.949 --> 00:10:10.959 align:start position:0% +uh we don't want you know the only + + +00:10:10.959 --> 00:10:12.389 align:start position:0% +uh we don't want you know the only +things<00:10:11.200> that<00:10:11.279> we<00:10:11.440> don't<00:10:11.600> want<00:10:11.839> like<00:10:12.079> maybe + +00:10:12.389 --> 00:10:12.399 align:start position:0% +things that we don't want like maybe + + +00:10:12.399 --> 00:10:13.750 align:start position:0% +things that we don't want like maybe +like<00:10:12.560> 15<00:10:13.040> of<00:10:13.120> you<00:10:13.360> to + +00:10:13.750 --> 00:10:13.760 align:start position:0% +like 15 of you to + + +00:10:13.760 --> 00:10:15.590 align:start position:0% +like 15 of you to +you<00:10:13.839> know<00:10:14.079> bundle<00:10:14.560> up<00:10:14.720> and<00:10:14.880> like<00:10:15.040> have<00:10:15.279> one + +00:10:15.590 --> 00:10:15.600 align:start position:0% +you know bundle up and like have one + + +00:10:15.600 --> 00:10:17.190 align:start position:0% +you know bundle up and like have one +project<00:10:16.000> so<00:10:16.160> because<00:10:16.480> in<00:10:16.560> that<00:10:16.720> case<00:10:16.959> it<00:10:17.040> will + +00:10:17.190 --> 00:10:17.200 align:start position:0% +project so because in that case it will + + +00:10:17.200 --> 00:10:18.069 align:start position:0% +project so because in that case it will +be<00:10:17.360> a<00:10:17.440> little<00:10:17.600> bit + +00:10:18.069 --> 00:10:18.079 align:start position:0% +be a little bit + + +00:10:18.079 --> 00:10:21.590 align:start position:0% +be a little bit +harder<00:10:18.560> for<00:10:18.800> us<00:10:19.040> to<00:10:19.760> evaluate<00:10:20.720> contributions + +00:10:21.590 --> 00:10:21.600 align:start position:0% +harder for us to evaluate contributions + + +00:10:21.600 --> 00:10:23.110 align:start position:0% +harder for us to evaluate contributions +of<00:10:21.760> each<00:10:22.079> student + +00:10:23.110 --> 00:10:23.120 align:start position:0% +of each student + + +00:10:23.120 --> 00:10:25.670 align:start position:0% +of each student +but<00:10:23.760> even<00:10:24.160> we<00:10:24.399> are<00:10:24.560> we<00:10:24.720> can<00:10:24.880> be<00:10:25.040> flexible<00:10:25.600> in + +00:10:25.670 --> 00:10:25.680 align:start position:0% +but even we are we can be flexible in + + +00:10:25.680 --> 00:10:27.350 align:start position:0% +but even we are we can be flexible in +that<00:10:25.839> case<00:10:26.079> if<00:10:26.240> you<00:10:26.320> think<00:10:26.480> that<00:10:26.720> okay<00:10:26.959> maybe + +00:10:27.350 --> 00:10:27.360 align:start position:0% +that case if you think that okay maybe + + +00:10:27.360 --> 00:10:28.470 align:start position:0% +that case if you think that okay maybe +five<00:10:27.680> of<00:10:27.839> you + +00:10:28.470 --> 00:10:28.480 align:start position:0% +five of you + + +00:10:28.480 --> 00:10:30.949 align:start position:0% +five of you +uh<00:10:28.800> have<00:10:29.120> some<00:10:29.360> really<00:10:29.839> interesting<00:10:30.320> ideas + +00:10:30.949 --> 00:10:30.959 align:start position:0% +uh have some really interesting ideas + + +00:10:30.959 --> 00:10:32.949 align:start position:0% +uh have some really interesting ideas +and<00:10:31.279> you<00:10:31.519> have<00:10:31.920> a<00:10:32.160> plan + +00:10:32.949 --> 00:10:32.959 align:start position:0% +and you have a plan + + +00:10:32.959 --> 00:10:35.829 align:start position:0% +and you have a plan +for<00:10:33.200> each<00:10:33.600> student<00:10:34.079> to<00:10:34.720> work<00:10:35.040> on<00:10:35.200> a<00:10:35.279> specific + +00:10:35.829 --> 00:10:35.839 align:start position:0% +for each student to work on a specific + + +00:10:35.839 --> 00:10:37.269 align:start position:0% +for each student to work on a specific +aspect<00:10:36.240> of<00:10:36.320> that<00:10:36.560> project + +00:10:37.269 --> 00:10:37.279 align:start position:0% +aspect of that project + + +00:10:37.279 --> 00:10:39.350 align:start position:0% +aspect of that project +you<00:10:37.440> can<00:10:37.600> definitely<00:10:38.160> talk<00:10:38.480> to<00:10:38.560> me<00:10:38.880> and<00:10:39.120> we'll + +00:10:39.350 --> 00:10:39.360 align:start position:0% +you can definitely talk to me and we'll + + +00:10:39.360 --> 00:10:40.949 align:start position:0% +you can definitely talk to me and we'll +work<00:10:39.600> it<00:10:39.760> out + +00:10:40.949 --> 00:10:40.959 align:start position:0% +work it out + + +00:10:40.959 --> 00:10:44.069 align:start position:0% +work it out +so<00:10:41.200> you<00:10:41.279> see<00:10:41.519> a<00:10:41.760> hand<00:10:42.000> is<00:10:42.240> raised<00:10:43.360> please<00:10:43.920> ask + +00:10:44.069 --> 00:10:44.079 align:start position:0% +so you see a hand is raised please ask + + +00:10:44.079 --> 00:10:45.509 align:start position:0% +so you see a hand is raised please ask +your<00:10:44.320> question + +00:10:45.509 --> 00:10:45.519 align:start position:0% +your question + + +00:10:45.519 --> 00:10:48.150 align:start position:0% +your question +oh<00:10:45.680> hi<00:10:45.920> so<00:10:46.160> dr<00:10:46.560> soheil<00:10:47.040> so<00:10:47.360> you<00:10:47.600> said<00:10:47.839> in<00:10:48.000> this + +00:10:48.150 --> 00:10:48.160 align:start position:0% +oh hi so dr soheil so you said in this + + +00:10:48.160 --> 00:10:48.790 align:start position:0% +oh hi so dr soheil so you said in this +class<00:10:48.399> we + +00:10:48.790 --> 00:10:48.800 align:start position:0% +class we + + +00:10:48.800 --> 00:10:50.790 align:start position:0% +class we +have<00:10:49.040> a<00:10:49.200> chance<00:10:49.600> actually<00:10:49.920> to<00:10:50.160> work<00:10:50.480> out<00:10:50.640> a + +00:10:50.790 --> 00:10:50.800 align:start position:0% +have a chance actually to work out a + + +00:10:50.800 --> 00:10:52.550 align:start position:0% +have a chance actually to work out a +paper<00:10:51.519> basically<00:10:51.920> in<00:10:52.079> your<00:10:52.240> group + +00:10:52.550 --> 00:10:52.560 align:start position:0% +paper basically in your group + + +00:10:52.560 --> 00:10:56.150 align:start position:0% +paper basically in your group +is<00:10:52.720> that<00:10:54.160> first<00:10:55.360> assignment + +00:10:56.150 --> 00:10:56.160 align:start position:0% +is that first assignment + + +00:10:56.160 --> 00:11:00.150 align:start position:0% +is that first assignment +yeah<00:10:58.399> sorry<00:10:58.959> can<00:10:59.120> you<00:10:59.200> repeat<00:10:59.519> your<00:10:59.680> question + +00:11:00.150 --> 00:11:00.160 align:start position:0% +yeah sorry can you repeat your question + + +00:11:00.160 --> 00:11:02.470 align:start position:0% +yeah sorry can you repeat your question +in<00:11:00.320> terms<00:11:00.560> of<00:11:00.720> the<00:11:01.279> project<00:11:01.760> or<00:11:02.000> in<00:11:02.079> terms<00:11:02.320> of + +00:11:02.470 --> 00:11:02.480 align:start position:0% +in terms of the project or in terms of + + +00:11:02.480 --> 00:11:04.470 align:start position:0% +in terms of the project or in terms of +the<00:11:02.560> papers<00:11:03.040> that<00:11:03.200> we'll<00:11:03.360> be<00:11:03.600> reading<00:11:04.079> for + +00:11:04.470 --> 00:11:04.480 align:start position:0% +the papers that we'll be reading for + + +00:11:04.480 --> 00:11:07.750 align:start position:0% +the papers that we'll be reading for +the<00:11:04.640> lectures<00:11:05.519> oh<00:11:05.839> so<00:11:06.480> you<00:11:06.800> uh<00:11:07.120> i<00:11:07.279> mean<00:11:07.440> so<00:11:07.680> i + +00:11:07.750 --> 00:11:07.760 align:start position:0% +the lectures oh so you uh i mean so i + + +00:11:07.760 --> 00:11:09.269 align:start position:0% +the lectures oh so you uh i mean so i +saw<00:11:08.000> actually<00:11:08.320> from<00:11:08.480> the<00:11:08.640> syllabus + +00:11:09.269 --> 00:11:09.279 align:start position:0% +saw actually from the syllabus + + +00:11:09.279 --> 00:11:11.190 align:start position:0% +saw actually from the syllabus +in<00:11:09.360> this<00:11:09.600> class<00:11:09.920> we<00:11:10.079> basically<00:11:10.480> have<00:11:10.720> a<00:11:10.880> chance + +00:11:11.190 --> 00:11:11.200 align:start position:0% +in this class we basically have a chance + + +00:11:11.200 --> 00:11:13.590 align:start position:0% +in this class we basically have a chance +to<00:11:11.440> actually<00:11:11.760> make<00:11:12.000> a<00:11:12.160> paper<00:11:12.560> by<00:11:12.800> ourselves + +00:11:13.590 --> 00:11:13.600 align:start position:0% +to actually make a paper by ourselves + + +00:11:13.600 --> 00:11:16.310 align:start position:0% +to actually make a paper by ourselves +is<00:11:13.760> that<00:11:14.240> the<00:11:14.959> idea<00:11:15.279> yes<00:11:15.760> so<00:11:15.920> that<00:11:16.160> is + +00:11:16.310 --> 00:11:16.320 align:start position:0% +is that the idea yes so that is + + +00:11:16.320 --> 00:11:17.910 align:start position:0% +is that the idea yes so that is +basically<00:11:16.959> a<00:11:17.200> project + +00:11:17.910 --> 00:11:17.920 align:start position:0% +basically a project + + +00:11:17.920 --> 00:11:20.470 align:start position:0% +basically a project +so<00:11:18.640> a<00:11:18.800> project<00:11:19.440> yeah<00:11:19.680> so<00:11:19.839> you<00:11:20.079> basically + +00:11:20.470 --> 00:11:20.480 align:start position:0% +so a project yeah so you basically + + +00:11:20.480 --> 00:11:22.150 align:start position:0% +so a project yeah so you basically +within<00:11:20.880> a<00:11:20.959> project<00:11:21.279> you<00:11:21.519> basically + +00:11:22.150 --> 00:11:22.160 align:start position:0% +within a project you basically + + +00:11:22.160 --> 00:11:25.110 align:start position:0% +within a project you basically +uh<00:11:22.720> you<00:11:22.800> know<00:11:23.040> write<00:11:23.279> a<00:11:23.360> paper<00:11:24.240> uh<00:11:24.640> for<00:11:24.880> this + +00:11:25.110 --> 00:11:25.120 align:start position:0% +uh you know write a paper uh for this + + +00:11:25.120 --> 00:11:25.670 align:start position:0% +uh you know write a paper uh for this +course + +00:11:25.670 --> 00:11:25.680 align:start position:0% +course + + +00:11:25.680 --> 00:11:28.150 align:start position:0% +course +and<00:11:25.839> as<00:11:26.000> i<00:11:26.320> i<00:11:26.480> should<00:11:26.720> mention<00:11:27.120> that<00:11:28.000> you + +00:11:28.150 --> 00:11:28.160 align:start position:0% +and as i i should mention that you + + +00:11:28.160 --> 00:11:28.790 align:start position:0% +and as i i should mention that you +cannot + +00:11:28.790 --> 00:11:28.800 align:start position:0% +cannot + + +00:11:28.800 --> 00:11:31.190 align:start position:0% +cannot +in<00:11:28.959> a<00:11:29.040> way<00:11:29.440> quote<00:11:29.680> unquote<00:11:30.160> recycle<00:11:30.800> your + +00:11:31.190 --> 00:11:31.200 align:start position:0% +in a way quote unquote recycle your + + +00:11:31.200 --> 00:11:32.630 align:start position:0% +in a way quote unquote recycle your +previous<00:11:31.680> uh + +00:11:32.630 --> 00:11:32.640 align:start position:0% +previous uh + + +00:11:32.640 --> 00:11:34.870 align:start position:0% +previous uh +projects<00:11:33.120> and<00:11:33.279> papers<00:11:33.680> right<00:11:33.839> so<00:11:34.320> i<00:11:34.480> know<00:11:34.640> most + +00:11:34.870 --> 00:11:34.880 align:start position:0% +projects and papers right so i know most + + +00:11:34.880 --> 00:11:36.389 align:start position:0% +projects and papers right so i know most +of<00:11:34.959> you<00:11:35.120> are<00:11:35.200> doing<00:11:35.519> already<00:11:35.920> some<00:11:36.160> great + +00:11:36.389 --> 00:11:36.399 align:start position:0% +of you are doing already some great + + +00:11:36.399 --> 00:11:37.750 align:start position:0% +of you are doing already some great +research<00:11:36.959> with<00:11:37.200> your + +00:11:37.750 --> 00:11:37.760 align:start position:0% +research with your + + +00:11:37.760 --> 00:11:39.930 align:start position:0% +research with your +you<00:11:37.839> know<00:11:38.079> advisors<00:11:38.720> and<00:11:38.800> supervisors + +00:11:39.930 --> 00:11:39.940 align:start position:0% +you know advisors and supervisors + + +00:11:39.940 --> 00:11:41.590 align:start position:0% +you know advisors and supervisors +[Music] + +00:11:41.590 --> 00:11:41.600 align:start position:0% +[Music] + + +00:11:41.600 --> 00:11:43.990 align:start position:0% +[Music] +for<00:11:41.839> this<00:11:42.079> course<00:11:42.320> you<00:11:42.480> cannot<00:11:43.040> resubmit<00:11:43.760> some + +00:11:43.990 --> 00:11:44.000 align:start position:0% +for this course you cannot resubmit some + + +00:11:44.000 --> 00:11:46.790 align:start position:0% +for this course you cannot resubmit some +of<00:11:44.160> your<00:11:44.480> previous<00:11:44.959> work<00:11:45.279> so<00:11:45.440> we<00:11:45.600> want<00:11:45.920> you<00:11:46.160> to + +00:11:46.790 --> 00:11:46.800 align:start position:0% +of your previous work so we want you to + + +00:11:46.800 --> 00:11:48.949 align:start position:0% +of your previous work so we want you to +you<00:11:46.959> know<00:11:47.279> think<00:11:47.519> about<00:11:47.839> some<00:11:48.079> new<00:11:48.320> problems + +00:11:48.949 --> 00:11:48.959 align:start position:0% +you know think about some new problems + + +00:11:48.959 --> 00:11:51.350 align:start position:0% +you know think about some new problems +or<00:11:49.200> some<00:11:49.360> related<00:11:49.839> problems<00:11:50.240> that<00:11:50.399> you<00:11:50.639> have + +00:11:51.350 --> 00:11:51.360 align:start position:0% +or some related problems that you have + + +00:11:51.360 --> 00:11:53.750 align:start position:0% +or some related problems that you have +not<00:11:51.760> you<00:11:51.920> know<00:11:52.160> conducted<00:11:52.720> the<00:11:53.040> research<00:11:53.519> for + +00:11:53.750 --> 00:11:53.760 align:start position:0% +not you know conducted the research for + + +00:11:53.760 --> 00:11:54.949 align:start position:0% +not you know conducted the research for +that<00:11:54.320> and + +00:11:54.949 --> 00:11:54.959 align:start position:0% +that and + + +00:11:54.959 --> 00:11:57.350 align:start position:0% +that and +do<00:11:55.120> it<00:11:55.279> within<00:11:55.600> this<00:11:55.839> course<00:11:56.160> in<00:11:56.240> order<00:11:56.560> to<00:11:57.120> be + +00:11:57.350 --> 00:11:57.360 align:start position:0% +do it within this course in order to be + + +00:11:57.360 --> 00:11:58.550 align:start position:0% +do it within this course in order to be +fair<00:11:57.839> um + +00:11:58.550 --> 00:11:58.560 align:start position:0% +fair um + + +00:11:58.560 --> 00:12:01.750 align:start position:0% +fair um +you<00:11:58.720> know<00:11:58.959> to<00:11:59.920> all<00:12:00.160> students<00:12:01.279> all<00:12:01.360> right<00:12:01.519> thank + +00:12:01.750 --> 00:12:01.760 align:start position:0% +you know to all students all right thank + + +00:12:01.760 --> 00:12:02.550 align:start position:0% +you know to all students all right thank +you + +00:12:02.550 --> 00:12:02.560 align:start position:0% +you + + +00:12:02.560 --> 00:12:04.550 align:start position:0% +you +very<00:12:02.800> much<00:12:03.040> so<00:12:03.200> i<00:12:03.279> see<00:12:03.519> some<00:12:03.680> more<00:12:03.920> questions + +00:12:04.550 --> 00:12:04.560 align:start position:0% +very much so i see some more questions + + +00:12:04.560 --> 00:12:05.670 align:start position:0% +very much so i see some more questions +um + +00:12:05.670 --> 00:12:05.680 align:start position:0% +um + + +00:12:05.680 --> 00:12:08.949 align:start position:0% +um +could<00:12:05.920> we<00:12:06.079> use<00:12:06.480> this<00:12:06.959> course<00:12:07.279> project<00:12:07.760> for<00:12:08.160> uh + +00:12:08.949 --> 00:12:08.959 align:start position:0% +could we use this course project for uh + + +00:12:08.959 --> 00:12:12.230 align:start position:0% +could we use this course project for uh +two<00:12:09.200> different<00:12:09.519> courses<00:12:10.240> no<00:12:10.720> unless + +00:12:12.230 --> 00:12:12.240 align:start position:0% +two different courses no unless + + +00:12:12.240 --> 00:12:14.790 align:start position:0% +two different courses no unless +you<00:12:12.399> know<00:12:12.959> you<00:12:13.279> basically<00:12:13.839> have<00:12:14.320> very + +00:12:14.790 --> 00:12:14.800 align:start position:0% +you know you basically have very + + +00:12:14.800 --> 00:12:17.430 align:start position:0% +you know you basically have very +separate<00:12:15.440> goals<00:12:16.000> and<00:12:16.160> separate + +00:12:17.430 --> 00:12:17.440 align:start position:0% +separate goals and separate + + +00:12:17.440 --> 00:12:18.550 align:start position:0% +separate goals and separate +you<00:12:17.519> know<00:12:17.680> techniques<00:12:18.160> that<00:12:18.320> you<00:12:18.480> are + +00:12:18.550 --> 00:12:18.560 align:start position:0% +you know techniques that you are + + +00:12:18.560 --> 00:12:21.590 align:start position:0% +you know techniques that you are +developing<00:12:20.240> for<00:12:20.560> different<00:12:20.880> courses<00:12:21.279> so<00:12:21.440> you + +00:12:21.590 --> 00:12:21.600 align:start position:0% +developing for different courses so you + + +00:12:21.600 --> 00:12:23.590 align:start position:0% +developing for different courses so you +cannot<00:12:21.920> submit<00:12:22.399> one<00:12:22.639> project<00:12:23.040> to<00:12:23.200> multiple + +00:12:23.590 --> 00:12:23.600 align:start position:0% +cannot submit one project to multiple + + +00:12:23.600 --> 00:12:25.110 align:start position:0% +cannot submit one project to multiple +courses + +00:12:25.110 --> 00:12:25.120 align:start position:0% +courses + + +00:12:25.120 --> 00:12:28.069 align:start position:0% +courses +there's<00:12:25.360> another<00:12:25.680> question<00:12:26.079> what<00:12:26.399> would<00:12:26.800> an + +00:12:28.069 --> 00:12:28.079 align:start position:0% +there's another question what would an + + +00:12:28.079 --> 00:12:30.949 align:start position:0% +there's another question what would an +example<00:12:28.560> of<00:12:28.720> a<00:12:28.880> project<00:12:29.360> be + +00:12:30.949 --> 00:12:30.959 align:start position:0% +example of a project be + + +00:12:30.959 --> 00:12:33.829 align:start position:0% +example of a project be +it's<00:12:31.279> very<00:12:31.519> open-ended<00:12:32.160> so<00:12:32.399> as<00:12:32.560> i<00:12:32.639> mentioned + +00:12:33.829 --> 00:12:33.839 align:start position:0% +it's very open-ended so as i mentioned + + +00:12:33.839 --> 00:12:35.110 align:start position:0% +it's very open-ended so as i mentioned +this<00:12:34.000> is<00:12:34.160> not<00:12:34.320> an<00:12:34.480> undergrad + +00:12:35.110 --> 00:12:35.120 align:start position:0% +this is not an undergrad + + +00:12:35.120 --> 00:12:36.710 align:start position:0% +this is not an undergrad +course<00:12:35.600> right<00:12:35.760> so<00:12:35.920> we<00:12:36.079> are<00:12:36.160> not<00:12:36.320> going<00:12:36.480> to<00:12:36.560> give + +00:12:36.710 --> 00:12:36.720 align:start position:0% +course right so we are not going to give + + +00:12:36.720 --> 00:12:39.430 align:start position:0% +course right so we are not going to give +you<00:12:36.959> like<00:12:37.680> specific<00:12:38.320> guidelines<00:12:38.880> of<00:12:39.040> like + +00:12:39.430 --> 00:12:39.440 align:start position:0% +you like specific guidelines of like + + +00:12:39.440 --> 00:12:41.910 align:start position:0% +you like specific guidelines of like +you<00:12:39.519> know<00:12:39.839> this<00:12:40.079> is<00:12:40.240> the<00:12:40.480> you<00:12:40.639> know<00:12:41.519> problem + +00:12:41.910 --> 00:12:41.920 align:start position:0% +you know this is the you know problem + + +00:12:41.920 --> 00:12:43.670 align:start position:0% +you know this is the you know problem +that<00:12:42.160> you<00:12:42.240> need<00:12:42.399> to<00:12:42.560> work<00:12:42.880> on<00:12:43.120> for<00:12:43.279> the<00:12:43.440> next + +00:12:43.670 --> 00:12:43.680 align:start position:0% +that you need to work on for the next + + +00:12:43.680 --> 00:12:44.470 align:start position:0% +that you need to work on for the next +week<00:12:44.000> so + +00:12:44.470 --> 00:12:44.480 align:start position:0% +week so + + +00:12:44.480 --> 00:12:46.310 align:start position:0% +week so +this<00:12:44.639> is<00:12:44.800> a<00:12:44.880> research<00:12:45.360> course<00:12:45.680> so<00:12:46.000> you<00:12:46.160> can + +00:12:46.310 --> 00:12:46.320 align:start position:0% +this is a research course so you can + + +00:12:46.320 --> 00:12:47.750 align:start position:0% +this is a research course so you can +take<00:12:46.560> a<00:12:46.880> look<00:12:47.120> into + +00:12:47.750 --> 00:12:47.760 align:start position:0% +take a look into + + +00:12:47.760 --> 00:12:50.150 align:start position:0% +take a look into +some<00:12:48.800> some<00:12:49.040> of<00:12:49.200> the<00:12:49.360> topics<00:12:49.680> some<00:12:49.920> of<00:12:50.000> the + +00:12:50.150 --> 00:12:50.160 align:start position:0% +some some of the topics some of the + + +00:12:50.160 --> 00:12:51.670 align:start position:0% +some some of the topics some of the +papers<00:12:50.480> that<00:12:50.639> we<00:12:50.800> are<00:12:50.880> covering<00:12:51.360> in<00:12:51.440> this + +00:12:51.670 --> 00:12:51.680 align:start position:0% +papers that we are covering in this + + +00:12:51.680 --> 00:12:52.389 align:start position:0% +papers that we are covering in this +course + +00:12:52.389 --> 00:12:52.399 align:start position:0% +course + + +00:12:52.399 --> 00:12:54.230 align:start position:0% +course +get<00:12:52.560> some<00:12:52.800> ideas<00:12:53.120> or<00:12:53.279> maybe<00:12:53.680> you<00:12:53.760> know<00:12:53.920> come<00:12:54.160> up + +00:12:54.230 --> 00:12:54.240 align:start position:0% +get some ideas or maybe you know come up + + +00:12:54.240 --> 00:12:55.910 align:start position:0% +get some ideas or maybe you know come up +with<00:12:54.480> different<00:12:55.279> topics + +00:12:55.910 --> 00:12:55.920 align:start position:0% +with different topics + + +00:12:55.920 --> 00:12:59.670 align:start position:0% +with different topics +so<00:12:56.079> we<00:12:56.240> are<00:12:56.399> very<00:12:56.639> flexible<00:12:57.360> in<00:12:57.440> that<00:12:57.760> sense + +00:12:59.670 --> 00:12:59.680 align:start position:0% +so we are very flexible in that sense + + +00:12:59.680 --> 00:13:01.590 align:start position:0% +so we are very flexible in that sense +so<00:12:59.920> there's<00:13:00.560> one<00:13:00.720> more<00:13:00.959> question<00:13:01.279> can + +00:13:01.590 --> 00:13:01.600 align:start position:0% +so there's one more question can + + +00:13:01.600 --> 00:13:03.350 align:start position:0% +so there's one more question can +students<00:13:02.079> who<00:13:02.240> are<00:13:02.399> auditing + +00:13:03.350 --> 00:13:03.360 align:start position:0% +students who are auditing + + +00:13:03.360 --> 00:13:05.910 align:start position:0% +students who are auditing +or<00:13:03.600> attending<00:13:04.079> the<00:13:04.240> course<00:13:04.560> submit<00:13:05.040> projects + +00:13:05.910 --> 00:13:05.920 align:start position:0% +or attending the course submit projects + + +00:13:05.920 --> 00:13:06.710 align:start position:0% +or attending the course submit projects +uh + +00:13:06.710 --> 00:13:06.720 align:start position:0% +uh + + +00:13:06.720 --> 00:13:09.990 align:start position:0% +uh +in<00:13:06.880> that<00:13:07.200> case<00:13:07.680> uh<00:13:08.320> you<00:13:08.560> should<00:13:09.200> contact<00:13:09.760> me + +00:13:09.990 --> 00:13:10.000 align:start position:0% +in that case uh you should contact me + + +00:13:10.000 --> 00:13:11.590 align:start position:0% +in that case uh you should contact me +and<00:13:10.240> course<00:13:10.560> tas<00:13:11.200> to + +00:13:11.590 --> 00:13:11.600 align:start position:0% +and course tas to + + +00:13:11.600 --> 00:13:13.750 align:start position:0% +and course tas to +discuss<00:13:12.480> you<00:13:12.639> know<00:13:12.880> because<00:13:13.200> this<00:13:13.360> is<00:13:13.519> a<00:13:13.600> big + +00:13:13.750 --> 00:13:13.760 align:start position:0% +discuss you know because this is a big + + +00:13:13.760 --> 00:13:15.509 align:start position:0% +discuss you know because this is a big +course<00:13:14.079> we<00:13:14.160> want<00:13:14.320> to<00:13:14.399> make<00:13:14.560> sure<00:13:14.800> that<00:13:14.959> we + +00:13:15.509 --> 00:13:15.519 align:start position:0% +course we want to make sure that we + + +00:13:15.519 --> 00:13:17.990 align:start position:0% +course we want to make sure that we +you<00:13:15.680> know<00:13:15.920> have<00:13:16.240> uh<00:13:16.959> you<00:13:17.040> know<00:13:17.279> resources<00:13:17.920> in + +00:13:17.990 --> 00:13:18.000 align:start position:0% +you know have uh you know resources in + + +00:13:18.000 --> 00:13:18.629 align:start position:0% +you know have uh you know resources in +terms<00:13:18.320> of + +00:13:18.629 --> 00:13:18.639 align:start position:0% +terms of + + +00:13:18.639 --> 00:13:22.629 align:start position:0% +terms of +uh<00:13:19.760> and<00:13:20.000> in<00:13:20.079> terms<00:13:20.399> of<00:13:20.880> handling<00:13:21.360> the + +00:13:22.629 --> 00:13:22.639 align:start position:0% +uh and in terms of handling the + + +00:13:22.639 --> 00:13:24.310 align:start position:0% +uh and in terms of handling the +you<00:13:22.800> know<00:13:23.040> different<00:13:23.360> projects<00:13:23.760> so<00:13:24.000> if<00:13:24.079> you're + +00:13:24.310 --> 00:13:24.320 align:start position:0% +you know different projects so if you're + + +00:13:24.320 --> 00:13:26.310 align:start position:0% +you know different projects so if you're +auditing<00:13:24.800> you're<00:13:25.040> not<00:13:25.519> you<00:13:25.680> know + +00:13:26.310 --> 00:13:26.320 align:start position:0% +auditing you're not you know + + +00:13:26.320 --> 00:13:29.350 align:start position:0% +auditing you're not you know +officially<00:13:27.360> registered<00:13:28.000> to<00:13:28.160> the<00:13:28.320> course + +00:13:29.350 --> 00:13:29.360 align:start position:0% +officially registered to the course + + +00:13:29.360 --> 00:13:32.710 align:start position:0% +officially registered to the course +contact<00:13:29.839> the<00:13:30.079> tas<00:13:30.560> and<00:13:30.720> myself<00:13:31.360> and<00:13:31.600> we<00:13:31.760> can + +00:13:32.710 --> 00:13:32.720 align:start position:0% +contact the tas and myself and we can + + +00:13:32.720 --> 00:13:35.910 align:start position:0% +contact the tas and myself and we can +see<00:13:32.959> if<00:13:33.200> it<00:13:33.360> is<00:13:33.519> a<00:13:33.680> good<00:13:33.920> fit<00:13:34.079> for<00:13:34.240> you<00:13:34.399> to + +00:13:35.910 --> 00:13:35.920 align:start position:0% +see if it is a good fit for you to + + +00:13:35.920 --> 00:13:39.990 align:start position:0% +see if it is a good fit for you to +work<00:13:36.240> on<00:13:36.399> a<00:13:36.560> class<00:13:36.880> project + +00:13:39.990 --> 00:13:40.000 align:start position:0% + + + +00:13:40.000 --> 00:13:41.990 align:start position:0% + +okay<00:13:40.240> so<00:13:40.399> one<00:13:40.639> more<00:13:40.880> question<00:13:41.199> is<00:13:41.360> there<00:13:41.600> some + +00:13:41.990 --> 00:13:42.000 align:start position:0% +okay so one more question is there some + + +00:13:42.000 --> 00:13:44.870 align:start position:0% +okay so one more question is there some +online<00:13:42.800> forum<00:13:43.199> where<00:13:43.519> we<00:13:43.680> can<00:13:44.079> try<00:13:44.399> to<00:13:44.560> form + +00:13:44.870 --> 00:13:44.880 align:start position:0% +online forum where we can try to form + + +00:13:44.880 --> 00:13:45.670 align:start position:0% +online forum where we can try to form +groups<00:13:45.199> yes + +00:13:45.670 --> 00:13:45.680 align:start position:0% +groups yes + + +00:13:45.680 --> 00:13:49.269 align:start position:0% +groups yes +i<00:13:45.839> think<00:13:46.079> i<00:13:46.399> skipped<00:13:46.800> it<00:13:48.079> there<00:13:48.240> is<00:13:48.399> a<00:13:48.560> piazza + +00:13:49.269 --> 00:13:49.279 align:start position:0% +i think i skipped it there is a piazza + + +00:13:49.279 --> 00:13:52.069 align:start position:0% +i think i skipped it there is a piazza +link + +00:13:52.069 --> 00:13:52.079 align:start position:0% + + + +00:13:52.079 --> 00:13:54.790 align:start position:0% + +here<00:13:52.399> so<00:13:52.639> you<00:13:52.800> should<00:13:53.199> just<00:13:53.519> there<00:13:54.240> to<00:13:54.480> that + +00:13:54.790 --> 00:13:54.800 align:start position:0% +here so you should just there to that + + +00:13:54.800 --> 00:13:55.590 align:start position:0% +here so you should just there to that +and<00:13:55.040> maybe + +00:13:55.590 --> 00:13:55.600 align:start position:0% +and maybe + + +00:13:55.600 --> 00:13:58.790 align:start position:0% +and maybe +you<00:13:55.760> know<00:13:56.000> there<00:13:56.240> you<00:13:56.320> can<00:13:56.560> post<00:13:56.880> your<00:13:57.760> ideas + +00:13:58.790 --> 00:13:58.800 align:start position:0% +you know there you can post your ideas + + +00:13:58.800 --> 00:14:01.910 align:start position:0% +you know there you can post your ideas +and<00:13:59.199> format<00:13:59.680> team<00:14:00.639> on<00:14:00.800> piazza<00:14:01.199> it's<00:14:01.360> very<00:14:01.600> easy + +00:14:01.910 --> 00:14:01.920 align:start position:0% +and format team on piazza it's very easy + + +00:14:01.920 --> 00:14:03.590 align:start position:0% +and format team on piazza it's very easy +to<00:14:02.000> use<00:14:02.240> if<00:14:02.399> you<00:14:02.480> haven't<00:14:02.800> used<00:14:03.120> it + +00:14:03.590 --> 00:14:03.600 align:start position:0% +to use if you haven't used it + + +00:14:03.600 --> 00:14:08.069 align:start position:0% +to use if you haven't used it +before<00:14:06.480> all<00:14:06.560> right<00:14:06.880> one<00:14:07.120> more<00:14:07.360> question + +00:14:08.069 --> 00:14:08.079 align:start position:0% +before all right one more question + + +00:14:08.079 --> 00:14:11.430 align:start position:0% +before all right one more question +is<00:14:08.240> there<00:14:08.480> a<00:14:08.639> metric<00:14:09.279> or<00:14:09.519> a<00:14:09.760> quiz<00:14:10.240> to<00:14:10.800> see<00:14:11.040> if<00:14:11.199> we + +00:14:11.430 --> 00:14:11.440 align:start position:0% +is there a metric or a quiz to see if we + + +00:14:11.440 --> 00:14:13.829 align:start position:0% +is there a metric or a quiz to see if we +know<00:14:11.680> enough<00:14:12.000> of<00:14:12.160> the<00:14:12.320> required<00:14:12.800> background + +00:14:13.829 --> 00:14:13.839 align:start position:0% +know enough of the required background + + +00:14:13.839 --> 00:14:17.910 align:start position:0% +know enough of the required background +will<00:14:14.160> potentially<00:14:14.880> you<00:14:15.040> know<00:14:15.360> post + +00:14:17.910 --> 00:14:17.920 align:start position:0% + + + +00:14:17.920 --> 00:14:21.509 align:start position:0% + +some<00:14:19.120> maybe<00:14:19.680> background<00:14:20.560> documents + +00:14:21.509 --> 00:14:21.519 align:start position:0% +some maybe background documents + + +00:14:21.519 --> 00:14:25.670 align:start position:0% +some maybe background documents +um<00:14:23.120> you<00:14:23.279> can<00:14:23.440> actually<00:14:23.920> take<00:14:24.160> a<00:14:24.320> look<00:14:24.560> into + +00:14:25.670 --> 00:14:25.680 align:start position:0% +um you can actually take a look into + + +00:14:25.680 --> 00:14:28.150 align:start position:0% +um you can actually take a look into +some<00:14:25.920> of<00:14:26.079> the<00:14:26.720> previous<00:14:27.199> courses<00:14:27.600> that<00:14:27.760> i<00:14:27.920> have + +00:14:28.150 --> 00:14:28.160 align:start position:0% +some of the previous courses that i have + + +00:14:28.160 --> 00:14:31.430 align:start position:0% +some of the previous courses that i have +taught<00:14:28.639> you<00:14:28.800> can<00:14:28.959> go<00:14:29.120> to<00:14:29.279> my<00:14:29.519> course<00:14:29.760> webpage + +00:14:31.430 --> 00:14:31.440 align:start position:0% +taught you can go to my course webpage + + +00:14:31.440 --> 00:14:33.990 align:start position:0% +taught you can go to my course webpage +726<00:14:32.320> a<00:14:32.399> machine<00:14:32.720> learning<00:14:33.040> course<00:14:33.680> take<00:14:33.920> a + +00:14:33.990 --> 00:14:34.000 align:start position:0% +726 a machine learning course take a + + +00:14:34.000 --> 00:14:35.670 align:start position:0% +726 a machine learning course take a +look<00:14:34.240> at<00:14:34.320> the<00:14:34.480> backgrounds<00:14:35.120> and<00:14:35.279> take<00:14:35.440> a<00:14:35.519> look + +00:14:35.670 --> 00:14:35.680 align:start position:0% +look at the backgrounds and take a look + + +00:14:35.680 --> 00:14:36.949 align:start position:0% +look at the backgrounds and take a look +at<00:14:35.839> the<00:14:35.920> lectures<00:14:36.399> of<00:14:36.560> that + +00:14:36.949 --> 00:14:36.959 align:start position:0% +at the lectures of that + + +00:14:36.959 --> 00:14:40.150 align:start position:0% +at the lectures of that +class<00:14:37.360> and<00:14:37.519> see<00:14:37.920> if + +00:14:40.150 --> 00:14:40.160 align:start position:0% +class and see if + + +00:14:40.160 --> 00:14:42.389 align:start position:0% +class and see if +it<00:14:40.399> is<00:14:40.639> you<00:14:40.720> know<00:14:40.880> easy<00:14:41.199> for<00:14:41.360> you<00:14:41.519> to<00:14:41.680> follow<00:14:42.000> it + +00:14:42.389 --> 00:14:42.399 align:start position:0% +it is you know easy for you to follow it + + +00:14:42.399 --> 00:14:44.069 align:start position:0% +it is you know easy for you to follow it +and<00:14:42.560> today<00:14:42.880> i'm<00:14:43.120> going<00:14:43.360> to + +00:14:44.069 --> 00:14:44.079 align:start position:0% +and today i'm going to + + +00:14:44.079 --> 00:14:46.629 align:start position:0% +and today i'm going to +briefly<00:14:44.639> cover<00:14:45.040> some<00:14:45.279> of<00:14:45.440> these<00:14:46.079> topics<00:14:46.560> and + +00:14:46.629 --> 00:14:46.639 align:start position:0% +briefly cover some of these topics and + + +00:14:46.639 --> 00:14:47.910 align:start position:0% +briefly cover some of these topics and +you'll<00:14:46.880> see<00:14:47.120> if + +00:14:47.910 --> 00:14:47.920 align:start position:0% +you'll see if + + +00:14:47.920 --> 00:14:51.590 align:start position:0% +you'll see if +it<00:14:48.160> is<00:14:49.440> easy<00:14:49.839> for<00:14:50.079> you<00:14:50.480> to + +00:14:51.590 --> 00:14:51.600 align:start position:0% +it is easy for you to + + +00:14:51.600 --> 00:14:53.269 align:start position:0% +it is easy for you to +follow<00:14:52.000> some<00:14:52.160> of<00:14:52.320> these<00:14:52.480> concepts<00:14:52.959> that<00:14:53.120> we'll + +00:14:53.269 --> 00:14:53.279 align:start position:0% +follow some of these concepts that we'll + + +00:14:53.279 --> 00:14:57.269 align:start position:0% +follow some of these concepts that we'll +be<00:14:53.440> covering + +00:14:57.269 --> 00:14:57.279 align:start position:0% + + + +00:14:57.279 --> 00:14:59.509 align:start position:0% + +okay<00:14:57.600> so<00:14:57.920> one<00:14:58.079> more<00:14:58.320> question<00:14:58.720> can<00:14:58.959> we<00:14:59.120> have<00:14:59.360> a + +00:14:59.509 --> 00:14:59.519 align:start position:0% +okay so one more question can we have a + + +00:14:59.519 --> 00:15:01.189 align:start position:0% +okay so one more question can we have a +list<00:14:59.760> of<00:14:59.920> potential<00:15:00.480> projects + +00:15:01.189 --> 00:15:01.199 align:start position:0% +list of potential projects + + +00:15:01.199 --> 00:15:04.870 align:start position:0% +list of potential projects +and<00:15:01.519> no<00:15:02.880> as<00:15:03.040> i<00:15:03.199> mentioned<00:15:04.000> you<00:15:04.160> know<00:15:04.399> i<00:15:04.639> want + +00:15:04.870 --> 00:15:04.880 align:start position:0% +and no as i mentioned you know i want + + +00:15:04.880 --> 00:15:05.590 align:start position:0% +and no as i mentioned you know i want +you<00:15:05.120> to + +00:15:05.590 --> 00:15:05.600 align:start position:0% +you to + + +00:15:05.600 --> 00:15:08.949 align:start position:0% +you to +think<00:15:05.839> about<00:15:06.240> these<00:15:06.639> uh<00:15:08.240> projects + +00:15:08.949 --> 00:15:08.959 align:start position:0% +think about these uh projects + + +00:15:08.959 --> 00:15:11.189 align:start position:0% +think about these uh projects +or<00:15:09.279> ideas<00:15:09.839> yourself<00:15:10.320> so<00:15:10.560> it<00:15:10.720> is<00:15:10.800> not<00:15:11.040> an + +00:15:11.189 --> 00:15:11.199 align:start position:0% +or ideas yourself so it is not an + + +00:15:11.199 --> 00:15:12.710 align:start position:0% +or ideas yourself so it is not an +undergrad<00:15:11.760> course + +00:15:12.710 --> 00:15:12.720 align:start position:0% +undergrad course + + +00:15:12.720 --> 00:15:15.829 align:start position:0% +undergrad course +and<00:15:12.959> we<00:15:13.120> are<00:15:13.360> very<00:15:13.680> flexible<00:15:14.800> uh<00:15:15.199> in + +00:15:15.829 --> 00:15:15.839 align:start position:0% +and we are very flexible uh in + + +00:15:15.839 --> 00:15:18.069 align:start position:0% +and we are very flexible uh in +you<00:15:16.000> know<00:15:16.160> basically<00:15:17.199> admitting<00:15:17.760> different + +00:15:18.069 --> 00:15:18.079 align:start position:0% +you know basically admitting different + + +00:15:18.079 --> 00:15:20.069 align:start position:0% +you know basically admitting different +types<00:15:18.399> of<00:15:18.480> projects<00:15:18.880> either<00:15:19.120> practical<00:15:19.680> or + +00:15:20.069 --> 00:15:20.079 align:start position:0% +types of projects either practical or + + +00:15:20.079 --> 00:15:24.069 align:start position:0% +types of projects either practical or +theoretical<00:15:22.560> and + +00:15:24.069 --> 00:15:24.079 align:start position:0% +theoretical and + + +00:15:24.079 --> 00:15:26.069 align:start position:0% +theoretical and +final<00:15:24.480> question<00:15:24.880> is<00:15:25.040> there<00:15:25.279> any<00:15:25.519> example<00:15:25.920> of + +00:15:26.069 --> 00:15:26.079 align:start position:0% +final question is there any example of + + +00:15:26.079 --> 00:15:27.910 align:start position:0% +final question is there any example of +lecture<00:15:26.399> describing<00:15:26.959> that<00:15:27.120> we<00:15:27.279> can + +00:15:27.910 --> 00:15:27.920 align:start position:0% +lecture describing that we can + + +00:15:27.920 --> 00:15:31.350 align:start position:0% +lecture describing that we can +refer<00:15:28.320> to<00:15:29.040> uh<00:15:29.600> i<00:15:29.759> think<00:15:30.079> today<00:15:30.880> we'll + +00:15:31.350 --> 00:15:31.360 align:start position:0% +refer to uh i think today we'll + + +00:15:31.360 --> 00:15:33.590 align:start position:0% +refer to uh i think today we'll +start<00:15:31.839> the<00:15:32.399> basically<00:15:32.800> there's<00:15:33.120> like<00:15:33.360> an + +00:15:33.590 --> 00:15:33.600 align:start position:0% +start the basically there's like an + + +00:15:33.600 --> 00:15:35.110 align:start position:0% +start the basically there's like an +overleaf + +00:15:35.110 --> 00:15:35.120 align:start position:0% +overleaf + + +00:15:35.120 --> 00:15:38.389 align:start position:0% +overleaf +link<00:15:35.519> that<00:15:35.759> we<00:15:35.920> have<00:15:36.160> already<00:15:36.639> shared + +00:15:38.389 --> 00:15:38.399 align:start position:0% +link that we have already shared + + +00:15:38.399 --> 00:15:40.150 align:start position:0% +link that we have already shared +i<00:15:38.480> think<00:15:38.720> after<00:15:39.040> today's<00:15:39.440> lecture<00:15:39.839> you<00:15:40.000> can + +00:15:40.150 --> 00:15:40.160 align:start position:0% +i think after today's lecture you can + + +00:15:40.160 --> 00:15:41.910 align:start position:0% +i think after today's lecture you can +just<00:15:40.399> like<00:15:40.639> follow<00:15:40.959> that<00:15:41.120> format<00:15:41.519> so<00:15:41.680> there's + +00:15:41.910 --> 00:15:41.920 align:start position:0% +just like follow that format so there's + + +00:15:41.920 --> 00:15:42.470 align:start position:0% +just like follow that format so there's +some + +00:15:42.470 --> 00:15:42.480 align:start position:0% +some + + +00:15:42.480 --> 00:15:44.629 align:start position:0% +some +formatting<00:15:43.199> already<00:15:43.759> integrated<00:15:44.320> in<00:15:44.399> that + +00:15:44.629 --> 00:15:44.639 align:start position:0% +formatting already integrated in that + + +00:15:44.639 --> 00:15:47.670 align:start position:0% +formatting already integrated in that +overleaf + +00:15:47.670 --> 00:15:47.680 align:start position:0% + + + +00:15:47.680 --> 00:15:49.990 align:start position:0% + +okay<00:15:48.240> so<00:15:48.959> i<00:15:49.040> think<00:15:49.199> that's<00:15:49.440> all<00:15:49.600> for<00:15:49.759> the + +00:15:49.990 --> 00:15:50.000 align:start position:0% +okay so i think that's all for the + + +00:15:50.000 --> 00:15:51.829 align:start position:0% +okay so i think that's all for the +questions<00:15:50.639> uh + +00:15:51.829 --> 00:15:51.839 align:start position:0% +questions uh + + +00:15:51.839 --> 00:15:53.829 align:start position:0% +questions uh +let's<00:15:52.240> start<00:15:52.560> the<00:15:52.800> technical<00:15:53.360> part<00:15:53.680> of + +00:15:53.829 --> 00:15:53.839 align:start position:0% +let's start the technical part of + + +00:15:53.839 --> 00:15:55.590 align:start position:0% +let's start the technical part of +today's<00:15:54.320> lecture + +00:15:55.590 --> 00:15:55.600 align:start position:0% +today's lecture + + +00:15:55.600 --> 00:15:58.150 align:start position:0% +today's lecture +so<00:15:55.920> as<00:15:56.160> you<00:15:56.560> know<00:15:56.959> most<00:15:57.199> of<00:15:57.360> you<00:15:57.519> know<00:15:57.839> i + +00:15:58.150 --> 00:15:58.160 align:start position:0% +so as you know most of you know i + + +00:15:58.160 --> 00:15:59.110 align:start position:0% +so as you know most of you know i +usually<00:15:58.639> like + +00:15:59.110 --> 00:15:59.120 align:start position:0% +usually like + + +00:15:59.120 --> 00:16:01.910 align:start position:0% +usually like +whiteboards<00:16:00.240> so<00:16:00.720> unfortunately<00:16:01.440> we<00:16:01.680> don't + +00:16:01.910 --> 00:16:01.920 align:start position:0% +whiteboards so unfortunately we don't + + +00:16:01.920 --> 00:16:03.030 align:start position:0% +whiteboards so unfortunately we don't +have + +00:16:03.030 --> 00:16:03.040 align:start position:0% +have + + +00:16:03.040 --> 00:16:05.829 align:start position:0% +have +a<00:16:03.920> physical<00:16:04.480> whiteboard<00:16:04.959> here<00:16:05.199> but<00:16:05.360> we'll<00:16:05.600> try + +00:16:05.829 --> 00:16:05.839 align:start position:0% +a physical whiteboard here but we'll try + + +00:16:05.839 --> 00:16:06.790 align:start position:0% +a physical whiteboard here but we'll try +to<00:16:06.000> use<00:16:06.320> some + +00:16:06.790 --> 00:16:06.800 align:start position:0% +to use some + + +00:16:06.800 --> 00:16:11.350 align:start position:0% +to use some +virtual<00:16:07.279> whiteboard<00:16:07.759> did<00:16:08.000> in<00:16:08.079> this<00:16:08.320> course + +00:16:11.350 --> 00:16:11.360 align:start position:0% + + + +00:16:11.360 --> 00:16:15.670 align:start position:0% + +all<00:16:11.519> right<00:16:11.759> so + +00:16:15.670 --> 00:16:15.680 align:start position:0% + + + +00:16:15.680 --> 00:16:19.509 align:start position:0% + +in<00:16:15.839> order<00:16:16.240> to<00:16:17.360> do<00:16:17.600> that<00:16:18.160> i<00:16:18.320> need<00:16:18.480> to + +00:16:19.509 --> 00:16:19.519 align:start position:0% +in order to do that i need to + + +00:16:19.519 --> 00:16:33.140 align:start position:0% +in order to do that i need to +connect<00:16:19.920> my<00:16:20.839> tablet + +00:16:33.140 --> 00:16:33.150 align:start position:0% + + + +00:16:33.150 --> 00:16:37.590 align:start position:0% + +[Music] + +00:16:37.590 --> 00:16:37.600 align:start position:0% + + + +00:16:37.600 --> 00:17:16.829 align:start position:0% + +okay<00:16:38.000> so<00:16:38.399> we<00:16:39.600> have<00:16:48.839> some + +00:17:16.829 --> 00:17:16.839 align:start position:0% + + + +00:17:16.839 --> 00:17:44.830 align:start position:0% + +this + +00:17:44.830 --> 00:17:44.840 align:start position:0% + + + +00:17:44.840 --> 00:17:57.590 align:start position:0% + +foreign + +00:17:57.590 --> 00:17:57.600 align:start position:0% + + + +00:17:57.600 --> 00:18:01.350 align:start position:0% + +okay<00:17:58.160> do<00:17:58.320> you<00:17:58.400> guys<00:17:58.720> see<00:17:58.960> my<00:17:59.360> uh<00:18:00.240> my<00:18:00.480> note<00:18:00.880> and + +00:18:01.350 --> 00:18:01.360 align:start position:0% +okay do you guys see my uh my note and + + +00:18:01.360 --> 00:18:04.630 align:start position:0% +okay do you guys see my uh my note and +do<00:18:01.440> you<00:18:01.600> hear<00:18:01.840> me<00:18:02.000> well<00:18:03.520> yes + +00:18:04.630 --> 00:18:04.640 align:start position:0% +do you hear me well yes + + +00:18:04.640 --> 00:18:08.390 align:start position:0% +do you hear me well yes +okay<00:18:05.039> fantastic<00:18:06.640> so<00:18:07.280> this<00:18:07.520> part<00:18:07.760> was<00:18:08.000> the + +00:18:08.390 --> 00:18:08.400 align:start position:0% +okay fantastic so this part was the + + +00:18:08.400 --> 00:18:10.390 align:start position:0% +okay fantastic so this part was the +hardest<00:18:08.799> part<00:18:09.120> so<00:18:09.280> i<00:18:09.360> wasn't<00:18:09.679> sure<00:18:10.000> if<00:18:10.160> it<00:18:10.320> is + +00:18:10.390 --> 00:18:10.400 align:start position:0% +hardest part so i wasn't sure if it is + + +00:18:10.400 --> 00:18:11.909 align:start position:0% +hardest part so i wasn't sure if it is +going<00:18:10.480> to<00:18:10.640> work<00:18:10.880> out<00:18:11.039> or<00:18:11.200> not + +00:18:11.909 --> 00:18:11.919 align:start position:0% +going to work out or not + + +00:18:11.919 --> 00:18:16.549 align:start position:0% +going to work out or not +but<00:18:12.160> it<00:18:12.240> seems<00:18:12.480> it<00:18:12.720> is<00:18:12.840> working<00:18:14.480> all<00:18:14.720> right + +00:18:16.549 --> 00:18:16.559 align:start position:0% +but it seems it is working all right + + +00:18:16.559 --> 00:18:20.070 align:start position:0% +but it seems it is working all right +so<00:18:17.039> um<00:18:17.760> so<00:18:18.000> today<00:18:18.640> we<00:18:18.799> are<00:18:18.960> going<00:18:19.280> to + +00:18:20.070 --> 00:18:20.080 align:start position:0% +so um so today we are going to + + +00:18:20.080 --> 00:18:23.110 align:start position:0% +so um so today we are going to +uh<00:18:20.640> talk<00:18:20.960> about<00:18:21.360> basic<00:18:21.919> deep<00:18:22.240> learning + +00:18:23.110 --> 00:18:23.120 align:start position:0% +uh talk about basic deep learning + + +00:18:23.120 --> 00:18:26.070 align:start position:0% +uh talk about basic deep learning +models<00:18:23.520> and<00:18:23.679> optimization<00:18:24.320> problems<00:18:25.440> as<00:18:25.840> i + +00:18:26.070 --> 00:18:26.080 align:start position:0% +models and optimization problems as i + + +00:18:26.080 --> 00:18:26.950 align:start position:0% +models and optimization problems as i +mentioned<00:18:26.559> i + +00:18:26.950 --> 00:18:26.960 align:start position:0% +mentioned i + + +00:18:26.960 --> 00:18:29.430 align:start position:0% +mentioned i +expect<00:18:27.440> you<00:18:27.600> to<00:18:28.080> already<00:18:28.559> be<00:18:28.799> familiar<00:18:29.280> with + +00:18:29.430 --> 00:18:29.440 align:start position:0% +expect you to already be familiar with + + +00:18:29.440 --> 00:18:30.230 align:start position:0% +expect you to already be familiar with +most<00:18:29.679> of<00:18:29.840> these + +00:18:30.230 --> 00:18:30.240 align:start position:0% +most of these + + +00:18:30.240 --> 00:18:33.510 align:start position:0% +most of these +uh<00:18:30.799> topics<00:18:31.840> uh<00:18:32.240> so<00:18:32.400> i'm<00:18:32.559> gonna<00:18:32.880> go<00:18:33.200> very + +00:18:33.510 --> 00:18:33.520 align:start position:0% +uh topics uh so i'm gonna go very + + +00:18:33.520 --> 00:18:34.230 align:start position:0% +uh topics uh so i'm gonna go very +quickly + +00:18:34.230 --> 00:18:34.240 align:start position:0% +quickly + + +00:18:34.240 --> 00:18:36.230 align:start position:0% +quickly +so<00:18:34.400> if<00:18:34.559> you<00:18:34.720> think<00:18:34.960> that<00:18:35.280> you<00:18:35.440> know<00:18:35.600> it<00:18:35.760> is<00:18:36.000> hard + +00:18:36.230 --> 00:18:36.240 align:start position:0% +so if you think that you know it is hard + + +00:18:36.240 --> 00:18:37.590 align:start position:0% +so if you think that you know it is hard +for<00:18:36.400> you<00:18:36.559> to<00:18:36.799> follow<00:18:37.120> it + +00:18:37.590 --> 00:18:37.600 align:start position:0% +for you to follow it + + +00:18:37.600 --> 00:18:40.390 align:start position:0% +for you to follow it +you<00:18:37.760> know<00:18:38.000> contact<00:18:38.400> tas<00:18:38.799> and<00:18:38.960> myself<00:18:39.600> to<00:18:40.000> see + +00:18:40.390 --> 00:18:40.400 align:start position:0% +you know contact tas and myself to see + + +00:18:40.400 --> 00:18:40.789 align:start position:0% +you know contact tas and myself to see +if + +00:18:40.789 --> 00:18:40.799 align:start position:0% +if + + +00:18:40.799 --> 00:18:44.150 align:start position:0% +if +this<00:18:41.039> course<00:18:41.280> is<00:18:41.440> a<00:18:41.520> good<00:18:41.679> fit<00:18:42.000> for<00:18:42.160> you + +00:18:44.150 --> 00:18:44.160 align:start position:0% +this course is a good fit for you + + +00:18:44.160 --> 00:18:48.390 align:start position:0% +this course is a good fit for you +all<00:18:44.240> right<00:18:44.480> so<00:18:44.720> this<00:18:44.960> is<00:18:45.280> cms<00:18:47.080> c828 + +00:18:48.390 --> 00:18:48.400 align:start position:0% +all right so this is cms c828 + + +00:18:48.400 --> 00:18:54.830 align:start position:0% +all right so this is cms c828 +w<00:18:51.039> foundations + +00:18:54.830 --> 00:18:54.840 align:start position:0% + + + +00:18:54.840 --> 00:19:00.789 align:start position:0% + +of<00:18:56.559> deep<00:18:56.799> learning + +00:19:00.789 --> 00:19:00.799 align:start position:0% + + + +00:19:00.799 --> 00:19:05.909 align:start position:0% + +okay<00:19:01.120> so<00:19:02.160> let's<00:19:02.559> start<00:19:02.799> with<00:19:03.360> with<00:19:03.600> a<00:19:03.760> note + +00:19:05.909 --> 00:19:05.919 align:start position:0% +okay so let's start with with a note + + +00:19:05.919 --> 00:19:10.150 align:start position:0% +okay so let's start with with a note +or<00:19:06.400> perhaps<00:19:06.880> a<00:19:06.960> remark<00:19:08.480> so<00:19:08.720> deep<00:19:09.039> learning<00:19:09.760> is + +00:19:10.150 --> 00:19:10.160 align:start position:0% +or perhaps a remark so deep learning is + + +00:19:10.160 --> 00:19:13.750 align:start position:0% +or perhaps a remark so deep learning is +a<00:19:10.960> very<00:19:11.440> fast + +00:19:13.750 --> 00:19:13.760 align:start position:0% +a very fast + + +00:19:13.760 --> 00:19:21.110 align:start position:0% +a very fast +moving<00:19:14.400> area + +00:19:21.110 --> 00:19:21.120 align:start position:0% + + + +00:19:21.120 --> 00:19:24.630 align:start position:0% + +probably<00:19:21.440> you<00:19:21.600> have<00:19:21.840> noticed<00:19:22.400> most<00:19:22.720> of<00:19:23.039> or + +00:19:24.630 --> 00:19:24.640 align:start position:0% +probably you have noticed most of or + + +00:19:24.640 --> 00:19:27.270 align:start position:0% +probably you have noticed most of or +most<00:19:24.960> of<00:19:25.039> the<00:19:25.280> papers<00:19:25.840> and<00:19:26.080> the<00:19:26.480> topics<00:19:27.039> that + +00:19:27.270 --> 00:19:27.280 align:start position:0% +most of the papers and the topics that + + +00:19:27.280 --> 00:19:28.789 align:start position:0% +most of the papers and the topics that +we'll<00:19:27.520> be<00:19:27.760> covering + +00:19:28.789 --> 00:19:28.799 align:start position:0% +we'll be covering + + +00:19:28.799 --> 00:19:32.150 align:start position:0% +we'll be covering +in<00:19:28.960> this<00:19:29.200> course + +00:19:32.150 --> 00:19:32.160 align:start position:0% + + + +00:19:32.160 --> 00:19:35.510 align:start position:0% + +of<00:19:32.320> this<00:19:32.559> course<00:19:34.000> they<00:19:34.320> are + +00:19:35.510 --> 00:19:35.520 align:start position:0% +of this course they are + + +00:19:35.520 --> 00:19:44.870 align:start position:0% +of this course they are +very<00:19:37.200> recent<00:19:38.160> topics<00:19:38.720> and<00:19:38.880> papers + +00:19:44.870 --> 00:19:44.880 align:start position:0% + + + +00:19:44.880 --> 00:19:47.750 align:start position:0% + +i<00:19:45.039> would<00:19:45.200> say<00:19:45.520> perhaps<00:19:46.080> within<00:19:47.039> most<00:19:47.360> of<00:19:47.520> them + +00:19:47.750 --> 00:19:47.760 align:start position:0% +i would say perhaps within most of them + + +00:19:47.760 --> 00:19:48.710 align:start position:0% +i would say perhaps within most of them +within + +00:19:48.710 --> 00:19:48.720 align:start position:0% +within + + +00:19:48.720 --> 00:19:52.070 align:start position:0% +within +the<00:19:48.880> last<00:19:49.200> five<00:19:49.520> years<00:19:50.880> so<00:19:51.520> one + +00:19:52.070 --> 00:19:52.080 align:start position:0% +the last five years so one + + +00:19:52.080 --> 00:19:55.990 align:start position:0% +the last five years so one +um<00:19:53.280> interesting<00:19:53.840> question<00:19:54.960> is + +00:19:55.990 --> 00:19:56.000 align:start position:0% +um interesting question is + + +00:19:56.000 --> 00:19:59.350 align:start position:0% +um interesting question is +to<00:19:56.240> see<00:19:56.880> uh<00:19:57.440> if<00:19:57.600> i<00:19:57.760> teach<00:19:58.080> this<00:19:58.320> course + +00:19:59.350 --> 00:19:59.360 align:start position:0% +to see uh if i teach this course + + +00:19:59.360 --> 00:20:02.870 align:start position:0% +to see uh if i teach this course +in<00:19:59.520> the<00:20:00.559> you<00:20:00.720> know<00:20:01.120> in<00:20:01.280> the<00:20:01.440> next<00:20:01.760> uh<00:20:02.320> in + +00:20:02.870 --> 00:20:02.880 align:start position:0% +in the you know in the next uh in + + +00:20:02.880 --> 00:20:06.789 align:start position:0% +in the you know in the next uh in +five<00:20:03.120> years<00:20:03.679> how<00:20:04.840> different + +00:20:06.789 --> 00:20:06.799 align:start position:0% +five years how different + + +00:20:06.799 --> 00:20:11.590 align:start position:0% +five years how different +different<00:20:07.360> these<00:20:07.679> topics + +00:20:11.590 --> 00:20:11.600 align:start position:0% + + + +00:20:11.600 --> 00:20:17.350 align:start position:0% + +and<00:20:11.840> papers<00:20:13.840> will<00:20:14.080> be + +00:20:17.350 --> 00:20:17.360 align:start position:0% + + + +00:20:17.360 --> 00:20:21.590 align:start position:0% + +let's<00:20:17.600> say<00:20:17.919> in<00:20:19.360> five<00:20:19.600> years + +00:20:21.590 --> 00:20:21.600 align:start position:0% +let's say in five years + + +00:20:21.600 --> 00:20:23.830 align:start position:0% +let's say in five years +so<00:20:21.760> it<00:20:21.919> is<00:20:22.080> a<00:20:22.159> very<00:20:22.559> so<00:20:23.120> the<00:20:23.280> reason<00:20:23.600> i'm + +00:20:23.830 --> 00:20:23.840 align:start position:0% +so it is a very so the reason i'm + + +00:20:23.840 --> 00:20:25.669 align:start position:0% +so it is a very so the reason i'm +mentioning<00:20:24.320> and<00:20:24.480> i'm<00:20:24.559> highlighting<00:20:25.200> this<00:20:25.520> is + +00:20:25.669 --> 00:20:25.679 align:start position:0% +mentioning and i'm highlighting this is + + +00:20:25.679 --> 00:20:27.750 align:start position:0% +mentioning and i'm highlighting this is +that<00:20:25.840> you<00:20:26.080> should<00:20:26.320> not + +00:20:27.750 --> 00:20:27.760 align:start position:0% +that you should not + + +00:20:27.760 --> 00:20:31.590 align:start position:0% +that you should not +take<00:20:28.400> any<00:20:28.799> of<00:20:29.039> these<00:20:30.000> concepts<00:20:30.640> and<00:20:30.799> topics + +00:20:31.590 --> 00:20:31.600 align:start position:0% +take any of these concepts and topics + + +00:20:31.600 --> 00:20:34.870 align:start position:0% +take any of these concepts and topics +as<00:20:32.840> uh<00:20:33.840> basically<00:20:34.400> a + +00:20:34.870 --> 00:20:34.880 align:start position:0% +as uh basically a + + +00:20:34.880 --> 00:20:37.990 align:start position:0% +as uh basically a +definite<00:20:35.600> uh<00:20:36.799> you<00:20:36.960> know<00:20:37.120> concept<00:20:37.679> that + +00:20:37.990 --> 00:20:38.000 align:start position:0% +definite uh you know concept that + + +00:20:38.000 --> 00:20:40.549 align:start position:0% +definite uh you know concept that +explains<00:20:38.640> deep<00:20:38.880> learning<00:20:39.280> so<00:20:39.440> it<00:20:39.600> is<00:20:39.760> a<00:20:39.840> very + +00:20:40.549 --> 00:20:40.559 align:start position:0% +explains deep learning so it is a very + + +00:20:40.559 --> 00:20:42.230 align:start position:0% +explains deep learning so it is a very +fast<00:20:40.960> moving<00:20:41.520> area + +00:20:42.230 --> 00:20:42.240 align:start position:0% +fast moving area + + +00:20:42.240 --> 00:20:45.830 align:start position:0% +fast moving area +every<00:20:43.039> year<00:20:43.440> we<00:20:43.600> are<00:20:43.840> learning<00:20:44.240> more<00:20:44.559> and<00:20:44.720> more + +00:20:45.830 --> 00:20:45.840 align:start position:0% +every year we are learning more and more + + +00:20:45.840 --> 00:20:48.950 align:start position:0% +every year we are learning more and more +why<00:20:46.159> deep<00:20:46.400> learning<00:20:46.880> is<00:20:47.120> powerful<00:20:48.159> and<00:20:48.640> also + +00:20:48.950 --> 00:20:48.960 align:start position:0% +why deep learning is powerful and also + + +00:20:48.960 --> 00:20:50.549 align:start position:0% +why deep learning is powerful and also +what<00:20:49.200> are<00:20:49.280> the<00:20:49.520> limitations<00:20:50.159> of<00:20:50.320> deep + +00:20:50.549 --> 00:20:50.559 align:start position:0% +what are the limitations of deep + + +00:20:50.559 --> 00:20:51.270 align:start position:0% +what are the limitations of deep +learning + +00:20:51.270 --> 00:20:51.280 align:start position:0% +learning + + +00:20:51.280 --> 00:20:54.390 align:start position:0% +learning +so<00:20:51.520> this<00:20:51.840> course<00:20:52.240> is<00:20:52.640> perhaps<00:20:53.200> you<00:20:53.280> know<00:20:53.679> a + +00:20:54.390 --> 00:20:54.400 align:start position:0% +so this course is perhaps you know a + + +00:20:54.400 --> 00:20:56.789 align:start position:0% +so this course is perhaps you know a +good<00:20:55.039> starting<00:20:55.600> point<00:20:55.919> to<00:20:56.240> think<00:20:56.480> a<00:20:56.559> little + +00:20:56.789 --> 00:20:56.799 align:start position:0% +good starting point to think a little + + +00:20:56.799 --> 00:20:58.630 align:start position:0% +good starting point to think a little +bit<00:20:57.039> about<00:20:57.360> different<00:20:57.760> aspects<00:20:58.240> of<00:20:58.400> deep + +00:20:58.630 --> 00:20:58.640 align:start position:0% +bit about different aspects of deep + + +00:20:58.640 --> 00:20:59.990 align:start position:0% +bit about different aspects of deep +learning + +00:20:59.990 --> 00:21:00.000 align:start position:0% +learning + + +00:21:00.000 --> 00:21:02.950 align:start position:0% +learning +that<00:21:00.240> potentially<00:21:01.120> can<00:21:01.600> evolve<00:21:02.240> and<00:21:02.480> change + +00:21:02.950 --> 00:21:02.960 align:start position:0% +that potentially can evolve and change + + +00:21:02.960 --> 00:21:03.750 align:start position:0% +that potentially can evolve and change +within<00:21:03.280> the + +00:21:03.750 --> 00:21:03.760 align:start position:0% +within the + + +00:21:03.760 --> 00:21:07.669 align:start position:0% +within the +next<00:21:04.000> decade<00:21:04.480> as<00:21:04.640> well<00:21:06.480> okay<00:21:06.880> so + +00:21:07.669 --> 00:21:07.679 align:start position:0% +next decade as well okay so + + +00:21:07.679 --> 00:21:14.470 align:start position:0% +next decade as well okay so +today's<00:21:08.159> agenda<00:21:08.960> is + +00:21:14.470 --> 00:21:14.480 align:start position:0% + + + +00:21:14.480 --> 00:21:18.549 align:start position:0% + +so<00:21:14.720> we<00:21:14.880> are<00:21:14.960> going<00:21:15.280> to<00:21:16.159> cover<00:21:16.720> basic + +00:21:18.549 --> 00:21:18.559 align:start position:0% +so we are going to cover basic + + +00:21:18.559 --> 00:21:23.350 align:start position:0% +so we are going to cover basic +deep<00:21:18.880> learning + +00:21:23.350 --> 00:21:23.360 align:start position:0% + + + +00:21:23.360 --> 00:21:29.669 align:start position:0% + +models<00:21:24.000> and<00:21:24.480> architectures + +00:21:29.669 --> 00:21:29.679 align:start position:0% + + + +00:21:29.679 --> 00:21:32.470 align:start position:0% + +we'll<00:21:30.080> start<00:21:30.320> with<00:21:30.559> the<00:21:30.799> erm<00:21:31.600> empirical<00:21:32.240> risk + +00:21:32.470 --> 00:21:32.480 align:start position:0% +we'll start with the erm empirical risk + + +00:21:32.480 --> 00:21:33.909 align:start position:0% +we'll start with the erm empirical risk +minimization + +00:21:33.909 --> 00:21:33.919 align:start position:0% +minimization + + +00:21:33.919 --> 00:21:36.950 align:start position:0% +minimization +we'll<00:21:34.320> then<00:21:34.640> talk<00:21:34.880> about<00:21:35.200> the<00:21:35.360> loss<00:21:35.760> functions + +00:21:36.950 --> 00:21:36.960 align:start position:0% +we'll then talk about the loss functions + + +00:21:36.960 --> 00:21:39.990 align:start position:0% +we'll then talk about the loss functions +that<00:21:37.200> we<00:21:37.440> use<00:21:38.240> in<00:21:38.720> erm + +00:21:39.990 --> 00:21:40.000 align:start position:0% +that we use in erm + + +00:21:40.000 --> 00:21:43.029 align:start position:0% +that we use in erm +then<00:21:40.159> we'll<00:21:40.480> talk<00:21:40.720> about<00:21:40.960> the<00:21:41.280> models<00:21:42.640> uh + +00:21:43.029 --> 00:21:43.039 align:start position:0% +then we'll talk about the models uh + + +00:21:43.039 --> 00:21:46.230 align:start position:0% +then we'll talk about the models uh +that<00:21:43.919> we<00:21:44.400> use<00:21:44.960> in<00:21:45.120> order<00:21:45.520> to + +00:21:46.230 --> 00:21:46.240 align:start position:0% +that we use in order to + + +00:21:46.240 --> 00:21:47.909 align:start position:0% +that we use in order to +have<00:21:46.480> a<00:21:46.640> mapping<00:21:47.039> from<00:21:47.360> inputs<00:21:47.679> to<00:21:47.760> the + +00:21:47.909 --> 00:21:47.919 align:start position:0% +have a mapping from inputs to the + + +00:21:47.919 --> 00:21:50.310 align:start position:0% +have a mapping from inputs to the +outputs<00:21:48.640> and<00:21:48.880> finally<00:21:49.360> we'll<00:21:49.600> look<00:21:49.840> into<00:21:50.159> the + +00:21:50.310 --> 00:21:50.320 align:start position:0% +outputs and finally we'll look into the + + +00:21:50.320 --> 00:21:52.230 align:start position:0% +outputs and finally we'll look into the +optimization<00:21:51.120> problems<00:21:51.679> and + +00:21:52.230 --> 00:21:52.240 align:start position:0% +optimization problems and + + +00:21:52.240 --> 00:21:54.789 align:start position:0% +optimization problems and +how<00:21:52.559> we<00:21:53.280> solve<00:21:53.600> these<00:21:53.840> optimizations<00:21:54.640> in + +00:21:54.789 --> 00:21:54.799 align:start position:0% +how we solve these optimizations in + + +00:21:54.799 --> 00:21:57.350 align:start position:0% +how we solve these optimizations in +practice + +00:21:57.350 --> 00:21:57.360 align:start position:0% + + + +00:21:57.360 --> 00:21:59.110 align:start position:0% + +all<00:21:57.440> right<00:21:57.600> so<00:21:57.919> we<00:21:58.159> start<00:21:58.320> with<00:21:58.480> supervised + +00:21:59.110 --> 00:21:59.120 align:start position:0% +all right so we start with supervised + + +00:21:59.120 --> 00:22:05.669 align:start position:0% +all right so we start with supervised +learning<00:21:59.520> setup + +00:22:05.669 --> 00:22:05.679 align:start position:0% + + + +00:22:05.679 --> 00:22:09.190 align:start position:0% + +in<00:22:05.919> supervised<00:22:06.559> learning<00:22:07.360> i'm<00:22:07.919> given<00:22:08.559> a<00:22:08.799> data + +00:22:09.190 --> 00:22:09.200 align:start position:0% +in supervised learning i'm given a data + + +00:22:09.200 --> 00:22:09.750 align:start position:0% +in supervised learning i'm given a data +set + +00:22:09.750 --> 00:22:09.760 align:start position:0% +set + + +00:22:09.760 --> 00:22:15.029 align:start position:0% +set +like<00:22:10.000> this<00:22:10.480> xi's<00:22:11.440> and<00:22:11.840> y<00:22:12.240> i's + +00:22:15.029 --> 00:22:15.039 align:start position:0% + + + +00:22:15.039 --> 00:22:19.190 align:start position:0% + +i<00:22:15.280> have<00:22:15.520> n<00:22:15.840> samples<00:22:16.880> x's<00:22:17.600> are<00:22:18.000> my + +00:22:19.190 --> 00:22:19.200 align:start position:0% +i have n samples x's are my + + +00:22:19.200 --> 00:22:30.230 align:start position:0% +i have n samples x's are my +features<00:22:21.360> and<00:22:22.000> y's<00:22:22.400> are<00:22:22.559> my<00:22:24.840> labels + +00:22:30.230 --> 00:22:30.240 align:start position:0% + + + +00:22:30.240 --> 00:22:31.990 align:start position:0% + +all<00:22:30.320> right<00:22:30.480> so<00:22:30.640> what<00:22:30.880> are<00:22:30.960> the<00:22:31.120> dimensions<00:22:31.840> you + +00:22:31.990 --> 00:22:32.000 align:start position:0% +all right so what are the dimensions you + + +00:22:32.000 --> 00:22:33.590 align:start position:0% +all right so what are the dimensions you +usually<00:22:32.480> assume<00:22:32.960> x<00:22:33.200> is + +00:22:33.590 --> 00:22:33.600 align:start position:0% +usually assume x is + + +00:22:33.600 --> 00:22:35.669 align:start position:0% +usually assume x is +in<00:22:33.760> a<00:22:34.000> high<00:22:34.159> dimensional<00:22:34.799> space<00:22:35.120> let's<00:22:35.360> say<00:22:35.520> in + +00:22:35.669 --> 00:22:35.679 align:start position:0% +in a high dimensional space let's say in + + +00:22:35.679 --> 00:22:37.590 align:start position:0% +in a high dimensional space let's say in +a<00:22:35.760> two<00:22:35.919> dimensional<00:22:36.559> space + +00:22:37.590 --> 00:22:37.600 align:start position:0% +a two dimensional space + + +00:22:37.600 --> 00:22:41.190 align:start position:0% +a two dimensional space +uh<00:22:37.919> why<00:22:38.240> is<00:22:38.400> there<00:22:39.520> can<00:22:39.679> be<00:22:40.240> a + +00:22:41.190 --> 00:22:41.200 align:start position:0% +uh why is there can be a + + +00:22:41.200 --> 00:22:44.470 align:start position:0% +uh why is there can be a +lower<00:22:41.520> dimension<00:22:42.720> uh<00:22:43.520> e<00:22:44.080> here + +00:22:44.470 --> 00:22:44.480 align:start position:0% +lower dimension uh e here + + +00:22:44.480 --> 00:22:46.870 align:start position:0% +lower dimension uh e here +assume<00:22:45.200> y's<00:22:45.520> are<00:22:45.679> in<00:22:45.840> one<00:22:46.080> dimension<00:22:46.559> but<00:22:46.720> you + +00:22:46.870 --> 00:22:46.880 align:start position:0% +assume y's are in one dimension but you + + +00:22:46.880 --> 00:22:48.230 align:start position:0% +assume y's are in one dimension but you +can<00:22:47.039> have<00:22:47.440> uh + +00:22:48.230 --> 00:22:48.240 align:start position:0% +can have uh + + +00:22:48.240 --> 00:22:51.190 align:start position:0% +can have uh +c<00:22:48.559> classes<00:22:49.520> if<00:22:49.679> you<00:22:49.840> do<00:22:50.000> like<00:22:50.240> this<00:22:50.640> you<00:22:50.720> know<00:22:50.960> c + +00:22:51.190 --> 00:22:51.200 align:start position:0% +c classes if you do like this you know c + + +00:22:51.200 --> 00:22:52.390 align:start position:0% +c classes if you do like this you know c +class + +00:22:52.390 --> 00:22:52.400 align:start position:0% +class + + +00:22:52.400 --> 00:22:54.230 align:start position:0% +class +classification<00:22:53.200> problem<00:22:53.520> so<00:22:53.760> you<00:22:53.840> can<00:22:54.000> also + +00:22:54.230 --> 00:22:54.240 align:start position:0% +classification problem so you can also + + +00:22:54.240 --> 00:22:56.710 align:start position:0% +classification problem so you can also +assume<00:22:54.640> y<00:22:55.120> is<00:22:55.280> in<00:22:55.440> cc<00:22:55.840> is<00:22:55.919> usually<00:22:56.320> like<00:22:56.640> you + +00:22:56.710 --> 00:22:56.720 align:start position:0% +assume y is in cc is usually like you + + +00:22:56.720 --> 00:22:58.630 align:start position:0% +assume y is in cc is usually like you +know<00:22:56.960> not<00:22:57.200> as<00:22:57.440> large<00:22:57.760> as<00:22:58.080> uh + +00:22:58.630 --> 00:22:58.640 align:start position:0% +know not as large as uh + + +00:22:58.640 --> 00:23:01.430 align:start position:0% +know not as large as uh +d<00:22:58.960> in<00:22:59.120> general<00:23:00.000> but<00:23:00.240> for<00:23:00.400> the<00:23:00.720> most<00:23:01.039> part<00:23:01.280> of + +00:23:01.430 --> 00:23:01.440 align:start position:0% +d in general but for the most part of + + +00:23:01.440 --> 00:23:03.190 align:start position:0% +d in general but for the most part of +today's<00:23:01.840> lectures<00:23:02.240> just<00:23:02.480> to<00:23:02.640> introduce<00:23:03.039> some + +00:23:03.190 --> 00:23:03.200 align:start position:0% +today's lectures just to introduce some + + +00:23:03.200 --> 00:23:05.029 align:start position:0% +today's lectures just to introduce some +of<00:23:03.360> the<00:23:03.440> concepts<00:23:03.840> i'm<00:23:04.000> going<00:23:04.159> to<00:23:04.320> assume + +00:23:05.029 --> 00:23:05.039 align:start position:0% +of the concepts i'm going to assume + + +00:23:05.039 --> 00:23:08.310 align:start position:0% +of the concepts i'm going to assume +y<00:23:05.520> is<00:23:05.919> in<00:23:06.320> reals<00:23:06.960> one<00:23:07.120> dimensional + +00:23:08.310 --> 00:23:08.320 align:start position:0% +y is in reals one dimensional + + +00:23:08.320 --> 00:23:11.750 align:start position:0% +y is in reals one dimensional +space<00:23:10.159> so<00:23:10.400> we<00:23:10.559> have<00:23:10.880> uh + +00:23:11.750 --> 00:23:11.760 align:start position:0% +space so we have uh + + +00:23:11.760 --> 00:23:14.789 align:start position:0% +space so we have uh +basically<00:23:12.320> two<00:23:12.559> types<00:23:12.880> of<00:23:13.120> problems<00:23:14.400> within + +00:23:14.789 --> 00:23:14.799 align:start position:0% +basically two types of problems within + + +00:23:14.799 --> 00:23:15.350 align:start position:0% +basically two types of problems within +this + +00:23:15.350 --> 00:23:15.360 align:start position:0% +this + + +00:23:15.360 --> 00:23:18.549 align:start position:0% +this +we<00:23:15.520> can<00:23:15.760> have<00:23:16.000> regression<00:23:16.640> problem + +00:23:18.549 --> 00:23:18.559 align:start position:0% +we can have regression problem + + +00:23:18.559 --> 00:23:23.909 align:start position:0% +we can have regression problem +and<00:23:18.720> we<00:23:18.880> can<00:23:19.120> have<00:23:19.840> classification<00:23:20.840> problem + +00:23:23.909 --> 00:23:23.919 align:start position:0% + + + +00:23:23.919 --> 00:23:27.510 align:start position:0% + +in<00:23:24.000> the<00:23:24.159> regression<00:23:24.799> problem<00:23:26.080> we<00:23:26.320> assume + +00:23:27.510 --> 00:23:27.520 align:start position:0% +in the regression problem we assume + + +00:23:27.520 --> 00:23:32.149 align:start position:0% +in the regression problem we assume +the<00:23:28.640> labels<00:23:29.120> that<00:23:29.360> we<00:23:29.520> have<00:23:29.760> there<00:23:30.080> in<00:23:30.240> reals + +00:23:32.149 --> 00:23:32.159 align:start position:0% +the labels that we have there in reals + + +00:23:32.159 --> 00:23:34.070 align:start position:0% +the labels that we have there in reals +for<00:23:32.320> instance<00:23:32.880> if<00:23:33.039> you<00:23:33.120> want<00:23:33.280> to<00:23:33.600> predict + +00:23:34.070 --> 00:23:34.080 align:start position:0% +for instance if you want to predict + + +00:23:34.080 --> 00:23:35.510 align:start position:0% +for instance if you want to predict +housing<00:23:34.559> prices<00:23:35.039> that<00:23:35.200> would<00:23:35.360> be<00:23:35.440> a + +00:23:35.510 --> 00:23:35.520 align:start position:0% +housing prices that would be a + + +00:23:35.520 --> 00:23:36.070 align:start position:0% +housing prices that would be a +regression + +00:23:36.070 --> 00:23:36.080 align:start position:0% +regression + + +00:23:36.080 --> 00:23:39.190 align:start position:0% +regression +problem<00:23:37.039> and<00:23:37.280> the<00:23:37.520> classification<00:23:38.400> problem + +00:23:39.190 --> 00:23:39.200 align:start position:0% +problem and the classification problem + + +00:23:39.200 --> 00:23:41.909 align:start position:0% +problem and the classification problem +we<00:23:39.520> assume<00:23:40.000> the<00:23:40.240> labels<00:23:40.799> are<00:23:41.120> in<00:23:41.279> a<00:23:41.520> finite + +00:23:41.909 --> 00:23:41.919 align:start position:0% +we assume the labels are in a finite + + +00:23:41.919 --> 00:23:43.430 align:start position:0% +we assume the labels are in a finite +field<00:23:42.320> for<00:23:42.480> instance + +00:23:43.430 --> 00:23:43.440 align:start position:0% +field for instance + + +00:23:43.440 --> 00:23:45.430 align:start position:0% +field for instance +in<00:23:43.600> a<00:23:43.679> binary<00:23:44.159> classification<00:23:44.880> problem<00:23:45.200> let's + +00:23:45.430 --> 00:23:45.440 align:start position:0% +in a binary classification problem let's + + +00:23:45.440 --> 00:23:47.110 align:start position:0% +in a binary classification problem let's +assume<00:23:45.760> the<00:23:45.919> labels<00:23:46.320> are<00:23:46.400> either + +00:23:47.110 --> 00:23:47.120 align:start position:0% +assume the labels are either + + +00:23:47.120 --> 00:23:49.830 align:start position:0% +assume the labels are either +minus<00:23:47.520> one<00:23:47.760> or<00:23:47.919> one<00:23:48.559> or<00:23:48.880> in<00:23:49.039> a<00:23:49.200> multi-label + +00:23:49.830 --> 00:23:49.840 align:start position:0% +minus one or one or in a multi-label + + +00:23:49.840 --> 00:23:52.070 align:start position:0% +minus one or one or in a multi-label +classification<00:23:50.640> problem<00:23:51.039> we<00:23:51.200> assume + +00:23:52.070 --> 00:23:52.080 align:start position:0% +classification problem we assume + + +00:23:52.080 --> 00:23:54.470 align:start position:0% +classification problem we assume +the<00:23:52.240> labels<00:23:52.720> are<00:23:53.360> let's<00:23:53.520> say<00:23:53.760> from<00:23:54.000> one<00:23:54.240> to + +00:23:54.470 --> 00:23:54.480 align:start position:0% +the labels are let's say from one to + + +00:23:54.480 --> 00:23:55.110 align:start position:0% +the labels are let's say from one to +three<00:23:54.799> the + +00:23:55.110 --> 00:23:55.120 align:start position:0% +three the + + +00:23:55.120 --> 00:23:59.510 align:start position:0% +three the +integer<00:23:55.600> numbers<00:23:57.919> all<00:23:58.080> right<00:23:58.240> so<00:23:58.400> what<00:23:58.640> is<00:23:58.799> the + +00:23:59.510 --> 00:23:59.520 align:start position:0% +integer numbers all right so what is the + + +00:23:59.520 --> 00:24:02.870 align:start position:0% +integer numbers all right so what is the +goal<00:23:59.840> here<00:24:00.080> so<00:24:00.240> our<00:24:00.400> goal<00:24:00.720> is<00:24:01.120> very<00:24:01.679> clear + +00:24:02.870 --> 00:24:02.880 align:start position:0% +goal here so our goal is very clear + + +00:24:02.880 --> 00:24:05.909 align:start position:0% +goal here so our goal is very clear +in<00:24:03.440> supervised<00:24:04.080> learning<00:24:05.200> so<00:24:05.440> we'll<00:24:05.679> talk + +00:24:05.909 --> 00:24:05.919 align:start position:0% +in supervised learning so we'll talk + + +00:24:05.919 --> 00:24:07.510 align:start position:0% +in supervised learning so we'll talk +about<00:24:06.159> reinforcement<00:24:06.799> learning<00:24:07.279> and + +00:24:07.510 --> 00:24:07.520 align:start position:0% +about reinforcement learning and + + +00:24:07.520 --> 00:24:09.350 align:start position:0% +about reinforcement learning and +unsupervised<00:24:08.320> learning<00:24:08.799> and<00:24:08.960> in<00:24:09.039> fact + +00:24:09.350 --> 00:24:09.360 align:start position:0% +unsupervised learning and in fact + + +00:24:09.360 --> 00:24:11.190 align:start position:0% +unsupervised learning and in fact +self-supervised<00:24:10.159> learning<00:24:10.640> later<00:24:10.960> on<00:24:11.120> in + +00:24:11.190 --> 00:24:11.200 align:start position:0% +self-supervised learning later on in + + +00:24:11.200 --> 00:24:11.830 align:start position:0% +self-supervised learning later on in +this + +00:24:11.830 --> 00:24:11.840 align:start position:0% +this + + +00:24:11.840 --> 00:24:13.430 align:start position:0% +this +in<00:24:12.000> this<00:24:12.240> course<00:24:12.559> but<00:24:12.720> let's<00:24:13.039> start<00:24:13.200> with<00:24:13.360> the + +00:24:13.430 --> 00:24:13.440 align:start position:0% +in this course but let's start with the + + +00:24:13.440 --> 00:24:15.909 align:start position:0% +in this course but let's start with the +supervised<00:24:14.000> learning<00:24:14.799> our<00:24:15.039> goal<00:24:15.360> is<00:24:15.520> to + +00:24:15.909 --> 00:24:15.919 align:start position:0% +supervised learning our goal is to + + +00:24:15.919 --> 00:24:20.230 align:start position:0% +supervised learning our goal is to +find<00:24:16.799> a<00:24:17.039> function<00:24:18.000> let's<00:24:18.240> say<00:24:18.559> fw + +00:24:20.230 --> 00:24:20.240 align:start position:0% +find a function let's say fw + + +00:24:20.240 --> 00:24:23.029 align:start position:0% +find a function let's say fw +from<00:24:20.559> my<00:24:20.960> input<00:24:21.440> space<00:24:22.000> d<00:24:22.159> dimensional<00:24:22.640> input + +00:24:23.029 --> 00:24:23.039 align:start position:0% +from my input space d dimensional input + + +00:24:23.039 --> 00:24:23.830 align:start position:0% +from my input space d dimensional input +space<00:24:23.360> to<00:24:23.520> my + +00:24:23.830 --> 00:24:23.840 align:start position:0% +space to my + + +00:24:23.840 --> 00:24:26.870 align:start position:0% +space to my +output<00:24:24.320> space<00:24:25.600> to<00:24:25.919> r + +00:24:26.870 --> 00:24:26.880 align:start position:0% +output space to r + + +00:24:26.880 --> 00:24:31.029 align:start position:0% +output space to r +such<00:24:27.200> that<00:24:27.520> if<00:24:27.840> w<00:24:28.400> of<00:24:28.720> x<00:24:28.960> i<00:24:29.520> is<00:24:29.840> a<00:24:30.080> good + +00:24:31.029 --> 00:24:31.039 align:start position:0% +such that if w of x i is a good + + +00:24:31.039 --> 00:24:33.990 align:start position:0% +such that if w of x i is a good +predictor<00:24:31.600> of<00:24:31.840> by<00:24:32.159> eye<00:24:32.640> for<00:24:32.960> all<00:24:33.200> of<00:24:33.440> my + +00:24:33.990 --> 00:24:34.000 align:start position:0% +predictor of by eye for all of my + + +00:24:34.000 --> 00:24:36.070 align:start position:0% +predictor of by eye for all of my +training<00:24:34.480> samples + +00:24:36.070 --> 00:24:36.080 align:start position:0% +training samples + + +00:24:36.080 --> 00:24:39.669 align:start position:0% +training samples +so<00:24:36.480> w<00:24:37.120> here<00:24:37.840> is + +00:24:39.669 --> 00:24:39.679 align:start position:0% +so w here is + + +00:24:39.679 --> 00:24:45.990 align:start position:0% +so w here is +representing<00:24:40.640> my<00:24:41.520> model<00:24:44.840> parameters + +00:24:45.990 --> 00:24:46.000 align:start position:0% +representing my model parameters + + +00:24:46.000 --> 00:24:48.070 align:start position:0% +representing my model parameters +right<00:24:46.159> so<00:24:46.480> that's<00:24:46.720> the<00:24:46.960> goal<00:24:47.200> so<00:24:47.279> i<00:24:47.440> want<00:24:47.600> to + +00:24:48.070 --> 00:24:48.080 align:start position:0% +right so that's the goal so i want to + + +00:24:48.080 --> 00:24:50.310 align:start position:0% +right so that's the goal so i want to +basically<00:24:48.720> find<00:24:48.960> this<00:24:49.279> function<00:24:49.679> mapping<00:24:50.000> my + +00:24:50.310 --> 00:24:50.320 align:start position:0% +basically find this function mapping my + + +00:24:50.320 --> 00:24:53.750 align:start position:0% +basically find this function mapping my +inputs<00:24:50.799> to<00:24:51.600> uh<00:24:52.000> outputs + +00:24:53.750 --> 00:24:53.760 align:start position:0% +inputs to uh outputs + + +00:24:53.760 --> 00:24:57.110 align:start position:0% +inputs to uh outputs +all<00:24:53.919> right<00:24:54.080> so<00:24:54.320> how<00:24:54.720> i<00:24:55.279> can<00:24:56.400> find<00:24:56.720> such<00:24:56.960> a + +00:24:57.110 --> 00:24:57.120 align:start position:0% +all right so how i can find such a + + +00:24:57.120 --> 00:24:57.909 align:start position:0% +all right so how i can find such a +function + +00:24:57.909 --> 00:24:57.919 align:start position:0% +function + + +00:24:57.919 --> 00:25:01.110 align:start position:0% +function +we<00:24:58.480> formulate<00:24:59.120> an<00:24:59.360> optimization<00:25:00.159> problem + +00:25:01.110 --> 00:25:01.120 align:start position:0% +we formulate an optimization problem + + +00:25:01.120 --> 00:25:05.669 align:start position:0% +we formulate an optimization problem +we<00:25:01.360> are<00:25:02.039> minimizing + +00:25:05.669 --> 00:25:05.679 align:start position:0% + + + +00:25:05.679 --> 00:25:08.549 align:start position:0% + +a<00:25:05.919> loss<00:25:06.240> function<00:25:06.799> an<00:25:07.039> average<00:25:07.520> loss<00:25:07.840> function + +00:25:08.549 --> 00:25:08.559 align:start position:0% +a loss function an average loss function + + +00:25:08.559 --> 00:25:08.950 align:start position:0% +a loss function an average loss function +over + +00:25:08.950 --> 00:25:08.960 align:start position:0% +over + + +00:25:08.960 --> 00:25:12.070 align:start position:0% +over +training<00:25:10.000> samples<00:25:10.799> one<00:25:11.039> over<00:25:11.360> m<00:25:11.520> so<00:25:11.679> i<00:25:11.760> have<00:25:11.919> a + +00:25:12.070 --> 00:25:12.080 align:start position:0% +training samples one over m so i have a + + +00:25:12.080 --> 00:25:13.110 align:start position:0% +training samples one over m so i have a +loss<00:25:12.320> function + +00:25:13.110 --> 00:25:13.120 align:start position:0% +loss function + + +00:25:13.120 --> 00:25:16.549 align:start position:0% +loss function +between<00:25:13.679> my<00:25:13.919> predictions<00:25:14.640> if<00:25:14.799> w<00:25:15.279> of<00:25:15.440> x<00:25:15.679> i + +00:25:16.549 --> 00:25:16.559 align:start position:0% +between my predictions if w of x i + + +00:25:16.559 --> 00:25:19.830 align:start position:0% +between my predictions if w of x i +and<00:25:16.799> y<00:25:17.120> i<00:25:19.279> and<00:25:19.520> i + +00:25:19.830 --> 00:25:19.840 align:start position:0% +and y i and i + + +00:25:19.840 --> 00:25:23.269 align:start position:0% +and y i and i +pick<00:25:20.480> a<00:25:20.799> model<00:25:21.520> i<00:25:21.679> pick<00:25:21.919> a<00:25:22.000> w<00:25:22.480> that<00:25:22.640> minimizes + +00:25:23.269 --> 00:25:23.279 align:start position:0% +pick a model i pick a w that minimizes + + +00:25:23.279 --> 00:25:25.510 align:start position:0% +pick a model i pick a w that minimizes +this<00:25:23.520> loss<00:25:23.760> function + +00:25:25.510 --> 00:25:25.520 align:start position:0% +this loss function + + +00:25:25.520 --> 00:25:28.230 align:start position:0% +this loss function +so<00:25:25.760> this<00:25:26.000> is<00:25:26.159> called<00:25:26.840> empirical<00:25:27.919> risk + +00:25:28.230 --> 00:25:28.240 align:start position:0% +so this is called empirical risk + + +00:25:28.240 --> 00:25:29.750 align:start position:0% +so this is called empirical risk +minimization<00:25:28.960> this<00:25:29.200> is<00:25:29.279> called + +00:25:29.750 --> 00:25:29.760 align:start position:0% +minimization this is called + + +00:25:29.760 --> 00:25:35.269 align:start position:0% +minimization this is called +e<00:25:30.320> r<00:25:30.720> m<00:25:32.400> empirical + +00:25:35.269 --> 00:25:35.279 align:start position:0% + + + +00:25:35.279 --> 00:25:40.310 align:start position:0% + +risk<00:25:36.960> minimization + +00:25:40.310 --> 00:25:40.320 align:start position:0% + + + +00:25:40.320 --> 00:25:43.990 align:start position:0% + +okay<00:25:40.640> so<00:25:40.960> why<00:25:41.360> in<00:25:42.559> uh + +00:25:43.990 --> 00:25:44.000 align:start position:0% +okay so why in uh + + +00:25:44.000 --> 00:25:45.909 align:start position:0% +okay so why in uh +empirical<00:25:44.559> risk<00:25:44.720> minimization<00:25:45.440> because<00:25:45.760> we + +00:25:45.909 --> 00:25:45.919 align:start position:0% +empirical risk minimization because we + + +00:25:45.919 --> 00:25:47.990 align:start position:0% +empirical risk minimization because we +are<00:25:46.240> uh<00:25:46.559> basically<00:25:47.039> looking<00:25:47.360> at<00:25:47.440> the<00:25:47.600> average + +00:25:47.990 --> 00:25:48.000 align:start position:0% +are uh basically looking at the average + + +00:25:48.000 --> 00:25:48.549 align:start position:0% +are uh basically looking at the average +loss<00:25:48.240> of + +00:25:48.549 --> 00:25:48.559 align:start position:0% +loss of + + +00:25:48.559 --> 00:25:50.470 align:start position:0% +loss of +or<00:25:48.799> training<00:25:49.120> sets<00:25:49.520> another<00:25:49.840> way<00:25:50.080> to<00:25:50.240> write + +00:25:50.470 --> 00:25:50.480 align:start position:0% +or training sets another way to write + + +00:25:50.480 --> 00:25:52.070 align:start position:0% +or training sets another way to write +this<00:25:50.880> optimization + +00:25:52.070 --> 00:25:52.080 align:start position:0% +this optimization + + +00:25:52.080 --> 00:25:53.750 align:start position:0% +this optimization +uh<00:25:52.240> the<00:25:52.400> objective<00:25:52.880> of<00:25:52.960> this<00:25:53.200> optimization + +00:25:53.750 --> 00:25:53.760 align:start position:0% +uh the objective of this optimization + + +00:25:53.760 --> 00:25:55.990 align:start position:0% +uh the objective of this optimization +which<00:25:54.000> is<00:25:54.159> uh<00:25:54.559> probably<00:25:55.039> easier<00:25:55.440> to<00:25:55.600> write<00:25:55.840> it + +00:25:55.990 --> 00:25:56.000 align:start position:0% +which is uh probably easier to write it + + +00:25:56.000 --> 00:25:56.549 align:start position:0% +which is uh probably easier to write it +up + +00:25:56.549 --> 00:25:56.559 align:start position:0% +up + + +00:25:56.559 --> 00:25:59.269 align:start position:0% +up +is<00:25:56.880> i'm<00:25:57.120> looking<00:25:57.360> at<00:25:57.440> the<00:25:57.679> expected<00:25:58.400> loss<00:25:58.960> over + +00:25:59.269 --> 00:25:59.279 align:start position:0% +is i'm looking at the expected loss over + + +00:25:59.279 --> 00:26:01.590 align:start position:0% +is i'm looking at the expected loss over +my<00:25:59.520> samples<00:26:00.159> coming<00:26:00.640> from + +00:26:01.590 --> 00:26:01.600 align:start position:0% +my samples coming from + + +00:26:01.600 --> 00:26:04.630 align:start position:0% +my samples coming from +my<00:26:01.919> training<00:26:02.320> set<00:26:02.559> i<00:26:02.720> have<00:26:03.440> p<00:26:03.679> x<00:26:03.919> y<00:26:04.240> n + +00:26:04.630 --> 00:26:04.640 align:start position:0% +my training set i have p x y n + + +00:26:04.640 --> 00:26:06.310 align:start position:0% +my training set i have p x y n +that's<00:26:04.880> my<00:26:05.120> training<00:26:05.440> set<00:26:05.679> i'm<00:26:05.840> looking<00:26:06.159> at + +00:26:06.310 --> 00:26:06.320 align:start position:0% +that's my training set i'm looking at + + +00:26:06.320 --> 00:26:08.470 align:start position:0% +that's my training set i'm looking at +the<00:26:07.360> average<00:26:07.919> loss + +00:26:08.470 --> 00:26:08.480 align:start position:0% +the average loss + + +00:26:08.480 --> 00:26:11.909 align:start position:0% +the average loss +from<00:26:09.440> between<00:26:10.000> f<00:26:10.159> of<00:26:10.400> x<00:26:10.799> and<00:26:11.039> y + +00:26:11.909 --> 00:26:11.919 align:start position:0% +from between f of x and y + + +00:26:11.919 --> 00:26:15.110 align:start position:0% +from between f of x and y +so<00:26:12.240> this<00:26:12.400> is<00:26:12.559> another<00:26:13.039> way<00:26:13.600> a<00:26:13.760> more<00:26:14.159> uh<00:26:14.559> compact + +00:26:15.110 --> 00:26:15.120 align:start position:0% +so this is another way a more uh compact + + +00:26:15.120 --> 00:26:16.310 align:start position:0% +so this is another way a more uh compact +way<00:26:15.440> of<00:26:15.760> uh + +00:26:16.310 --> 00:26:16.320 align:start position:0% +way of uh + + +00:26:16.320 --> 00:26:19.110 align:start position:0% +way of uh +writing<00:26:17.360> uh<00:26:17.679> the<00:26:18.000> objective<00:26:18.559> of<00:26:18.640> this<00:26:18.880> loss + +00:26:19.110 --> 00:26:19.120 align:start position:0% +writing uh the objective of this loss + + +00:26:19.120 --> 00:26:20.549 align:start position:0% +writing uh the objective of this loss +function + +00:26:20.549 --> 00:26:20.559 align:start position:0% +function + + +00:26:20.559 --> 00:26:23.990 align:start position:0% +function +all<00:26:20.640> right<00:26:20.880> so<00:26:21.919> next<00:26:22.159> i'm<00:26:22.400> going<00:26:22.640> to<00:26:23.440> dissect + +00:26:23.990 --> 00:26:24.000 align:start position:0% +all right so next i'm going to dissect + + +00:26:24.000 --> 00:26:25.669 align:start position:0% +all right so next i'm going to dissect +this<00:26:24.320> optimization<00:26:24.960> i'm<00:26:25.120> going<00:26:25.279> to<00:26:25.360> first + +00:26:25.669 --> 00:26:25.679 align:start position:0% +this optimization i'm going to first + + +00:26:25.679 --> 00:26:27.350 align:start position:0% +this optimization i'm going to first +talk<00:26:25.840> about<00:26:26.159> the<00:26:26.320> loss + +00:26:27.350 --> 00:26:27.360 align:start position:0% +talk about the loss + + +00:26:27.360 --> 00:26:29.669 align:start position:0% +talk about the loss +function<00:26:27.840> here<00:26:28.640> and<00:26:28.720> then<00:26:28.960> i'm<00:26:29.120> going<00:26:29.279> to<00:26:29.440> talk + +00:26:29.669 --> 00:26:29.679 align:start position:0% +function here and then i'm going to talk + + +00:26:29.679 --> 00:26:30.870 align:start position:0% +function here and then i'm going to talk +about<00:26:29.919> the<00:26:30.159> models + +00:26:30.870 --> 00:26:30.880 align:start position:0% +about the models + + +00:26:30.880 --> 00:26:34.230 align:start position:0% +about the models +fw<00:26:31.919> here<00:26:32.720> and<00:26:32.960> finally<00:26:33.440> i'm<00:26:33.679> going<00:26:33.840> to<00:26:34.000> talk + +00:26:34.230 --> 00:26:34.240 align:start position:0% +fw here and finally i'm going to talk + + +00:26:34.240 --> 00:26:37.190 align:start position:0% +fw here and finally i'm going to talk +about<00:26:34.480> this<00:26:34.640> whole<00:26:34.880> optimization<00:26:35.679> and<00:26:35.919> how<00:26:36.159> we + +00:26:37.190 --> 00:26:37.200 align:start position:0% +about this whole optimization and how we + + +00:26:37.200 --> 00:26:40.230 align:start position:0% +about this whole optimization and how we +solve<00:26:37.440> this<00:26:37.679> optimization<00:26:38.400> problem + +00:26:40.230 --> 00:26:40.240 align:start position:0% +solve this optimization problem + + +00:26:40.240 --> 00:26:43.110 align:start position:0% +solve this optimization problem +but<00:26:40.400> before<00:26:40.799> going<00:26:41.039> forward<00:26:42.320> are<00:26:42.559> there<00:26:42.799> some + +00:26:43.110 --> 00:26:43.120 align:start position:0% +but before going forward are there some + + +00:26:43.120 --> 00:26:48.710 align:start position:0% +but before going forward are there some +questions + +00:26:48.710 --> 00:26:48.720 align:start position:0% + + + +00:26:48.720 --> 00:26:57.190 align:start position:0% + +okay<00:26:49.039> so<00:26:49.200> let<00:26:49.360> me<00:26:50.840> see + +00:26:57.190 --> 00:26:57.200 align:start position:0% + + + +00:26:57.200 --> 00:26:58.710 align:start position:0% + +okay<00:26:57.440> so<00:26:57.679> one<00:26:57.840> question<00:26:58.159> is<00:26:58.320> about<00:26:58.559> the + +00:26:58.710 --> 00:26:58.720 align:start position:0% +okay so one question is about the + + +00:26:58.720 --> 00:27:00.789 align:start position:0% +okay so one question is about the +definition<00:26:59.360> of<00:26:59.520> deep<00:26:59.760> learning + +00:27:00.789 --> 00:27:00.799 align:start position:0% +definition of deep learning + + +00:27:00.799 --> 00:27:04.549 align:start position:0% +definition of deep learning +uh<00:27:01.279> so<00:27:01.679> i'll<00:27:02.080> talk<00:27:02.320> about<00:27:02.640> that<00:27:03.200> in<00:27:03.360> a<00:27:03.440> bit + +00:27:04.549 --> 00:27:04.559 align:start position:0% +uh so i'll talk about that in a bit + + +00:27:04.559 --> 00:27:05.990 align:start position:0% +uh so i'll talk about that in a bit +is<00:27:04.720> there<00:27:04.880> any<00:27:05.200> example<00:27:05.600> of<00:27:05.679> lecture + +00:27:05.990 --> 00:27:06.000 align:start position:0% +is there any example of lecture + + +00:27:06.000 --> 00:27:08.870 align:start position:0% +is there any example of lecture +describing<00:27:06.559> so<00:27:06.720> i<00:27:06.799> think<00:27:06.960> we<00:27:07.120> covered<00:27:07.520> that<00:27:07.840> uh + +00:27:08.870 --> 00:27:08.880 align:start position:0% +describing so i think we covered that uh + + +00:27:08.880 --> 00:27:12.310 align:start position:0% +describing so i think we covered that uh +questions<00:27:09.279> already<00:27:10.799> all<00:27:10.960> right<00:27:11.120> so<00:27:11.360> let's<00:27:11.760> uh + +00:27:12.310 --> 00:27:12.320 align:start position:0% +questions already all right so let's uh + + +00:27:12.320 --> 00:27:16.789 align:start position:0% +questions already all right so let's uh +first<00:27:12.640> talk<00:27:12.960> about<00:27:13.360> choices<00:27:14.320> for + +00:27:16.789 --> 00:27:16.799 align:start position:0% + + + +00:27:16.799 --> 00:27:19.990 align:start position:0% + +um<00:27:17.919> actually<00:27:18.240> there's<00:27:18.840> a + +00:27:19.990 --> 00:27:20.000 align:start position:0% +um actually there's a + + +00:27:20.000 --> 00:27:23.590 align:start position:0% +um actually there's a +hand<00:27:20.320> raised<00:27:20.799> um<00:27:22.240> yes<00:27:22.559> please<00:27:22.880> go<00:27:23.039> ahead + +00:27:23.590 --> 00:27:23.600 align:start position:0% +hand raised um yes please go ahead + + +00:27:23.600 --> 00:27:26.630 align:start position:0% +hand raised um yes please go ahead +causal<00:27:24.559> uh<00:27:24.799> awesome<00:27:25.039> thank<00:27:25.200> you<00:27:25.840> so<00:27:26.240> uh + +00:27:26.630 --> 00:27:26.640 align:start position:0% +causal uh awesome thank you so uh + + +00:27:26.640 --> 00:27:29.830 align:start position:0% +causal uh awesome thank you so uh +despite<00:27:27.039> the<00:27:27.600> statement<00:27:28.320> expected<00:27:29.120> uh + +00:27:29.830 --> 00:27:29.840 align:start position:0% +despite the statement expected uh + + +00:27:29.840 --> 00:27:32.950 align:start position:0% +despite the statement expected uh +loss<00:27:30.960> of<00:27:31.200> f<00:27:31.840> uh<00:27:32.399> just<00:27:32.640> the + +00:27:32.950 --> 00:27:32.960 align:start position:0% +loss of f uh just the + + +00:27:32.960 --> 00:27:36.470 align:start position:0% +loss of f uh just the +right<00:27:33.279> portion<00:27:33.840> so<00:27:34.720> um<00:27:35.120> say<00:27:35.440> that<00:27:35.840> uh + +00:27:36.470 --> 00:27:36.480 align:start position:0% +right portion so um say that uh + + +00:27:36.480 --> 00:27:39.590 align:start position:0% +right portion so um say that uh +your<00:27:37.840> problem<00:27:38.399> is<00:27:38.799> classifying + +00:27:39.590 --> 00:27:39.600 align:start position:0% +your problem is classifying + + +00:27:39.600 --> 00:27:41.830 align:start position:0% +your problem is classifying +some<00:27:39.840> data<00:27:40.159> into<00:27:40.399> two<00:27:40.880> different<00:27:41.200> categories + +00:27:41.830 --> 00:27:41.840 align:start position:0% +some data into two different categories + + +00:27:41.840 --> 00:27:43.190 align:start position:0% +some data into two different categories +like<00:27:42.080> zero<00:27:42.399> or<00:27:42.559> one + +00:27:43.190 --> 00:27:43.200 align:start position:0% +like zero or one + + +00:27:43.200 --> 00:27:46.149 align:start position:0% +like zero or one +uh<00:27:43.520> but<00:27:43.760> the<00:27:44.000> zero<00:27:44.240> category<00:27:44.799> happens<00:27:45.440> 99<00:27:46.080> of + +00:27:46.149 --> 00:27:46.159 align:start position:0% +uh but the zero category happens 99 of + + +00:27:46.159 --> 00:27:47.029 align:start position:0% +uh but the zero category happens 99 of +the<00:27:46.320> time + +00:27:47.029 --> 00:27:47.039 align:start position:0% +the time + + +00:27:47.039 --> 00:27:48.710 align:start position:0% +the time +the<00:27:47.200> one<00:27:47.440> category<00:27:47.840> happens<00:27:48.159> one<00:27:48.320> percent<00:27:48.640> of + +00:27:48.710 --> 00:27:48.720 align:start position:0% +the one category happens one percent of + + +00:27:48.720 --> 00:27:50.950 align:start position:0% +the one category happens one percent of +the<00:27:48.799> time<00:27:49.520> so + +00:27:50.950 --> 00:27:50.960 align:start position:0% +the time so + + +00:27:50.960 --> 00:27:54.070 align:start position:0% +the time so +would<00:27:51.200> we<00:27:51.679> do<00:27:51.840> this<00:27:52.240> expectation<00:27:53.520> given<00:27:53.840> that + +00:27:54.070 --> 00:27:54.080 align:start position:0% +would we do this expectation given that + + +00:27:54.080 --> 00:27:55.190 align:start position:0% +would we do this expectation given that +data<00:27:54.399> distribution + +00:27:55.190 --> 00:27:55.200 align:start position:0% +data distribution + + +00:27:55.200 --> 00:27:58.549 align:start position:0% +data distribution +or<00:27:56.559> were<00:27:56.720> we<00:27:56.960> trying<00:27:57.360> to<00:27:57.679> not + +00:27:58.549 --> 00:27:58.559 align:start position:0% +or were we trying to not + + +00:27:58.559 --> 00:28:00.149 align:start position:0% +or were we trying to not +take<00:27:58.799> into<00:27:59.039> account<00:27:59.360> the<00:27:59.520> underlying + +00:28:00.149 --> 00:28:00.159 align:start position:0% +take into account the underlying + + +00:28:00.159 --> 00:28:01.669 align:start position:0% +take into account the underlying +distribution<00:28:00.880> of + +00:28:01.669 --> 00:28:01.679 align:start position:0% +distribution of + + +00:28:01.679 --> 00:28:03.269 align:start position:0% +distribution of +one<00:28:01.840> of<00:28:01.919> them<00:28:02.159> actually<00:28:02.480> occurs<00:28:02.880> a<00:28:02.960> lot<00:28:03.120> more + +00:28:03.269 --> 00:28:03.279 align:start position:0% +one of them actually occurs a lot more + + +00:28:03.279 --> 00:28:04.870 align:start position:0% +one of them actually occurs a lot more +than<00:28:03.440> the<00:28:03.600> other<00:28:04.000> category + +00:28:04.870 --> 00:28:04.880 align:start position:0% +than the other category + + +00:28:04.880 --> 00:28:07.909 align:start position:0% +than the other category +okay<00:28:05.120> that's<00:28:05.440> a<00:28:05.520> very<00:28:05.760> good<00:28:06.080> question<00:28:07.200> um + +00:28:07.909 --> 00:28:07.919 align:start position:0% +okay that's a very good question um + + +00:28:07.919 --> 00:28:11.750 align:start position:0% +okay that's a very good question um +so<00:28:08.880> the<00:28:09.520> formulations<00:28:10.320> of<00:28:10.880> erm + +00:28:11.750 --> 00:28:11.760 align:start position:0% +so the formulations of erm + + +00:28:11.760 --> 00:28:14.710 align:start position:0% +so the formulations of erm +is<00:28:11.919> that<00:28:12.159> we<00:28:12.320> are<00:28:13.120> weighing<00:28:13.520> samples<00:28:14.159> equally + +00:28:14.710 --> 00:28:14.720 align:start position:0% +is that we are weighing samples equally + + +00:28:14.720 --> 00:28:16.310 align:start position:0% +is that we are weighing samples equally +so<00:28:14.880> we<00:28:15.039> don't<00:28:15.279> care<00:28:15.520> about<00:28:15.760> the<00:28:15.919> types<00:28:16.159> of + +00:28:16.310 --> 00:28:16.320 align:start position:0% +so we don't care about the types of + + +00:28:16.320 --> 00:28:17.350 align:start position:0% +so we don't care about the types of +labels<00:28:16.720> that<00:28:16.799> they<00:28:17.039> have + +00:28:17.350 --> 00:28:17.360 align:start position:0% +labels that they have + + +00:28:17.360 --> 00:28:19.750 align:start position:0% +labels that they have +you<00:28:17.520> have<00:28:17.760> like<00:28:18.240> 99<00:28:18.799> percent<00:28:19.279> positive + +00:28:19.750 --> 00:28:19.760 align:start position:0% +you have like 99 percent positive + + +00:28:19.760 --> 00:28:22.149 align:start position:0% +you have like 99 percent positive +samples<00:28:20.240> and<00:28:20.480> one<00:28:20.640> percent<00:28:20.960> negative<00:28:21.360> samples + +00:28:22.149 --> 00:28:22.159 align:start position:0% +samples and one percent negative samples + + +00:28:22.159 --> 00:28:23.990 align:start position:0% +samples and one percent negative samples +anyway<00:28:22.480> so<00:28:22.640> you<00:28:22.799> are<00:28:22.960> weighing<00:28:23.200> them<00:28:23.520> equally + +00:28:23.990 --> 00:28:24.000 align:start position:0% +anyway so you are weighing them equally + + +00:28:24.000 --> 00:28:25.430 align:start position:0% +anyway so you are weighing them equally +but<00:28:24.240> if<00:28:24.399> you<00:28:24.559> solve<00:28:24.880> them + +00:28:25.430 --> 00:28:25.440 align:start position:0% +but if you solve them + + +00:28:25.440 --> 00:28:28.789 align:start position:0% +but if you solve them +then<00:28:25.919> your<00:28:26.799> majority<00:28:27.360> class<00:28:27.679> may + +00:28:28.789 --> 00:28:28.799 align:start position:0% +then your majority class may + + +00:28:28.799 --> 00:28:31.830 align:start position:0% +then your majority class may +outpower<00:28:29.520> your<00:28:29.760> minority<00:28:30.399> class<00:28:31.039> so<00:28:31.279> we'll + +00:28:31.830 --> 00:28:31.840 align:start position:0% +outpower your minority class so we'll + + +00:28:31.840 --> 00:28:34.710 align:start position:0% +outpower your minority class so we'll +talk<00:28:32.399> in<00:28:32.799> uh<00:28:33.120> in<00:28:33.279> fact<00:28:33.520> the<00:28:33.840> fairness<00:28:34.320> lecture + +00:28:34.710 --> 00:28:34.720 align:start position:0% +talk in uh in fact the fairness lecture + + +00:28:34.720 --> 00:28:35.990 align:start position:0% +talk in uh in fact the fairness lecture +and<00:28:34.880> some<00:28:35.039> other<00:28:35.360> lectures + +00:28:35.990 --> 00:28:36.000 align:start position:0% +and some other lectures + + +00:28:36.000 --> 00:28:38.470 align:start position:0% +and some other lectures +how<00:28:36.240> we<00:28:36.480> can<00:28:36.640> renormalize<00:28:37.679> some<00:28:37.840> of<00:28:38.000> the<00:28:38.159> ways + +00:28:38.470 --> 00:28:38.480 align:start position:0% +how we can renormalize some of the ways + + +00:28:38.480 --> 00:28:39.750 align:start position:0% +how we can renormalize some of the ways +in<00:28:38.640> order<00:28:39.039> to + +00:28:39.750 --> 00:28:39.760 align:start position:0% +in order to + + +00:28:39.760 --> 00:28:43.350 align:start position:0% +in order to +reduce<00:28:40.320> biases<00:28:41.679> of<00:28:42.240> basically<00:28:42.880> uh + +00:28:43.350 --> 00:28:43.360 align:start position:0% +reduce biases of basically uh + + +00:28:43.360 --> 00:28:46.470 align:start position:0% +reduce biases of basically uh +the<00:28:43.600> erm<00:28:44.080> optimization<00:28:44.799> problem<00:28:45.440> but<00:28:45.840> for<00:28:46.080> now + +00:28:46.470 --> 00:28:46.480 align:start position:0% +the erm optimization problem but for now + + +00:28:46.480 --> 00:28:47.350 align:start position:0% +the erm optimization problem but for now +let's<00:28:46.799> assume + +00:28:47.350 --> 00:28:47.360 align:start position:0% +let's assume + + +00:28:47.360 --> 00:28:49.830 align:start position:0% +let's assume +all<00:28:47.600> samples<00:28:48.159> are<00:28:48.399> made<00:28:48.799> equally<00:28:49.279> so<00:28:49.520> that's + +00:28:49.830 --> 00:28:49.840 align:start position:0% +all samples are made equally so that's + + +00:28:49.840 --> 00:28:51.669 align:start position:0% +all samples are made equally so that's +the<00:28:50.080> formulation<00:28:50.640> that<00:28:50.799> we<00:28:50.960> consider<00:28:51.440> in<00:28:51.520> this + +00:28:51.669 --> 00:28:51.679 align:start position:0% +the formulation that we consider in this + + +00:28:51.679 --> 00:28:52.310 align:start position:0% +the formulation that we consider in this +lecture + +00:28:52.310 --> 00:28:52.320 align:start position:0% +lecture + + +00:28:52.320 --> 00:28:54.470 align:start position:0% +lecture +okay<00:28:52.559> awesome<00:28:52.880> thanks<00:28:53.440> very<00:28:53.679> good<00:28:54.240> good + +00:28:54.470 --> 00:28:54.480 align:start position:0% +okay awesome thanks very good good + + +00:28:54.480 --> 00:28:55.830 align:start position:0% +okay awesome thanks very good good +question + +00:28:55.830 --> 00:28:55.840 align:start position:0% +question + + +00:28:55.840 --> 00:28:59.110 align:start position:0% +question +all<00:28:56.000> right<00:28:56.240> so<00:28:56.960> uh<00:28:57.360> what<00:28:57.600> are<00:28:57.760> the<00:28:58.080> choices + +00:28:59.110 --> 00:28:59.120 align:start position:0% +all right so uh what are the choices + + +00:28:59.120 --> 00:29:04.630 align:start position:0% +all right so uh what are the choices +for<00:28:59.840> the<00:29:00.399> loss<00:29:00.720> function<00:29:02.080> so<00:29:02.320> that's<00:29:02.640> the + +00:29:04.630 --> 00:29:04.640 align:start position:0% +for the loss function so that's the + + +00:29:04.640 --> 00:29:09.590 align:start position:0% +for the loss function so that's the +start<00:29:05.039> source<00:29:05.440> choices<00:29:06.000> of<00:29:06.960> the + +00:29:09.590 --> 00:29:09.600 align:start position:0% + + + +00:29:09.600 --> 00:29:15.110 align:start position:0% + +loss<00:29:10.840> function + +00:29:15.110 --> 00:29:15.120 align:start position:0% + + + +00:29:15.120 --> 00:29:20.230 align:start position:0% + +so<00:29:15.360> in<00:29:15.520> the<00:29:15.760> regression<00:29:16.480> problem + +00:29:20.230 --> 00:29:20.240 align:start position:0% + + + +00:29:20.240 --> 00:29:24.789 align:start position:0% + +when<00:29:20.559> my<00:29:21.520> y<00:29:22.240> is<00:29:22.480> in<00:29:22.720> real<00:29:23.039> domain + +00:29:24.789 --> 00:29:24.799 align:start position:0% +when my y is in real domain + + +00:29:24.799 --> 00:29:27.269 align:start position:0% +when my y is in real domain +so<00:29:24.960> a<00:29:25.120> good<00:29:25.360> loss<00:29:25.600> function<00:29:26.240> is<00:29:26.640> quadratic + +00:29:27.269 --> 00:29:27.279 align:start position:0% +so a good loss function is quadratic + + +00:29:27.279 --> 00:29:31.029 align:start position:0% +so a good loss function is quadratic +loss<00:29:27.679> so<00:29:28.000> here<00:29:28.320> we<00:29:28.480> have<00:29:28.799> fw<00:29:29.440> of<00:29:29.679> x + +00:29:31.029 --> 00:29:31.039 align:start position:0% +loss so here we have fw of x + + +00:29:31.039 --> 00:29:38.830 align:start position:0% +loss so here we have fw of x +is<00:29:31.200> in<00:29:31.440> real<00:29:32.559> so<00:29:32.720> i'm<00:29:32.880> gonna<00:29:33.200> use<00:29:33.600> a<00:29:33.760> quadratic + +00:29:38.830 --> 00:29:38.840 align:start position:0% + + + +00:29:38.840 --> 00:29:40.549 align:start position:0% + +loss + +00:29:40.549 --> 00:29:40.559 align:start position:0% +loss + + +00:29:40.559 --> 00:29:43.830 align:start position:0% +loss +which<00:29:40.799> is<00:29:40.880> defined<00:29:41.520> as<00:29:42.799> loss<00:29:43.200> between + +00:29:43.830 --> 00:29:43.840 align:start position:0% +which is defined as loss between + + +00:29:43.840 --> 00:29:50.230 align:start position:0% +which is defined as loss between +f<00:29:44.159> w<00:29:44.720> of<00:29:44.960> x<00:29:45.360> and<00:29:45.520> y + +00:29:50.230 --> 00:29:50.240 align:start position:0% + + + +00:29:50.240 --> 00:29:52.950 align:start position:0% + +the<00:29:50.480> quadratic<00:29:51.039> norm<00:29:51.279> between<00:29:51.760> f<00:29:51.840> w<00:29:52.240> and<00:29:52.559> x<00:29:52.799> and + +00:29:52.950 --> 00:29:52.960 align:start position:0% +the quadratic norm between f w and x and + + +00:29:52.960 --> 00:29:55.909 align:start position:0% +the quadratic norm between f w and x and +y<00:29:53.279> it's<00:29:53.440> very<00:29:53.679> simple + +00:29:55.909 --> 00:29:55.919 align:start position:0% + + + +00:29:55.919 --> 00:29:58.470 align:start position:0% + +so<00:29:56.320> we<00:29:56.799> like<00:29:57.039> this<00:29:57.279> loss<00:29:57.600> function<00:29:58.000> for<00:29:58.159> many + +00:29:58.470 --> 00:29:58.480 align:start position:0% +so we like this loss function for many + + +00:29:58.480 --> 00:29:59.190 align:start position:0% +so we like this loss function for many +reasons + +00:29:59.190 --> 00:29:59.200 align:start position:0% +reasons + + +00:29:59.200 --> 00:30:02.389 align:start position:0% +reasons +um<00:30:00.159> one<00:30:00.399> of<00:30:00.480> the<00:30:00.640> reason<00:30:01.120> is<00:30:01.360> that<00:30:02.159> the + +00:30:02.389 --> 00:30:02.399 align:start position:0% +um one of the reason is that the + + +00:30:02.399 --> 00:30:04.470 align:start position:0% +um one of the reason is that the +derivative<00:30:03.120> of<00:30:03.279> a<00:30:03.360> quadratic<00:30:03.919> function<00:30:04.320> is + +00:30:04.470 --> 00:30:04.480 align:start position:0% +derivative of a quadratic function is + + +00:30:04.480 --> 00:30:08.470 align:start position:0% +derivative of a quadratic function is +linear<00:30:05.039> so<00:30:05.279> it<00:30:05.440> makes<00:30:05.760> the<00:30:06.000> analysis<00:30:06.840> easy + +00:30:08.470 --> 00:30:08.480 align:start position:0% +linear so it makes the analysis easy + + +00:30:08.480 --> 00:30:11.669 align:start position:0% +linear so it makes the analysis easy +other<00:30:08.799> reasons<00:30:09.279> are<00:30:09.520> it<00:30:09.760> is<00:30:10.799> differentiable + +00:30:11.669 --> 00:30:11.679 align:start position:0% +other reasons are it is differentiable + + +00:30:11.679 --> 00:30:14.149 align:start position:0% +other reasons are it is differentiable +it<00:30:11.840> is<00:30:12.000> a<00:30:12.159> smooth<00:30:12.559> function + +00:30:14.149 --> 00:30:14.159 align:start position:0% +it is a smooth function + + +00:30:14.159 --> 00:30:16.230 align:start position:0% +it is a smooth function +so<00:30:14.320> that's<00:30:14.640> why<00:30:14.880> quadratic<00:30:15.520> loss<00:30:15.840> is<00:30:15.919> one<00:30:16.159> of + +00:30:16.230 --> 00:30:16.240 align:start position:0% +so that's why quadratic loss is one of + + +00:30:16.240 --> 00:30:18.470 align:start position:0% +so that's why quadratic loss is one of +the<00:30:16.559> very<00:30:16.880> popular<00:30:17.440> laws<00:30:17.760> especially + +00:30:18.470 --> 00:30:18.480 align:start position:0% +the very popular laws especially + + +00:30:18.480 --> 00:30:21.990 align:start position:0% +the very popular laws especially +in<00:30:18.559> the<00:30:18.880> regression<00:30:20.840> problem + +00:30:21.990 --> 00:30:22.000 align:start position:0% +in the regression problem + + +00:30:22.000 --> 00:30:32.470 align:start position:0% +in the regression problem +let's<00:30:22.240> look<00:30:22.480> at<00:30:22.640> the<00:30:22.880> classification<00:30:23.760> problem + +00:30:32.470 --> 00:30:32.480 align:start position:0% + + + +00:30:32.480 --> 00:30:35.190 align:start position:0% + +okay<00:30:32.799> in<00:30:32.880> the<00:30:33.120> classification<00:30:34.000> problem<00:30:34.640> also + +00:30:35.190 --> 00:30:35.200 align:start position:0% +okay in the classification problem also + + +00:30:35.200 --> 00:30:35.510 align:start position:0% +okay in the classification problem also +the + +00:30:35.510 --> 00:30:35.520 align:start position:0% +the + + +00:30:35.520 --> 00:30:38.710 align:start position:0% +the +output<00:30:36.000> of<00:30:36.240> my<00:30:37.039> network<00:30:37.760> fw + +00:30:38.710 --> 00:30:38.720 align:start position:0% +output of my network fw + + +00:30:38.720 --> 00:30:42.070 align:start position:0% +output of my network fw +will<00:30:38.960> be<00:30:39.200> in<00:30:39.840> reals<00:30:40.880> but<00:30:41.279> my + +00:30:42.070 --> 00:30:42.080 align:start position:0% +will be in reals but my + + +00:30:42.080 --> 00:30:45.909 align:start position:0% +will be in reals but my +labels<00:30:42.640> are<00:30:42.880> in<00:30:43.600> let's<00:30:43.840> say<00:30:44.640> in<00:30:44.880> a + +00:30:45.909 --> 00:30:45.919 align:start position:0% +labels are in let's say in a + + +00:30:45.919 --> 00:30:48.310 align:start position:0% +labels are in let's say in a +binary<00:30:46.399> field<00:30:46.720> so<00:30:46.960> it<00:30:47.120> is<00:30:47.200> either<00:30:47.679> minus<00:30:48.080> one + +00:30:48.310 --> 00:30:48.320 align:start position:0% +binary field so it is either minus one + + +00:30:48.320 --> 00:30:49.269 align:start position:0% +binary field so it is either minus one +or<00:30:48.480> plus<00:30:48.880> one + +00:30:49.269 --> 00:30:49.279 align:start position:0% +or plus one + + +00:30:49.279 --> 00:30:51.750 align:start position:0% +or plus one +so<00:30:49.520> first<00:30:49.840> i<00:30:50.000> need<00:30:50.159> to<00:30:50.640> convert<00:30:51.200> these<00:30:51.440> real + +00:30:51.750 --> 00:30:51.760 align:start position:0% +so first i need to convert these real + + +00:30:51.760 --> 00:30:52.870 align:start position:0% +so first i need to convert these real +numbers + +00:30:52.870 --> 00:30:52.880 align:start position:0% +numbers + + +00:30:52.880 --> 00:30:56.070 align:start position:0% +numbers +to<00:30:53.679> some<00:30:54.399> uh<00:30:54.720> binary<00:30:55.279> labels + +00:30:56.070 --> 00:30:56.080 align:start position:0% +to some uh binary labels + + +00:30:56.080 --> 00:30:59.990 align:start position:0% +to some uh binary labels +so<00:30:56.240> in<00:30:56.399> order<00:30:56.640> to<00:30:56.799> do<00:30:56.960> that<00:30:57.279> one<00:30:57.519> way<00:30:57.840> is<00:30:58.640> to + +00:30:59.990 --> 00:31:00.000 align:start position:0% +so in order to do that one way is to + + +00:31:00.000 --> 00:31:02.230 align:start position:0% +so in order to do that one way is to +threshold<00:31:00.559> it<00:31:00.720> so<00:31:00.880> if<00:31:01.039> you<00:31:01.200> look<00:31:01.440> at<00:31:01.840> the + +00:31:02.230 --> 00:31:02.240 align:start position:0% +threshold it so if you look at the + + +00:31:02.240 --> 00:31:03.750 align:start position:0% +threshold it so if you look at the +output<00:31:02.640> that<00:31:02.799> you<00:31:03.039> have + +00:31:03.750 --> 00:31:03.760 align:start position:0% +output that you have + + +00:31:03.760 --> 00:31:07.990 align:start position:0% +output that you have +if<00:31:04.080> w<00:31:05.120> of<00:31:05.279> x<00:31:05.600> if<00:31:05.760> it<00:31:06.000> is<00:31:06.640> positive<00:31:07.279> you<00:31:07.519> predict + +00:31:07.990 --> 00:31:08.000 align:start position:0% +if w of x if it is positive you predict + + +00:31:08.000 --> 00:31:09.110 align:start position:0% +if w of x if it is positive you predict +the<00:31:08.080> label<00:31:08.480> one + +00:31:09.110 --> 00:31:09.120 align:start position:0% +the label one + + +00:31:09.120 --> 00:31:12.710 align:start position:0% +the label one +otherwise<00:31:09.760> you<00:31:10.000> predict<00:31:10.399> the<00:31:11.039> label<00:31:11.519> zero<00:31:11.919> or + +00:31:12.710 --> 00:31:12.720 align:start position:0% +otherwise you predict the label zero or + + +00:31:12.720 --> 00:31:15.430 align:start position:0% +otherwise you predict the label zero or +minus<00:31:13.120> one<00:31:13.840> and<00:31:14.080> my<00:31:14.320> last<00:31:14.640> function<00:31:15.039> would<00:31:15.200> be + +00:31:15.430 --> 00:31:15.440 align:start position:0% +minus one and my last function would be + + +00:31:15.440 --> 00:31:16.149 align:start position:0% +minus one and my last function would be +just + +00:31:16.149 --> 00:31:16.159 align:start position:0% +just + + +00:31:16.159 --> 00:31:18.389 align:start position:0% +just +comparing<00:31:16.880> the<00:31:17.039> labels<00:31:17.519> that<00:31:17.679> i'm<00:31:17.840> predicting + +00:31:18.389 --> 00:31:18.399 align:start position:0% +comparing the labels that i'm predicting + + +00:31:18.399 --> 00:31:19.350 align:start position:0% +comparing the labels that i'm predicting +with<00:31:18.640> my + +00:31:19.350 --> 00:31:19.360 align:start position:0% +with my + + +00:31:19.360 --> 00:31:22.389 align:start position:0% +with my +true<00:31:19.600> labels<00:31:20.640> so<00:31:20.880> in<00:31:20.960> that<00:31:21.200> case<00:31:21.519> my<00:31:21.840> loss + +00:31:22.389 --> 00:31:22.399 align:start position:0% +true labels so in that case my loss + + +00:31:22.399 --> 00:31:25.509 align:start position:0% +true labels so in that case my loss +would<00:31:22.640> be<00:31:23.600> something<00:31:24.000> like<00:31:24.240> this + +00:31:25.509 --> 00:31:25.519 align:start position:0% +would be something like this + + +00:31:25.519 --> 00:31:29.110 align:start position:0% +would be something like this +so<00:31:25.679> i<00:31:25.919> have<00:31:26.799> uh<00:31:27.200> fw + +00:31:29.110 --> 00:31:29.120 align:start position:0% +so i have uh fw + + +00:31:29.120 --> 00:31:33.029 align:start position:0% +so i have uh fw +of<00:31:29.279> x<00:31:29.600> times<00:31:30.000> y<00:31:30.720> if<00:31:31.200> both<00:31:31.600> f<00:31:31.760> w<00:31:32.159> and<00:31:32.320> y<00:31:32.559> they<00:31:32.799> have + +00:31:33.029 --> 00:31:33.039 align:start position:0% +of x times y if both f w and y they have + + +00:31:33.039 --> 00:31:35.430 align:start position:0% +of x times y if both f w and y they have +the<00:31:33.200> same<00:31:33.519> signs<00:31:33.840> it<00:31:34.000> means<00:31:34.320> i<00:31:34.480> do<00:31:34.720> not<00:31:35.039> have + +00:31:35.430 --> 00:31:35.440 align:start position:0% +the same signs it means i do not have + + +00:31:35.440 --> 00:31:38.470 align:start position:0% +the same signs it means i do not have +any<00:31:35.840> error<00:31:36.399> because<00:31:37.039> my<00:31:37.200> prediction<00:31:37.679> will<00:31:37.840> be + +00:31:38.470 --> 00:31:38.480 align:start position:0% +any error because my prediction will be + + +00:31:38.480 --> 00:31:41.350 align:start position:0% +any error because my prediction will be +uh<00:31:38.799> plus<00:31:39.120> one<00:31:39.360> my<00:31:39.679> y<00:31:40.000> my<00:31:40.320> y<00:31:40.559> is<00:31:40.720> also<00:31:40.960> plus<00:31:41.200> one + +00:31:41.350 --> 00:31:41.360 align:start position:0% +uh plus one my y my y is also plus one + + +00:31:41.360 --> 00:31:43.430 align:start position:0% +uh plus one my y my y is also plus one +so<00:31:41.519> i<00:31:41.600> don't<00:31:41.760> have<00:31:41.919> any<00:31:42.159> error + +00:31:43.430 --> 00:31:43.440 align:start position:0% +so i don't have any error + + +00:31:43.440 --> 00:31:46.149 align:start position:0% +so i don't have any error +if<00:31:43.840> f<00:31:44.080> w<00:31:44.399> times<00:31:44.720> y<00:31:45.039> is<00:31:45.200> negative<00:31:45.679> it<00:31:45.840> means<00:31:46.000> the + +00:31:46.149 --> 00:31:46.159 align:start position:0% +if f w times y is negative it means the + + +00:31:46.159 --> 00:31:48.230 align:start position:0% +if f w times y is negative it means the +signs<00:31:46.559> are<00:31:46.640> different<00:31:47.039> so<00:31:47.200> i'll<00:31:47.440> have<00:31:47.760> one + +00:31:48.230 --> 00:31:48.240 align:start position:0% +signs are different so i'll have one + + +00:31:48.240 --> 00:31:51.669 align:start position:0% +signs are different so i'll have one +unit<00:31:48.640> of<00:31:48.880> error<00:31:50.080> so<00:31:50.320> this<00:31:50.559> is<00:31:50.720> the + +00:31:51.669 --> 00:31:51.679 align:start position:0% +unit of error so this is the + + +00:31:51.679 --> 00:31:56.070 align:start position:0% +unit of error so this is the +0<00:31:52.720> 1<00:31:54.480> loss<00:31:54.799> function<00:31:55.200> so + +00:31:56.070 --> 00:31:56.080 align:start position:0% +0 1 loss function so + + +00:31:56.080 --> 00:31:59.669 align:start position:0% +0 1 loss function so +0<00:31:58.320> 1 + +00:31:59.669 --> 00:31:59.679 align:start position:0% +0 1 + + +00:31:59.679 --> 00:32:05.509 align:start position:0% +0 1 +loss<00:31:59.919> function<00:32:02.240> okay<00:32:02.559> so + +00:32:05.509 --> 00:32:05.519 align:start position:0% + + + +00:32:05.519 --> 00:32:07.509 align:start position:0% + +the<00:32:05.679> zero<00:32:06.000> one<00:32:06.240> loss<00:32:06.480> function<00:32:06.880> we<00:32:07.039> don't<00:32:07.279> like + +00:32:07.509 --> 00:32:07.519 align:start position:0% +the zero one loss function we don't like + + +00:32:07.519 --> 00:32:09.750 align:start position:0% +the zero one loss function we don't like +it<00:32:07.679> as<00:32:07.840> much<00:32:08.000> as<00:32:08.240> quadratic<00:32:08.880> loss + +00:32:09.750 --> 00:32:09.760 align:start position:0% +it as much as quadratic loss + + +00:32:09.760 --> 00:32:12.710 align:start position:0% +it as much as quadratic loss +um<00:32:10.559> there<00:32:10.880> are<00:32:11.120> a<00:32:11.200> lot<00:32:11.440> of<00:32:11.600> reasons<00:32:12.159> one<00:32:12.399> reason + +00:32:12.710 --> 00:32:12.720 align:start position:0% +um there are a lot of reasons one reason + + +00:32:12.720 --> 00:32:14.789 align:start position:0% +um there are a lot of reasons one reason +is<00:32:12.880> that<00:32:13.039> it<00:32:13.200> is<00:32:13.360> a<00:32:13.440> non-convex<00:32:14.159> function<00:32:14.559> as + +00:32:14.789 --> 00:32:14.799 align:start position:0% +is that it is a non-convex function as + + +00:32:14.799 --> 00:32:17.990 align:start position:0% +is that it is a non-convex function as +you<00:32:14.880> can<00:32:15.039> see<00:32:15.279> here + +00:32:17.990 --> 00:32:18.000 align:start position:0% + + + +00:32:18.000 --> 00:32:20.789 align:start position:0% + +and<00:32:18.320> another<00:32:18.720> reason<00:32:19.279> is<00:32:19.440> that<00:32:19.760> the<00:32:20.159> gradient + +00:32:20.789 --> 00:32:20.799 align:start position:0% +and another reason is that the gradient + + +00:32:20.799 --> 00:32:22.149 align:start position:0% +and another reason is that the gradient +of<00:32:20.880> this<00:32:21.120> function<00:32:21.519> if<00:32:21.679> you<00:32:21.760> look<00:32:21.919> at<00:32:22.000> the + +00:32:22.149 --> 00:32:22.159 align:start position:0% +of this function if you look at the + + +00:32:22.159 --> 00:32:23.509 align:start position:0% +of this function if you look at the +derivative + +00:32:23.509 --> 00:32:23.519 align:start position:0% +derivative + + +00:32:23.519 --> 00:32:26.310 align:start position:0% +derivative +in<00:32:23.679> different<00:32:24.159> points<00:32:24.640> it<00:32:24.880> is<00:32:25.919> almost + +00:32:26.310 --> 00:32:26.320 align:start position:0% +in different points it is almost + + +00:32:26.320 --> 00:32:26.950 align:start position:0% +in different points it is almost +everywhere + +00:32:26.950 --> 00:32:26.960 align:start position:0% +everywhere + + +00:32:26.960 --> 00:32:31.509 align:start position:0% +everywhere +zero<00:32:27.360> so<00:32:27.760> that<00:32:28.080> means<00:32:28.880> the<00:32:29.279> gradient<00:32:30.000> doesn't + +00:32:31.509 --> 00:32:31.519 align:start position:0% +zero so that means the gradient doesn't + + +00:32:31.519 --> 00:32:34.950 align:start position:0% +zero so that means the gradient doesn't +provide<00:32:32.080> information<00:32:33.679> especially<00:32:34.320> if<00:32:34.480> we<00:32:34.640> use + +00:32:34.950 --> 00:32:34.960 align:start position:0% +provide information especially if we use + + +00:32:34.960 --> 00:32:37.430 align:start position:0% +provide information especially if we use +the<00:32:35.120> gradient<00:32:35.600> information<00:32:36.159> in<00:32:36.320> order<00:32:36.640> to + +00:32:37.430 --> 00:32:37.440 align:start position:0% +the gradient information in order to + + +00:32:37.440 --> 00:32:40.389 align:start position:0% +the gradient information in order to +update<00:32:37.840> model<00:32:38.240> parameters<00:32:39.440> so<00:32:39.679> that's<00:32:40.000> one<00:32:40.240> of + +00:32:40.389 --> 00:32:40.399 align:start position:0% +update model parameters so that's one of + + +00:32:40.399 --> 00:32:41.990 align:start position:0% +update model parameters so that's one of +a<00:32:40.640> couple<00:32:40.960> of<00:32:41.120> drawbacks<00:32:41.600> for<00:32:41.760> the + +00:32:41.990 --> 00:32:42.000 align:start position:0% +a couple of drawbacks for the + + +00:32:42.000 --> 00:32:43.669 align:start position:0% +a couple of drawbacks for the +zero<00:32:42.240> one<00:32:42.399> loss<00:32:42.640> function<00:32:43.039> it<00:32:43.200> is<00:32:43.360> also + +00:32:43.669 --> 00:32:43.679 align:start position:0% +zero one loss function it is also + + +00:32:43.679 --> 00:32:45.190 align:start position:0% +zero one loss function it is also +non-smooth + +00:32:45.190 --> 00:32:45.200 align:start position:0% +non-smooth + + +00:32:45.200 --> 00:32:48.389 align:start position:0% +non-smooth +so<00:32:45.360> in<00:32:45.519> order<00:32:45.760> to<00:32:46.000> deal<00:32:46.399> with<00:32:46.640> these<00:32:47.760> problems + +00:32:48.389 --> 00:32:48.399 align:start position:0% +so in order to deal with these problems + + +00:32:48.399 --> 00:32:49.750 align:start position:0% +so in order to deal with these problems +we + +00:32:49.750 --> 00:32:49.760 align:start position:0% +we + + +00:32:49.760 --> 00:32:53.190 align:start position:0% +we +approximated<00:32:50.960> we<00:32:51.200> approximated<00:32:52.000> using<00:32:52.720> a + +00:32:53.190 --> 00:32:53.200 align:start position:0% +approximated we approximated using a + + +00:32:53.200 --> 00:32:56.789 align:start position:0% +approximated we approximated using a +function<00:32:53.600> called<00:32:53.919> hinge<00:32:54.320> loss<00:32:55.200> so<00:32:55.440> we + +00:32:56.789 --> 00:32:56.799 align:start position:0% +function called hinge loss so we + + +00:32:56.799 --> 00:32:59.669 align:start position:0% +function called hinge loss so we +have<00:32:57.760> instead<00:32:58.080> of<00:32:58.240> the<00:32:58.399> zero<00:32:58.720> one<00:32:58.960> loss<00:32:59.279> so<00:32:59.519> we + +00:32:59.669 --> 00:32:59.679 align:start position:0% +have instead of the zero one loss so we + + +00:32:59.679 --> 00:33:02.630 align:start position:0% +have instead of the zero one loss so we +have<00:32:59.919> a<00:33:00.000> hinge<00:33:00.399> loss + +00:33:02.630 --> 00:33:02.640 align:start position:0% +have a hinge loss + + +00:33:02.640 --> 00:33:06.389 align:start position:0% +have a hinge loss +so<00:33:02.880> hinge<00:33:03.519> when<00:33:03.840> fw<00:33:04.399> times<00:33:04.720> y<00:33:05.039> is<00:33:05.279> above<00:33:05.760> one + +00:33:06.389 --> 00:33:06.399 align:start position:0% +so hinge when fw times y is above one + + +00:33:06.399 --> 00:33:09.110 align:start position:0% +so hinge when fw times y is above one +it<00:33:06.640> is<00:33:06.880> zero<00:33:07.360> so<00:33:07.679> that<00:33:07.919> part<00:33:08.240> is<00:33:08.559> basically<00:33:08.960> the + +00:33:09.110 --> 00:33:09.120 align:start position:0% +it is zero so that part is basically the + + +00:33:09.120 --> 00:33:10.630 align:start position:0% +it is zero so that part is basically the +same<00:33:09.360> as<00:33:09.519> the<00:33:09.679> zero<00:33:09.919> one<00:33:10.159> laws + +00:33:10.630 --> 00:33:10.640 align:start position:0% +same as the zero one laws + + +00:33:10.640 --> 00:33:13.830 align:start position:0% +same as the zero one laws +but<00:33:10.880> before<00:33:11.440> that<00:33:12.080> it<00:33:12.480> linearly<00:33:13.120> increases + +00:33:13.830 --> 00:33:13.840 align:start position:0% +but before that it linearly increases + + +00:33:13.840 --> 00:33:14.310 align:start position:0% +but before that it linearly increases +the + +00:33:14.310 --> 00:33:14.320 align:start position:0% +the + + +00:33:14.320 --> 00:33:18.310 align:start position:0% +the +uh<00:33:15.200> the<00:33:15.440> loss<00:33:15.760> that<00:33:16.000> we<00:33:16.320> see<00:33:16.799> between<00:33:17.200> fw + +00:33:18.310 --> 00:33:18.320 align:start position:0% +uh the loss that we see between fw + + +00:33:18.320 --> 00:33:21.830 align:start position:0% +uh the loss that we see between fw +of<00:33:18.559> x<00:33:18.880> and<00:33:19.120> y<00:33:20.240> so<00:33:20.559> formally<00:33:21.120> speaking + +00:33:21.830 --> 00:33:21.840 align:start position:0% +of x and y so formally speaking + + +00:33:21.840 --> 00:33:27.990 align:start position:0% +of x and y so formally speaking +the<00:33:22.480> hinge<00:33:22.840> loss + +00:33:27.990 --> 00:33:28.000 align:start position:0% + + + +00:33:28.000 --> 00:33:31.909 align:start position:0% + +is<00:33:28.240> defined<00:33:29.679> as<00:33:30.159> max<00:33:30.640> between + +00:33:31.909 --> 00:33:31.919 align:start position:0% +is defined as max between + + +00:33:31.919 --> 00:33:35.029 align:start position:0% +is defined as max between +0<00:33:32.399> and<00:33:32.640> 1<00:33:32.880> minus<00:33:33.600> y<00:33:33.919> times<00:33:34.320> fw + +00:33:35.029 --> 00:33:35.039 align:start position:0% +0 and 1 minus y times fw + + +00:33:35.039 --> 00:33:38.470 align:start position:0% +0 and 1 minus y times fw +of<00:33:35.279> x<00:33:36.399> you<00:33:36.559> can<00:33:36.720> see<00:33:36.960> here<00:33:37.279> if<00:33:37.600> this + +00:33:38.470 --> 00:33:38.480 align:start position:0% +of x you can see here if this + + +00:33:38.480 --> 00:33:41.669 align:start position:0% +of x you can see here if this +quantity<00:33:39.440> y<00:33:39.679> times<00:33:39.919> fw<00:33:40.399> of<00:33:40.559> x<00:33:40.799> is<00:33:40.960> above<00:33:41.279> one + +00:33:41.669 --> 00:33:41.679 align:start position:0% +quantity y times fw of x is above one + + +00:33:41.679 --> 00:33:44.710 align:start position:0% +quantity y times fw of x is above one +so<00:33:41.919> this<00:33:42.080> whole<00:33:42.320> thing<00:33:42.559> is<00:33:42.720> negative<00:33:43.360> then<00:33:44.240> max + +00:33:44.710 --> 00:33:44.720 align:start position:0% +so this whole thing is negative then max + + +00:33:44.720 --> 00:33:46.630 align:start position:0% +so this whole thing is negative then max +of<00:33:44.960> zero<00:33:45.200> and<00:33:45.279> that<00:33:45.440> negative<00:33:45.840> number<00:33:46.240> will<00:33:46.399> be + +00:33:46.630 --> 00:33:46.640 align:start position:0% +of zero and that negative number will be + + +00:33:46.640 --> 00:33:49.029 align:start position:0% +of zero and that negative number will be +zero<00:33:47.440> uh<00:33:47.679> that's<00:33:48.000> this<00:33:48.240> part<00:33:48.559> and<00:33:48.640> the<00:33:48.799> other + +00:33:49.029 --> 00:33:49.039 align:start position:0% +zero uh that's this part and the other + + +00:33:49.039 --> 00:33:49.750 align:start position:0% +zero uh that's this part and the other +part + +00:33:49.750 --> 00:33:49.760 align:start position:0% +part + + +00:33:49.760 --> 00:33:52.149 align:start position:0% +part +will<00:33:50.000> basically<00:33:50.559> a<00:33:51.039> linear<00:33:51.519> function<00:33:52.000> with + +00:33:52.149 --> 00:33:52.159 align:start position:0% +will basically a linear function with + + +00:33:52.159 --> 00:33:53.269 align:start position:0% +will basically a linear function with +respect<00:33:52.640> to + +00:33:53.269 --> 00:33:53.279 align:start position:0% +respect to + + +00:33:53.279 --> 00:33:56.950 align:start position:0% +respect to +f<00:33:53.440> w<00:33:54.159> of<00:33:54.399> x<00:33:54.559> times<00:33:54.960> y<00:33:55.919> and<00:33:56.080> there<00:33:56.399> are<00:33:56.559> some + +00:33:56.950 --> 00:33:56.960 align:start position:0% +f w of x times y and there are some + + +00:33:56.960 --> 00:33:59.350 align:start position:0% +f w of x times y and there are some +you<00:33:57.120> know<00:33:57.679> other<00:33:58.080> justifications<00:33:58.960> for<00:33:59.200> the + +00:33:59.350 --> 00:33:59.360 align:start position:0% +you know other justifications for the + + +00:33:59.360 --> 00:34:01.990 align:start position:0% +you know other justifications for the +hinge<00:33:59.679> loss<00:34:00.080> based<00:34:00.399> on<00:34:00.640> the + +00:34:01.990 --> 00:34:02.000 align:start position:0% +hinge loss based on the + + +00:34:02.000 --> 00:34:05.269 align:start position:0% +hinge loss based on the +margin<00:34:02.480> maximization<00:34:03.919> uh<00:34:04.559> take<00:34:04.799> a<00:34:04.960> look<00:34:05.200> at + +00:34:05.269 --> 00:34:05.279 align:start position:0% +margin maximization uh take a look at + + +00:34:05.279 --> 00:34:07.509 align:start position:0% +margin maximization uh take a look at +the<00:34:05.600> svm + +00:34:07.509 --> 00:34:07.519 align:start position:0% +the svm + + +00:34:07.519 --> 00:34:11.109 align:start position:0% +the svm +literature<00:34:08.960> in<00:34:09.119> order<00:34:09.520> to + +00:34:11.109 --> 00:34:11.119 align:start position:0% +literature in order to + + +00:34:11.119 --> 00:34:13.349 align:start position:0% +literature in order to +see<00:34:11.440> that<00:34:11.679> justification<00:34:12.480> for<00:34:12.720> the<00:34:13.040> hinge + +00:34:13.349 --> 00:34:13.359 align:start position:0% +see that justification for the hinge + + +00:34:13.359 --> 00:34:17.270 align:start position:0% +see that justification for the hinge +loss<00:34:13.679> using<00:34:14.000> margin<00:34:14.399> maximization + +00:34:17.270 --> 00:34:17.280 align:start position:0% + + + +00:34:17.280 --> 00:34:20.710 align:start position:0% + +okay<00:34:17.599> so<00:34:17.919> i<00:34:18.480> see<00:34:18.720> there's<00:34:19.040> a<00:34:19.200> question<00:34:20.000> let<00:34:20.159> me + +00:34:20.710 --> 00:34:20.720 align:start position:0% +okay so i see there's a question let me + + +00:34:20.720 --> 00:34:24.629 align:start position:0% +okay so i see there's a question let me +see<00:34:22.000> is<00:34:22.159> the<00:34:22.399> quadratic<00:34:23.040> loss<00:34:23.679> equation + +00:34:24.629 --> 00:34:24.639 align:start position:0% +see is the quadratic loss equation + + +00:34:24.639 --> 00:34:27.270 align:start position:0% +see is the quadratic loss equation +in<00:34:24.800> the<00:34:24.879> quadratic<00:34:25.520> loss<00:34:25.760> equation<00:34:26.399> why<00:34:26.720> do<00:34:26.960> we + +00:34:27.270 --> 00:34:27.280 align:start position:0% +in the quadratic loss equation why do we + + +00:34:27.280 --> 00:34:28.069 align:start position:0% +in the quadratic loss equation why do we +have + +00:34:28.069 --> 00:34:28.079 align:start position:0% +have + + +00:34:28.079 --> 00:34:31.030 align:start position:0% +have +the<00:34:28.320> one-half<00:34:29.280> good<00:34:29.599> question<00:34:30.480> it<00:34:30.639> is<00:34:30.800> just + +00:34:31.030 --> 00:34:31.040 align:start position:0% +the one-half good question it is just + + +00:34:31.040 --> 00:34:33.589 align:start position:0% +the one-half good question it is just +for<00:34:31.280> mathematical<00:34:32.079> convenience<00:34:32.800> because + +00:34:33.589 --> 00:34:33.599 align:start position:0% +for mathematical convenience because + + +00:34:33.599 --> 00:34:36.710 align:start position:0% +for mathematical convenience because +when<00:34:33.760> you<00:34:34.000> take<00:34:34.240> the<00:34:34.399> derivative<00:34:35.919> here + +00:34:36.710 --> 00:34:36.720 align:start position:0% +when you take the derivative here + + +00:34:36.720 --> 00:34:39.909 align:start position:0% +when you take the derivative here +so<00:34:36.960> this<00:34:37.280> 2<00:34:37.679> comes<00:34:38.079> down<00:34:38.399> here<00:34:38.720> so<00:34:39.119> 2<00:34:39.359> and<00:34:39.679> 1 + +00:34:39.909 --> 00:34:39.919 align:start position:0% +so this 2 comes down here so 2 and 1 + + +00:34:39.919 --> 00:34:41.190 align:start position:0% +so this 2 comes down here so 2 and 1 +over<00:34:40.159> 2<00:34:40.399> cancel + +00:34:41.190 --> 00:34:41.200 align:start position:0% +over 2 cancel + + +00:34:41.200 --> 00:34:43.190 align:start position:0% +over 2 cancel +cancels<00:34:41.760> out<00:34:42.000> so<00:34:42.159> you<00:34:42.320> don't<00:34:42.480> actually<00:34:42.879> need + +00:34:43.190 --> 00:34:43.200 align:start position:0% +cancels out so you don't actually need + + +00:34:43.200 --> 00:34:45.430 align:start position:0% +cancels out so you don't actually need +like<00:34:43.520> you<00:34:43.679> know<00:34:44.159> this<00:34:44.399> one<00:34:44.639> over<00:34:44.960> half + +00:34:45.430 --> 00:34:45.440 align:start position:0% +like you know this one over half + + +00:34:45.440 --> 00:34:48.470 align:start position:0% +like you know this one over half +it's<00:34:45.520> just<00:34:45.839> a<00:34:45.919> linear<00:34:47.280> constant<00:34:47.760> scaling<00:34:48.320> it + +00:34:48.470 --> 00:34:48.480 align:start position:0% +it's just a linear constant scaling it + + +00:34:48.480 --> 00:34:49.750 align:start position:0% +it's just a linear constant scaling it +doesn't<00:34:48.720> change<00:34:49.040> anything + +00:34:49.750 --> 00:34:49.760 align:start position:0% +doesn't change anything + + +00:34:49.760 --> 00:34:52.550 align:start position:0% +doesn't change anything +but<00:34:50.000> it<00:34:50.159> is<00:34:50.639> convenient<00:34:51.119> to<00:34:51.280> not<00:34:51.520> carry<00:34:52.240> these + +00:34:52.550 --> 00:34:52.560 align:start position:0% +but it is convenient to not carry these + + +00:34:52.560 --> 00:34:54.230 align:start position:0% +but it is convenient to not carry these +constants<00:34:53.119> when<00:34:53.280> you<00:34:53.440> are<00:34:53.520> looking<00:34:53.839> at<00:34:54.000> the + +00:34:54.230 --> 00:34:54.240 align:start position:0% +constants when you are looking at the + + +00:34:54.240 --> 00:34:54.950 align:start position:0% +constants when you are looking at the +gradient<00:34:54.720> of + +00:34:54.950 --> 00:34:54.960 align:start position:0% +gradient of + + +00:34:54.960 --> 00:34:58.870 align:start position:0% +gradient of +loss<00:34:56.480> good<00:34:56.720> question + +00:34:58.870 --> 00:34:58.880 align:start position:0% +loss good question + + +00:34:58.880 --> 00:35:01.030 align:start position:0% +loss good question +all<00:34:59.040> right<00:34:59.200> so<00:34:59.839> regression<00:35:00.800> in<00:35:00.880> the + +00:35:01.030 --> 00:35:01.040 align:start position:0% +all right so regression in the + + +00:35:01.040 --> 00:35:02.550 align:start position:0% +all right so regression in the +regression<00:35:01.440> we<00:35:01.599> have<00:35:01.760> quadratic + +00:35:02.550 --> 00:35:02.560 align:start position:0% +regression we have quadratic + + +00:35:02.560 --> 00:35:04.710 align:start position:0% +regression we have quadratic +loss<00:35:03.119> in<00:35:03.359> classification<00:35:04.160> we<00:35:04.320> talked<00:35:04.560> about + +00:35:04.710 --> 00:35:04.720 align:start position:0% +loss in classification we talked about + + +00:35:04.720 --> 00:35:05.910 align:start position:0% +loss in classification we talked about +the<00:35:04.800> hinge<00:35:05.119> loss + +00:35:05.910 --> 00:35:05.920 align:start position:0% +the hinge loss + + +00:35:05.920 --> 00:35:08.630 align:start position:0% +the hinge loss +and<00:35:06.320> let's<00:35:06.880> talk<00:35:07.119> about<00:35:07.440> the<00:35:07.680> cross<00:35:08.000> entropy + +00:35:08.630 --> 00:35:08.640 align:start position:0% +and let's talk about the cross entropy + + +00:35:08.640 --> 00:35:10.710 align:start position:0% +and let's talk about the cross entropy +loss<00:35:08.960> as<00:35:09.200> well + +00:35:10.710 --> 00:35:10.720 align:start position:0% +loss as well + + +00:35:10.720 --> 00:35:13.109 align:start position:0% +loss as well +which<00:35:10.960> is<00:35:11.200> another<00:35:11.760> very<00:35:12.079> popular<00:35:12.640> loss<00:35:12.960> in + +00:35:13.109 --> 00:35:13.119 align:start position:0% +which is another very popular loss in + + +00:35:13.119 --> 00:35:14.870 align:start position:0% +which is another very popular loss in +the<00:35:13.359> classification + +00:35:14.870 --> 00:35:14.880 align:start position:0% +the classification + + +00:35:14.880 --> 00:35:18.230 align:start position:0% +the classification +problem<00:35:16.480> cross + +00:35:18.230 --> 00:35:18.240 align:start position:0% +problem cross + + +00:35:18.240 --> 00:35:19.380 align:start position:0% +problem cross +entropy + +00:35:19.380 --> 00:35:19.390 align:start position:0% +entropy + + +00:35:19.390 --> 00:35:22.069 align:start position:0% +entropy +[Music] + +00:35:22.069 --> 00:35:22.079 align:start position:0% +[Music] + + +00:35:22.079 --> 00:35:27.030 align:start position:0% +[Music] +loss<00:35:23.839> okay<00:35:24.240> so<00:35:24.560> here + +00:35:27.030 --> 00:35:27.040 align:start position:0% + + + +00:35:27.040 --> 00:35:29.750 align:start position:0% + +or<00:35:27.440> analysis<00:35:28.079> is<00:35:28.240> based<00:35:28.560> on<00:35:29.280> maximum + +00:35:29.750 --> 00:35:29.760 align:start position:0% +or analysis is based on maximum + + +00:35:29.760 --> 00:35:32.390 align:start position:0% +or analysis is based on maximum +likelihood<00:35:30.240> optimization + +00:35:32.390 --> 00:35:32.400 align:start position:0% +likelihood optimization + + +00:35:32.400 --> 00:35:35.670 align:start position:0% +likelihood optimization +based<00:35:32.720> on<00:35:33.440> maximum<00:35:34.000> likelihood + +00:35:35.670 --> 00:35:35.680 align:start position:0% +based on maximum likelihood + + +00:35:35.680 --> 00:35:39.829 align:start position:0% +based on maximum likelihood +optimization<00:35:38.240> okay<00:35:38.560> so<00:35:38.800> first<00:35:39.119> we<00:35:39.280> need<00:35:39.440> to + +00:35:39.829 --> 00:35:39.839 align:start position:0% +optimization okay so first we need to + + +00:35:39.839 --> 00:35:43.589 align:start position:0% +optimization okay so first we need to +convert<00:35:41.200> these<00:35:41.920> outputs<00:35:42.320> of<00:35:42.400> my<00:35:42.640> network<00:35:43.040> fw + +00:35:43.589 --> 00:35:43.599 align:start position:0% +convert these outputs of my network fw + + +00:35:43.599 --> 00:35:44.630 align:start position:0% +convert these outputs of my network fw +of<00:35:43.920> x2 + +00:35:44.630 --> 00:35:44.640 align:start position:0% +of x2 + + +00:35:44.640 --> 00:35:47.109 align:start position:0% +of x2 +probabilities<00:35:45.839> and<00:35:46.079> in<00:35:46.160> order<00:35:46.480> to<00:35:46.640> do<00:35:46.800> that<00:35:46.960> we + +00:35:47.109 --> 00:35:47.119 align:start position:0% +probabilities and in order to do that we + + +00:35:47.119 --> 00:35:50.150 align:start position:0% +probabilities and in order to do that we +are<00:35:47.200> going<00:35:47.440> to<00:35:47.680> use<00:35:48.000> a<00:35:48.240> sigmoid<00:35:48.800> function + +00:35:50.150 --> 00:35:50.160 align:start position:0% +are going to use a sigmoid function + + +00:35:50.160 --> 00:35:53.670 align:start position:0% +are going to use a sigmoid function +sigmoid<00:35:50.720> is<00:35:50.800> a<00:35:50.960> function<00:35:51.599> that<00:35:52.079> maps + +00:35:53.670 --> 00:35:53.680 align:start position:0% +sigmoid is a function that maps + + +00:35:53.680 --> 00:35:58.870 align:start position:0% +sigmoid is a function that maps +the<00:35:53.839> real<00:35:54.400> to<00:35:55.920> a<00:35:56.079> number<00:35:56.480> between<00:35:56.880> 0<00:35:57.200> and<00:35:57.359> 1 + +00:35:58.870 --> 00:35:58.880 align:start position:0% +the real to a number between 0 and 1 + + +00:35:58.880 --> 00:36:01.589 align:start position:0% +the real to a number between 0 and 1 +so<00:35:59.040> this<00:35:59.200> is<00:35:59.359> the<00:35:59.440> sigmoid<00:36:00.000> function<00:36:00.720> g<00:36:01.040> of<00:36:01.280> z + +00:36:01.589 --> 00:36:01.599 align:start position:0% +so this is the sigmoid function g of z + + +00:36:01.599 --> 00:36:02.069 align:start position:0% +so this is the sigmoid function g of z +is + +00:36:02.069 --> 00:36:02.079 align:start position:0% +is + + +00:36:02.079 --> 00:36:05.910 align:start position:0% +is +1<00:36:02.320> over<00:36:03.839> 1<00:36:04.079> plus<00:36:04.320> e<00:36:04.480> to<00:36:04.640> the<00:36:04.720> minus<00:36:05.119> c + +00:36:05.910 --> 00:36:05.920 align:start position:0% +1 over 1 plus e to the minus c + + +00:36:05.920 --> 00:36:09.109 align:start position:0% +1 over 1 plus e to the minus c +if<00:36:06.800> z<00:36:07.200> is<00:36:07.440> like<00:36:08.079> plus<00:36:08.400> infinity + +00:36:09.109 --> 00:36:09.119 align:start position:0% +if z is like plus infinity + + +00:36:09.119 --> 00:36:12.550 align:start position:0% +if z is like plus infinity +this<00:36:09.359> function<00:36:09.839> approaches<00:36:10.320> 1<00:36:11.040> if<00:36:11.680> z<00:36:12.079> is + +00:36:12.550 --> 00:36:12.560 align:start position:0% +this function approaches 1 if z is + + +00:36:12.560 --> 00:36:13.990 align:start position:0% +this function approaches 1 if z is +minus<00:36:12.960> infinity<00:36:13.440> this<00:36:13.599> function<00:36:13.920> is + +00:36:13.990 --> 00:36:14.000 align:start position:0% +minus infinity this function is + + +00:36:14.000 --> 00:36:16.150 align:start position:0% +minus infinity this function is +approaches<00:36:14.480> zero<00:36:15.359> and<00:36:15.599> at<00:36:15.839> the + +00:36:16.150 --> 00:36:16.160 align:start position:0% +approaches zero and at the + + +00:36:16.160 --> 00:36:19.750 align:start position:0% +approaches zero and at the +zero<00:36:16.720> it<00:36:16.960> is<00:36:17.440> in<00:36:17.599> fact<00:36:18.000> one<00:36:18.240> over<00:36:18.560> half + +00:36:19.750 --> 00:36:19.760 align:start position:0% +zero it is in fact one over half + + +00:36:19.760 --> 00:36:21.589 align:start position:0% +zero it is in fact one over half +right<00:36:20.000> so<00:36:20.160> we<00:36:20.320> use<00:36:20.480> this<00:36:20.720> function<00:36:21.119> in<00:36:21.200> order + +00:36:21.589 --> 00:36:21.599 align:start position:0% +right so we use this function in order + + +00:36:21.599 --> 00:36:23.990 align:start position:0% +right so we use this function in order +to<00:36:21.839> turn<00:36:22.480> this<00:36:22.800> f<00:36:22.960> w<00:36:23.359> of<00:36:23.599> x + +00:36:23.990 --> 00:36:24.000 align:start position:0% +to turn this f w of x + + +00:36:24.000 --> 00:36:26.790 align:start position:0% +to turn this f w of x +which<00:36:24.160> is<00:36:24.320> the<00:36:24.400> output<00:36:24.880> of<00:36:24.960> my<00:36:25.119> network<00:36:25.520> into<00:36:26.480> a + +00:36:26.790 --> 00:36:26.800 align:start position:0% +which is the output of my network into a + + +00:36:26.800 --> 00:36:27.910 align:start position:0% +which is the output of my network into a +probability + +00:36:27.910 --> 00:36:27.920 align:start position:0% +probability + + +00:36:27.920 --> 00:36:29.270 align:start position:0% +probability +and<00:36:28.079> then<00:36:28.320> i'm<00:36:28.480> going<00:36:28.640> to<00:36:28.800> use<00:36:29.040> this + +00:36:29.270 --> 00:36:29.280 align:start position:0% +and then i'm going to use this + + +00:36:29.280 --> 00:36:31.670 align:start position:0% +and then i'm going to use this +probability<00:36:30.240> if<00:36:30.480> this<00:36:30.640> probability<00:36:31.200> is<00:36:31.280> above + +00:36:31.670 --> 00:36:31.680 align:start position:0% +probability if this probability is above + + +00:36:31.680 --> 00:36:32.870 align:start position:0% +probability if this probability is above +half + +00:36:32.870 --> 00:36:32.880 align:start position:0% +half + + +00:36:32.880 --> 00:36:35.990 align:start position:0% +half +in<00:36:33.040> other<00:36:33.200> words<00:36:33.520> if<00:36:33.760> fw<00:36:34.720> of<00:36:34.880> x<00:36:35.200> is<00:36:35.359> positive + +00:36:35.990 --> 00:36:36.000 align:start position:0% +in other words if fw of x is positive + + +00:36:36.000 --> 00:36:37.510 align:start position:0% +in other words if fw of x is positive +then<00:36:36.240> i'm<00:36:36.320> going<00:36:36.400> to<00:36:36.560> predict<00:36:36.960> a<00:36:37.040> positive + +00:36:37.510 --> 00:36:37.520 align:start position:0% +then i'm going to predict a positive + + +00:36:37.520 --> 00:36:38.150 align:start position:0% +then i'm going to predict a positive +label + +00:36:38.150 --> 00:36:38.160 align:start position:0% +label + + +00:36:38.160 --> 00:36:39.510 align:start position:0% +label +otherwise<00:36:38.720> i'm<00:36:38.800> going<00:36:38.880> to<00:36:39.040> predict<00:36:39.440> a + +00:36:39.510 --> 00:36:39.520 align:start position:0% +otherwise i'm going to predict a + + +00:36:39.520 --> 00:36:41.430 align:start position:0% +otherwise i'm going to predict a +negative<00:36:40.000> label<00:36:41.040> so + +00:36:41.430 --> 00:36:41.440 align:start position:0% +negative label so + + +00:36:41.440 --> 00:36:45.270 align:start position:0% +negative label so +we<00:36:41.680> basically<00:36:42.240> say<00:36:43.119> we<00:36:43.280> have<00:36:43.520> this<00:36:44.079> uh + +00:36:45.270 --> 00:36:45.280 align:start position:0% +we basically say we have this uh + + +00:36:45.280 --> 00:36:48.310 align:start position:0% +we basically say we have this uh +generative<00:36:45.920> model<00:36:46.480> for<00:36:46.880> labels<00:36:47.680> the<00:36:47.920> property + +00:36:48.310 --> 00:36:48.320 align:start position:0% +generative model for labels the property + + +00:36:48.320 --> 00:36:48.950 align:start position:0% +generative model for labels the property +of<00:36:48.480> y + +00:36:48.950 --> 00:36:48.960 align:start position:0% +of y + + +00:36:48.960 --> 00:36:52.870 align:start position:0% +of y +given<00:36:49.920> x<00:36:50.320> and<00:36:50.480> a<00:36:50.640> particular<00:36:51.200> w + +00:36:52.870 --> 00:36:52.880 align:start position:0% +given x and a particular w + + +00:36:52.880 --> 00:36:55.990 align:start position:0% +given x and a particular w +is<00:36:53.520> g<00:36:54.000> of<00:36:54.880> if + +00:36:55.990 --> 00:36:56.000 align:start position:0% +is g of if + + +00:36:56.000 --> 00:37:00.950 align:start position:0% +is g of if +w<00:36:56.480> of<00:36:56.640> x<00:36:58.800> and<00:36:58.960> the<00:36:59.119> probability<00:36:59.760> of<00:36:59.920> y + +00:37:00.950 --> 00:37:00.960 align:start position:0% +w of x and the probability of y + + +00:37:00.960 --> 00:37:04.870 align:start position:0% +w of x and the probability of y +to<00:37:01.119> be<00:37:01.359> 0<00:37:02.240> is<00:37:02.480> just<00:37:03.200> the<00:37:03.440> complement<00:37:04.000> of<00:37:04.160> this + +00:37:04.870 --> 00:37:04.880 align:start position:0% +to be 0 is just the complement of this + + +00:37:04.880 --> 00:37:07.990 align:start position:0% +to be 0 is just the complement of this +1<00:37:05.119> minus<00:37:05.520> g<00:37:05.839> of<00:37:06.079> f<00:37:06.240> w<00:37:06.560> of<00:37:06.800> x<00:37:07.680> that's + +00:37:07.990 --> 00:37:08.000 align:start position:0% +1 minus g of f w of x that's + + +00:37:08.000 --> 00:37:11.510 align:start position:0% +1 minus g of f w of x that's +it<00:37:09.920> so<00:37:10.079> in<00:37:10.640> other<00:37:10.880> words<00:37:11.200> you + +00:37:11.510 --> 00:37:11.520 align:start position:0% +it so in other words you + + +00:37:11.520 --> 00:37:15.990 align:start position:0% +it so in other words you +you<00:37:11.680> can<00:37:12.000> combine<00:37:12.480> both<00:37:12.720> of<00:37:12.880> these<00:37:13.520> equations + +00:37:15.990 --> 00:37:16.000 align:start position:0% +you can combine both of these equations + + +00:37:16.000 --> 00:37:18.950 align:start position:0% +you can combine both of these equations +we<00:37:16.160> can<00:37:16.400> write<00:37:16.720> probability<00:37:17.119> of<00:37:17.280> y<00:37:17.760> given<00:37:18.400> x + +00:37:18.950 --> 00:37:18.960 align:start position:0% +we can write probability of y given x + + +00:37:18.960 --> 00:37:20.150 align:start position:0% +we can write probability of y given x +and<00:37:19.119> w + +00:37:20.150 --> 00:37:20.160 align:start position:0% +and w + + +00:37:20.160 --> 00:37:23.589 align:start position:0% +and w +is<00:37:20.400> nothing<00:37:20.839> but<00:37:22.160> g<00:37:22.640> of<00:37:22.960> f<00:37:23.200> w + +00:37:23.589 --> 00:37:23.599 align:start position:0% +is nothing but g of f w + + +00:37:23.599 --> 00:37:27.910 align:start position:0% +is nothing but g of f w +of<00:37:23.760> x<00:37:24.240> to<00:37:24.400> the<00:37:24.560> power<00:37:24.960> of<00:37:25.119> y<00:37:26.000> times<00:37:26.320> 1<00:37:26.560> minus<00:37:26.960> g + +00:37:27.910 --> 00:37:27.920 align:start position:0% +of x to the power of y times 1 minus g + + +00:37:27.920 --> 00:37:31.829 align:start position:0% +of x to the power of y times 1 minus g +of<00:37:28.240> f<00:37:28.480> of<00:37:28.560> w<00:37:28.960> of<00:37:29.119> x<00:37:29.599> to<00:37:29.760> the<00:37:29.920> power<00:37:30.240> 1<00:37:30.480> minus<00:37:30.800> y + +00:37:31.829 --> 00:37:31.839 align:start position:0% +of f of w of x to the power 1 minus y + + +00:37:31.839 --> 00:37:35.030 align:start position:0% +of f of w of x to the power 1 minus y +why<00:37:32.079> is<00:37:32.400> this<00:37:32.640> the<00:37:32.800> case<00:37:33.200> because<00:37:33.520> when<00:37:33.760> y<00:37:33.920> is<00:37:34.079> 0 + +00:37:35.030 --> 00:37:35.040 align:start position:0% +why is this the case because when y is 0 + + +00:37:35.040 --> 00:37:37.910 align:start position:0% +why is this the case because when y is 0 +then<00:37:36.000> this<00:37:36.160> time<00:37:36.400> goes<00:37:36.720> away<00:37:37.040> so<00:37:37.359> i<00:37:37.440> got<00:37:37.680> 1 + +00:37:37.910 --> 00:37:37.920 align:start position:0% +then this time goes away so i got 1 + + +00:37:37.920 --> 00:37:38.550 align:start position:0% +then this time goes away so i got 1 +minus<00:37:38.320> g + +00:37:38.550 --> 00:37:38.560 align:start position:0% +minus g + + +00:37:38.560 --> 00:37:41.589 align:start position:0% +minus g +of<00:37:38.720> f<00:37:38.800> w<00:37:39.119> of<00:37:39.280> x<00:37:39.599> then<00:37:39.839> y<00:37:40.000> is<00:37:40.160> 1<00:37:40.800> the<00:37:41.040> second<00:37:41.359> time + +00:37:41.589 --> 00:37:41.599 align:start position:0% +of f w of x then y is 1 the second time + + +00:37:41.599 --> 00:37:42.550 align:start position:0% +of f w of x then y is 1 the second time +goes<00:37:41.839> away + +00:37:42.550 --> 00:37:42.560 align:start position:0% +goes away + + +00:37:42.560 --> 00:37:45.829 align:start position:0% +goes away +and<00:37:42.880> the<00:37:43.040> first<00:37:43.280> term<00:37:43.599> is<00:37:44.160> g<00:37:44.400> of<00:37:44.720> f<00:37:44.960> w<00:37:45.280> of<00:37:45.440> x<00:37:45.520> so<00:37:45.680> i + +00:37:45.829 --> 00:37:45.839 align:start position:0% +and the first term is g of f w of x so i + + +00:37:45.839 --> 00:37:47.670 align:start position:0% +and the first term is g of f w of x so i +basically<00:37:46.240> summarize<00:37:46.720> these<00:37:46.880> two<00:37:47.119> equations + +00:37:47.670 --> 00:37:47.680 align:start position:0% +basically summarize these two equations + + +00:37:47.680 --> 00:37:50.230 align:start position:0% +basically summarize these two equations +in<00:37:47.839> one<00:37:48.240> equation<00:37:48.839> here + +00:37:50.230 --> 00:37:50.240 align:start position:0% +in one equation here + + +00:37:50.240 --> 00:37:52.230 align:start position:0% +in one equation here +all<00:37:50.400> right<00:37:50.640> so<00:37:51.040> then<00:37:51.440> we<00:37:51.680> just<00:37:51.920> write<00:37:52.160> the + +00:37:52.230 --> 00:37:52.240 align:start position:0% +all right so then we just write the + + +00:37:52.240 --> 00:37:54.790 align:start position:0% +all right so then we just write the +maximum<00:37:52.640> likelihood<00:37:53.280> optimization + +00:37:54.790 --> 00:37:54.800 align:start position:0% +maximum likelihood optimization + + +00:37:54.800 --> 00:37:57.990 align:start position:0% +maximum likelihood optimization +for<00:37:55.040> this + +00:37:57.990 --> 00:37:58.000 align:start position:0% + + + +00:37:58.000 --> 00:38:00.790 align:start position:0% + +the<00:37:58.160> maximum<00:37:58.640> likelihood<00:37:59.520> optimization + +00:38:00.790 --> 00:38:00.800 align:start position:0% +the maximum likelihood optimization + + +00:38:00.800 --> 00:38:02.150 align:start position:0% +the maximum likelihood optimization +would<00:38:01.119> maximize + +00:38:02.150 --> 00:38:02.160 align:start position:0% +would maximize + + +00:38:02.160 --> 00:38:05.109 align:start position:0% +would maximize +over<00:38:02.480> my<00:38:02.720> model<00:38:03.040> parameters<00:38:03.920> the<00:38:04.160> likelihood + +00:38:05.109 --> 00:38:05.119 align:start position:0% +over my model parameters the likelihood + + +00:38:05.119 --> 00:38:07.030 align:start position:0% +over my model parameters the likelihood +of<00:38:05.359> observing<00:38:06.000> this<00:38:06.240> data<00:38:06.640> set + +00:38:07.030 --> 00:38:07.040 align:start position:0% +of observing this data set + + +00:38:07.040 --> 00:38:09.510 align:start position:0% +of observing this data set +from<00:38:07.280> this<00:38:07.520> model<00:38:08.160> and<00:38:08.320> we<00:38:08.480> are<00:38:08.560> also<00:38:08.880> assuming + +00:38:09.510 --> 00:38:09.520 align:start position:0% +from this model and we are also assuming + + +00:38:09.520 --> 00:38:10.390 align:start position:0% +from this model and we are also assuming +independence + +00:38:10.390 --> 00:38:10.400 align:start position:0% +independence + + +00:38:10.400 --> 00:38:12.069 align:start position:0% +independence +among<00:38:10.800> samples<00:38:11.200> that<00:38:11.280> we<00:38:11.440> are<00:38:11.599> drawing<00:38:11.920> from + +00:38:12.069 --> 00:38:12.079 align:start position:0% +among samples that we are drawing from + + +00:38:12.079 --> 00:38:13.829 align:start position:0% +among samples that we are drawing from +this<00:38:12.320> model<00:38:12.720> so<00:38:13.280> the + +00:38:13.829 --> 00:38:13.839 align:start position:0% +this model so the + + +00:38:13.839 --> 00:38:15.750 align:start position:0% +this model so the +probability<00:38:14.400> of<00:38:14.640> observing<00:38:14.960> the<00:38:15.119> status<00:38:15.599> will + +00:38:15.750 --> 00:38:15.760 align:start position:0% +probability of observing the status will + + +00:38:15.760 --> 00:38:17.270 align:start position:0% +probability of observing the status will +be<00:38:15.920> the<00:38:16.079> product<00:38:16.640> of + +00:38:17.270 --> 00:38:17.280 align:start position:0% +be the product of + + +00:38:17.280 --> 00:38:20.470 align:start position:0% +be the product of +observing<00:38:18.480> individual<00:38:19.760> samples<00:38:20.240> of<00:38:20.320> the + +00:38:20.470 --> 00:38:20.480 align:start position:0% +observing individual samples of the + + +00:38:20.480 --> 00:38:21.270 align:start position:0% +observing individual samples of the +dataset + +00:38:21.270 --> 00:38:21.280 align:start position:0% +dataset + + +00:38:21.280 --> 00:38:26.470 align:start position:0% +dataset +from<00:38:21.599> this<00:38:21.920> particular<00:38:23.839> model + +00:38:26.470 --> 00:38:26.480 align:start position:0% + + + +00:38:26.480 --> 00:38:28.950 align:start position:0% + +so<00:38:26.960> this<00:38:27.200> is<00:38:27.280> the<00:38:27.920> optimization<00:38:28.560> for<00:38:28.720> the + +00:38:28.950 --> 00:38:28.960 align:start position:0% +so this is the optimization for the + + +00:38:28.960 --> 00:38:31.510 align:start position:0% +so this is the optimization for the +maximum<00:38:29.359> likelihood + +00:38:31.510 --> 00:38:31.520 align:start position:0% +maximum likelihood + + +00:38:31.520 --> 00:38:34.310 align:start position:0% +maximum likelihood +because<00:38:31.920> log<00:38:32.400> is<00:38:32.720> a<00:38:33.599> monotonically + +00:38:34.310 --> 00:38:34.320 align:start position:0% +because log is a monotonically + + +00:38:34.320 --> 00:38:35.910 align:start position:0% +because log is a monotonically +increasing<00:38:34.800> function<00:38:35.200> without<00:38:35.520> loss<00:38:35.760> of + +00:38:35.910 --> 00:38:35.920 align:start position:0% +increasing function without loss of + + +00:38:35.920 --> 00:38:37.910 align:start position:0% +increasing function without loss of +generality<00:38:36.640> this<00:38:36.880> is<00:38:37.040> equivalent + +00:38:37.910 --> 00:38:37.920 align:start position:0% +generality this is equivalent + + +00:38:37.920 --> 00:38:39.589 align:start position:0% +generality this is equivalent +if<00:38:38.079> i<00:38:38.320> apply<00:38:38.640> a<00:38:38.800> log<00:38:39.040> function<00:38:39.359> to<00:38:39.520> the + +00:38:39.589 --> 00:38:39.599 align:start position:0% +if i apply a log function to the + + +00:38:39.599 --> 00:38:41.270 align:start position:0% +if i apply a log function to the +objective<00:38:40.079> function<00:38:40.560> because<00:38:40.800> i<00:38:40.880> don't<00:38:41.040> like + +00:38:41.270 --> 00:38:41.280 align:start position:0% +objective function because i don't like + + +00:38:41.280 --> 00:38:42.310 align:start position:0% +objective function because i don't like +the<00:38:41.440> product + +00:38:42.310 --> 00:38:42.320 align:start position:0% +the product + + +00:38:42.320 --> 00:38:45.109 align:start position:0% +the product +my<00:38:42.560> product<00:38:42.960> of<00:38:43.119> a<00:38:43.200> lot<00:38:43.440> of<00:38:44.000> small<00:38:44.720> numbers + +00:38:45.109 --> 00:38:45.119 align:start position:0% +my product of a lot of small numbers + + +00:38:45.119 --> 00:38:46.790 align:start position:0% +my product of a lot of small numbers +these<00:38:45.280> properties<00:38:45.760> are<00:38:45.839> between<00:38:46.160> 0<00:38:46.400> and<00:38:46.560> 1 + +00:38:46.790 --> 00:38:46.800 align:start position:0% +these properties are between 0 and 1 + + +00:38:46.800 --> 00:38:47.109 align:start position:0% +these properties are between 0 and 1 +then + +00:38:47.109 --> 00:38:47.119 align:start position:0% +then + + +00:38:47.119 --> 00:38:49.109 align:start position:0% +then +n<00:38:47.280> is<00:38:47.440> large<00:38:47.839> it<00:38:48.000> may<00:38:48.480> introduce<00:38:48.960> some + +00:38:49.109 --> 00:38:49.119 align:start position:0% +n is large it may introduce some + + +00:38:49.119 --> 00:38:50.470 align:start position:0% +n is large it may introduce some +numerical + +00:38:50.470 --> 00:38:50.480 align:start position:0% +numerical + + +00:38:50.480 --> 00:38:53.510 align:start position:0% +numerical +issues<00:38:51.200> so<00:38:51.359> i<00:38:51.680> apply<00:38:52.160> a<00:38:52.320> log<00:38:52.640> function + +00:38:53.510 --> 00:38:53.520 align:start position:0% +issues so i apply a log function + + +00:38:53.520 --> 00:38:56.790 align:start position:0% +issues so i apply a log function +and<00:38:53.680> i<00:38:53.839> get<00:38:54.960> sum<00:38:55.440> 1<00:38:55.680> over<00:38:56.000> n + +00:38:56.790 --> 00:38:56.800 align:start position:0% +and i get sum 1 over n + + +00:38:56.800 --> 00:39:01.109 align:start position:0% +and i get sum 1 over n +log<00:38:57.920> of<00:38:58.560> probability<00:38:59.359> of<00:38:59.599> y<00:38:59.920> i + +00:39:01.109 --> 00:39:01.119 align:start position:0% +log of probability of y i + + +00:39:01.119 --> 00:39:05.430 align:start position:0% +log of probability of y i +given<00:39:01.520> x<00:39:01.680> i<00:39:01.920> and<00:39:02.000> w<00:39:03.599> and<00:39:03.839> i<00:39:03.920> have<00:39:04.160> already + +00:39:05.430 --> 00:39:05.440 align:start position:0% +given x i and w and i have already + + +00:39:05.440 --> 00:39:07.109 align:start position:0% +given x i and w and i have already +have<00:39:05.680> a<00:39:05.760> formula<00:39:06.320> for<00:39:06.480> this<00:39:06.640> so<00:39:06.800> i'm<00:39:06.880> just + +00:39:07.109 --> 00:39:07.119 align:start position:0% +have a formula for this so i'm just + + +00:39:07.119 --> 00:39:09.349 align:start position:0% +have a formula for this so i'm just +going<00:39:07.200> to<00:39:07.359> replace<00:39:07.839> it<00:39:08.000> here + +00:39:09.349 --> 00:39:09.359 align:start position:0% +going to replace it here + + +00:39:09.359 --> 00:39:12.470 align:start position:0% +going to replace it here +so<00:39:10.160> log<00:39:10.640> of<00:39:10.800> this<00:39:11.119> term<00:39:11.599> will<00:39:11.839> be<00:39:12.079> y + +00:39:12.470 --> 00:39:12.480 align:start position:0% +so log of this term will be y + + +00:39:12.480 --> 00:39:15.589 align:start position:0% +so log of this term will be y +times<00:39:12.880> log<00:39:13.200> of<00:39:13.599> g<00:39:13.839> f<00:39:14.000> w<00:39:14.320> of<00:39:14.480> x<00:39:14.640> plus<00:39:14.880> 1<00:39:15.119> minus + +00:39:15.589 --> 00:39:15.599 align:start position:0% +times log of g f w of x plus 1 minus + + +00:39:15.599 --> 00:39:18.790 align:start position:0% +times log of g f w of x plus 1 minus +y<00:39:15.920> times<00:39:16.240> log<00:39:16.480> of<00:39:16.560> 1<00:39:16.720> minus<00:39:17.040> g<00:39:17.359> f<00:39:17.440> w<00:39:17.839> of<00:39:18.000> x + +00:39:18.790 --> 00:39:18.800 align:start position:0% +y times log of 1 minus g f w of x + + +00:39:18.800 --> 00:39:20.390 align:start position:0% +y times log of 1 minus g f w of x +and<00:39:18.960> i'm<00:39:19.119> going<00:39:19.200> to<00:39:19.359> change<00:39:19.680> this<00:39:19.920> max<00:39:20.160> to<00:39:20.320> a + +00:39:20.390 --> 00:39:20.400 align:start position:0% +and i'm going to change this max to a + + +00:39:20.400 --> 00:39:22.630 align:start position:0% +and i'm going to change this max to a +min<00:39:20.640> and<00:39:20.800> put<00:39:21.040> a<00:39:21.119> minus<00:39:21.520> sign + +00:39:22.630 --> 00:39:22.640 align:start position:0% +min and put a minus sign + + +00:39:22.640 --> 00:39:25.829 align:start position:0% +min and put a minus sign +it<00:39:22.800> is<00:39:23.200> basically<00:39:23.760> the<00:39:23.920> same<00:39:24.160> thing<00:39:25.359> so<00:39:25.599> this + +00:39:25.829 --> 00:39:25.839 align:start position:0% +it is basically the same thing so this + + +00:39:25.839 --> 00:39:26.710 align:start position:0% +it is basically the same thing so this +will<00:39:26.000> be + +00:39:26.710 --> 00:39:26.720 align:start position:0% +will be + + +00:39:26.720 --> 00:39:30.069 align:start position:0% +will be +mean<00:39:27.040> over<00:39:27.359> w<00:39:28.079> minus + +00:39:30.069 --> 00:39:30.079 align:start position:0% +mean over w minus + + +00:39:30.079 --> 00:39:33.109 align:start position:0% +mean over w minus +some<00:39:30.960> 1<00:39:31.200> to<00:39:31.520> n<00:39:32.640> y + +00:39:33.109 --> 00:39:33.119 align:start position:0% +some 1 to n y + + +00:39:33.119 --> 00:39:36.150 align:start position:0% +some 1 to n y +times<00:39:33.599> log<00:39:34.320> of + +00:39:36.150 --> 00:39:36.160 align:start position:0% +times log of + + +00:39:36.160 --> 00:39:39.750 align:start position:0% +times log of +g<00:39:36.720> f<00:39:37.040> w<00:39:37.680> of<00:39:38.000> x<00:39:39.119> plus + +00:39:39.750 --> 00:39:39.760 align:start position:0% +g f w of x plus + + +00:39:39.760 --> 00:39:43.510 align:start position:0% +g f w of x plus +1<00:39:40.000> minus<00:39:40.839> y<00:39:42.560> times + +00:39:43.510 --> 00:39:43.520 align:start position:0% +1 minus y times + + +00:39:43.520 --> 00:39:46.870 align:start position:0% +1 minus y times +log<00:39:44.640> of<00:39:45.040> 1<00:39:45.280> minus<00:39:45.839> g + +00:39:46.870 --> 00:39:46.880 align:start position:0% +log of 1 minus g + + +00:39:46.880 --> 00:39:50.550 align:start position:0% +log of 1 minus g +if<00:39:47.280> w<00:39:47.760> of<00:39:48.000> x<00:39:49.760> and<00:39:49.920> that's<00:39:50.240> it + +00:39:50.550 --> 00:39:50.560 align:start position:0% +if w of x and that's it + + +00:39:50.560 --> 00:39:53.270 align:start position:0% +if w of x and that's it +congratulations<00:39:51.760> this<00:39:52.079> is<00:39:52.400> the<00:39:52.800> cross + +00:39:53.270 --> 00:39:53.280 align:start position:0% +congratulations this is the cross + + +00:39:53.280 --> 00:39:54.710 align:start position:0% +congratulations this is the cross +entropy + +00:39:54.710 --> 00:39:54.720 align:start position:0% +entropy + + +00:39:54.720 --> 00:39:57.829 align:start position:0% +entropy +loss<00:39:54.960> function<00:39:57.440> cross + +00:39:57.829 --> 00:39:57.839 align:start position:0% +loss function cross + + +00:39:57.839 --> 00:40:00.790 align:start position:0% +loss function cross +entropy<00:39:58.960> plus<00:39:59.839> why<00:40:00.079> is<00:40:00.160> it<00:40:00.320> called<00:40:00.560> cross + +00:40:00.790 --> 00:40:00.800 align:start position:0% +entropy plus why is it called cross + + +00:40:00.800 --> 00:40:02.950 align:start position:0% +entropy plus why is it called cross +entropy<00:40:01.200> loss<00:40:01.440> because<00:40:01.760> it<00:40:01.920> is<00:40:02.160> uh + +00:40:02.950 --> 00:40:02.960 align:start position:0% +entropy loss because it is uh + + +00:40:02.960 --> 00:40:05.829 align:start position:0% +entropy loss because it is uh +very<00:40:03.280> similar<00:40:03.920> if<00:40:04.000> you<00:40:04.160> change<00:40:04.400> these<00:40:05.040> y's<00:40:05.359> to + +00:40:05.829 --> 00:40:05.839 align:start position:0% +very similar if you change these y's to + + +00:40:05.839 --> 00:40:07.589 align:start position:0% +very similar if you change these y's to +the<00:40:06.000> same<00:40:06.160> probabilities<00:40:06.960> as<00:40:07.119> the + +00:40:07.589 --> 00:40:07.599 align:start position:0% +the same probabilities as the + + +00:40:07.599 --> 00:40:08.950 align:start position:0% +the same probabilities as the +arguments<00:40:08.079> of<00:40:08.160> the<00:40:08.319> log<00:40:08.480> you'll<00:40:08.640> get<00:40:08.800> the + +00:40:08.950 --> 00:40:08.960 align:start position:0% +arguments of the log you'll get the + + +00:40:08.960 --> 00:40:11.270 align:start position:0% +arguments of the log you'll get the +shannon<00:40:09.359> entropy<00:40:10.480> that's<00:40:10.720> why<00:40:10.960> this<00:40:11.200> is + +00:40:11.270 --> 00:40:11.280 align:start position:0% +shannon entropy that's why this is + + +00:40:11.280 --> 00:40:11.670 align:start position:0% +shannon entropy that's why this is +called + +00:40:11.670 --> 00:40:11.680 align:start position:0% +called + + +00:40:11.680 --> 00:40:15.910 align:start position:0% +called +cross-entropy<00:40:14.000> loss<00:40:14.319> function + +00:40:15.910 --> 00:40:15.920 align:start position:0% +cross-entropy loss function + + +00:40:15.920 --> 00:40:17.990 align:start position:0% +cross-entropy loss function +so<00:40:16.079> i<00:40:16.240> just<00:40:16.400> want<00:40:16.560> to<00:40:16.640> add<00:40:16.880> a<00:40:16.960> note<00:40:17.280> here<00:40:17.760> so + +00:40:17.990 --> 00:40:18.000 align:start position:0% +so i just want to add a note here so + + +00:40:18.000 --> 00:40:20.230 align:start position:0% +so i just want to add a note here so +this<00:40:18.240> is<00:40:18.400> for<00:40:18.720> the + +00:40:20.230 --> 00:40:20.240 align:start position:0% +this is for the + + +00:40:20.240 --> 00:40:22.309 align:start position:0% +this is for the +binary<00:40:20.800> classification<00:40:21.599> problem<00:40:22.000> that<00:40:22.160> we + +00:40:22.309 --> 00:40:22.319 align:start position:0% +binary classification problem that we + + +00:40:22.319 --> 00:40:23.510 align:start position:0% +binary classification problem that we +use<00:40:22.480> the<00:40:22.640> sigmoid + +00:40:23.510 --> 00:40:23.520 align:start position:0% +use the sigmoid + + +00:40:23.520 --> 00:40:26.069 align:start position:0% +use the sigmoid +to<00:40:23.839> turn<00:40:24.160> these<00:40:24.480> numbers<00:40:25.440> uh<00:40:25.760> into + +00:40:26.069 --> 00:40:26.079 align:start position:0% +to turn these numbers uh into + + +00:40:26.079 --> 00:40:27.109 align:start position:0% +to turn these numbers uh into +probabilities + +00:40:27.109 --> 00:40:27.119 align:start position:0% +probabilities + + +00:40:27.119 --> 00:40:29.750 align:start position:0% +probabilities +in<00:40:27.280> the<00:40:27.520> multi-label<00:40:28.319> case<00:40:28.960> uh<00:40:29.359> instead<00:40:29.680> of + +00:40:29.750 --> 00:40:29.760 align:start position:0% +in the multi-label case uh instead of + + +00:40:29.760 --> 00:40:30.950 align:start position:0% +in the multi-label case uh instead of +the<00:40:29.920> sigmoid<00:40:30.400> you + +00:40:30.950 --> 00:40:30.960 align:start position:0% +the sigmoid you + + +00:40:30.960 --> 00:40:34.470 align:start position:0% +the sigmoid you +we<00:40:31.200> often<00:40:31.599> use<00:40:32.240> soft<00:40:32.560> max<00:40:32.960> function<00:40:33.920> so<00:40:34.079> let<00:40:34.319> me + +00:40:34.470 --> 00:40:34.480 align:start position:0% +we often use soft max function so let me + + +00:40:34.480 --> 00:40:37.109 align:start position:0% +we often use soft max function so let me +just<00:40:34.640> add<00:40:34.880> that<00:40:35.119> note + +00:40:37.109 --> 00:40:37.119 align:start position:0% +just add that note + + +00:40:37.119 --> 00:40:43.589 align:start position:0% +just add that note +in<00:40:37.280> the<00:40:41.480> multi-label + +00:40:43.589 --> 00:40:43.599 align:start position:0% +in the multi-label + + +00:40:43.599 --> 00:40:48.230 align:start position:0% +in the multi-label +classification<00:40:44.880> problem + +00:40:48.230 --> 00:40:48.240 align:start position:0% + + + +00:40:48.240 --> 00:40:51.430 align:start position:0% + +we<00:40:48.400> use<00:40:48.720> softmax<00:40:50.480> it's<00:40:50.640> a<00:40:50.720> function<00:40:51.200> very + +00:40:51.430 --> 00:40:51.440 align:start position:0% +we use softmax it's a function very + + +00:40:51.440 --> 00:40:53.030 align:start position:0% +we use softmax it's a function very +similar<00:40:51.760> to<00:40:51.920> the<00:40:52.079> sigmoid + +00:40:53.030 --> 00:40:53.040 align:start position:0% +similar to the sigmoid + + +00:40:53.040 --> 00:40:56.550 align:start position:0% +similar to the sigmoid +uh<00:40:54.240> to<00:40:54.640> turn<00:40:55.440> real<00:40:55.760> numbers<00:40:56.160> to + +00:40:56.550 --> 00:40:56.560 align:start position:0% +uh to turn real numbers to + + +00:40:56.560 --> 00:40:59.589 align:start position:0% +uh to turn real numbers to +probability<00:40:57.200> softmax<00:40:58.480> to + +00:40:59.589 --> 00:40:59.599 align:start position:0% +probability softmax to + + +00:40:59.599 --> 00:41:03.030 align:start position:0% +probability softmax to +turn<00:41:01.119> logits + +00:41:03.030 --> 00:41:03.040 align:start position:0% +turn logits + + +00:41:03.040 --> 00:41:06.309 align:start position:0% +turn logits +to<00:41:04.160> probabilities<00:41:05.440> if<00:41:05.599> you<00:41:05.680> are<00:41:05.760> not<00:41:06.000> happy + +00:41:06.309 --> 00:41:06.319 align:start position:0% +to probabilities if you are not happy + + +00:41:06.319 --> 00:41:07.910 align:start position:0% +to probabilities if you are not happy +with<00:41:06.480> soft<00:41:06.720> max<00:41:07.040> i<00:41:07.119> think<00:41:07.280> you<00:41:07.440> should + +00:41:07.910 --> 00:41:07.920 align:start position:0% +with soft max i think you should + + +00:41:07.920 --> 00:41:12.630 align:start position:0% +with soft max i think you should +take<00:41:08.160> a<00:41:08.240> look<00:41:08.480> into<00:41:08.800> the<00:41:08.960> definition<00:41:10.240> of<00:41:10.400> that + +00:41:12.630 --> 00:41:12.640 align:start position:0% +take a look into the definition of that + + +00:41:12.640 --> 00:41:15.270 align:start position:0% +take a look into the definition of that +okay<00:41:13.040> so<00:41:13.599> uh<00:41:14.000> that's<00:41:14.319> all<00:41:14.560> for<00:41:14.720> the<00:41:14.960> loss + +00:41:15.270 --> 00:41:15.280 align:start position:0% +okay so uh that's all for the loss + + +00:41:15.280 --> 00:41:16.550 align:start position:0% +okay so uh that's all for the loss +functions + +00:41:16.550 --> 00:41:16.560 align:start position:0% +functions + + +00:41:16.560 --> 00:41:19.589 align:start position:0% +functions +i<00:41:17.920> see<00:41:18.240> there's<00:41:18.560> a<00:41:18.720> question + +00:41:19.589 --> 00:41:19.599 align:start position:0% +i see there's a question + + +00:41:19.599 --> 00:41:23.990 align:start position:0% +i see there's a question +let<00:41:19.839> me<00:41:20.000> take<00:41:20.240> that + +00:41:23.990 --> 00:41:24.000 align:start position:0% + + + +00:41:24.000 --> 00:41:25.910 align:start position:0% + +so<00:41:24.160> the<00:41:24.319> question<00:41:24.800> isn't<00:41:25.119> the<00:41:25.280> likelihood + +00:41:25.910 --> 00:41:25.920 align:start position:0% +so the question isn't the likelihood + + +00:41:25.920 --> 00:41:27.589 align:start position:0% +so the question isn't the likelihood +function<00:41:26.480> probability<00:41:27.280> of + +00:41:27.589 --> 00:41:27.599 align:start position:0% +function probability of + + +00:41:27.599 --> 00:41:30.950 align:start position:0% +function probability of +uh<00:41:28.480> data<00:41:29.119> given<00:41:29.599> x<00:41:30.400> uh + +00:41:30.950 --> 00:41:30.960 align:start position:0% +uh data given x uh + + +00:41:30.960 --> 00:41:34.150 align:start position:0% +uh data given x uh +okay<00:41:31.280> so<00:41:31.920> there<00:41:32.079> are<00:41:32.400> basically<00:41:32.960> uh<00:41:33.760> different + +00:41:34.150 --> 00:41:34.160 align:start position:0% +okay so there are basically uh different + + +00:41:34.160 --> 00:41:35.910 align:start position:0% +okay so there are basically uh different +ways<00:41:34.400> of<00:41:34.560> looking<00:41:34.800> at<00:41:34.880> this<00:41:35.200> problem + +00:41:35.910 --> 00:41:35.920 align:start position:0% +ways of looking at this problem + + +00:41:35.920 --> 00:41:38.950 align:start position:0% +ways of looking at this problem +so<00:41:36.240> here<00:41:36.560> we<00:41:36.720> are<00:41:37.040> we<00:41:37.200> wanna<00:41:38.000> say<00:41:38.480> if<00:41:38.640> you + +00:41:38.950 --> 00:41:38.960 align:start position:0% +so here we are we wanna say if you + + +00:41:38.960 --> 00:41:41.990 align:start position:0% +so here we are we wanna say if you +have<00:41:39.280> a<00:41:39.520> model<00:41:40.720> then<00:41:41.359> what<00:41:41.599> is<00:41:41.760> the + +00:41:41.990 --> 00:41:42.000 align:start position:0% +have a model then what is the + + +00:41:42.000 --> 00:41:44.550 align:start position:0% +have a model then what is the +probability<00:41:42.800> of<00:41:43.119> observing<00:41:43.760> a<00:41:43.839> particular + +00:41:44.550 --> 00:41:44.560 align:start position:0% +probability of observing a particular + + +00:41:44.560 --> 00:41:45.750 align:start position:0% +probability of observing a particular +data + +00:41:45.750 --> 00:41:45.760 align:start position:0% +data + + +00:41:45.760 --> 00:41:48.870 align:start position:0% +data +from<00:41:46.000> the<00:41:46.160> model<00:41:46.640> so<00:41:46.960> my<00:41:47.280> model<00:41:48.079> has<00:41:48.480> both + +00:41:48.870 --> 00:41:48.880 align:start position:0% +from the model so my model has both + + +00:41:48.880 --> 00:41:51.910 align:start position:0% +from the model so my model has both +labels<00:41:49.599> and<00:41:50.079> features<00:41:51.119> but + +00:41:51.910 --> 00:41:51.920 align:start position:0% +labels and features but + + +00:41:51.920 --> 00:41:55.670 align:start position:0% +labels and features but +since<00:41:52.319> the<00:41:53.119> probability<00:41:53.760> of<00:41:53.920> the<00:41:54.079> features<00:41:54.720> uh + +00:41:55.670 --> 00:41:55.680 align:start position:0% +since the probability of the features uh + + +00:41:55.680 --> 00:41:58.069 align:start position:0% +since the probability of the features uh +you<00:41:55.760> know<00:41:56.000> we<00:41:56.240> actually<00:41:56.640> don't<00:41:56.880> care<00:41:57.119> about<00:41:57.520> it + +00:41:58.069 --> 00:41:58.079 align:start position:0% +you know we actually don't care about it + + +00:41:58.079 --> 00:41:58.710 align:start position:0% +you know we actually don't care about it +so<00:41:58.319> you + +00:41:58.710 --> 00:41:58.720 align:start position:0% +so you + + +00:41:58.720 --> 00:42:01.589 align:start position:0% +so you +can<00:41:59.200> see<00:41:59.599> that<00:42:00.319> becomes<00:42:00.800> constant<00:42:01.359> in<00:42:01.440> my + +00:42:01.589 --> 00:42:01.599 align:start position:0% +can see that becomes constant in my + + +00:42:01.599 --> 00:42:03.190 align:start position:0% +can see that becomes constant in my +optimization<00:42:02.400> problem + +00:42:03.190 --> 00:42:03.200 align:start position:0% +optimization problem + + +00:42:03.200 --> 00:42:06.230 align:start position:0% +optimization problem +and<00:42:03.520> that's<00:42:03.839> why<00:42:04.160> we<00:42:04.480> uh<00:42:05.280> don't<00:42:05.520> need<00:42:05.760> to + +00:42:06.230 --> 00:42:06.240 align:start position:0% +and that's why we uh don't need to + + +00:42:06.240 --> 00:42:08.309 align:start position:0% +and that's why we uh don't need to +write<00:42:06.560> it<00:42:06.720> up<00:42:07.200> so<00:42:07.359> this<00:42:07.599> is<00:42:07.760> basically + +00:42:08.309 --> 00:42:08.319 align:start position:0% +write it up so this is basically + + +00:42:08.319 --> 00:42:09.349 align:start position:0% +write it up so this is basically +probability<00:42:08.880> of<00:42:09.040> y + +00:42:09.349 --> 00:42:09.359 align:start position:0% +probability of y + + +00:42:09.359 --> 00:42:12.550 align:start position:0% +probability of y +i<00:42:09.839> comma<00:42:10.240> x<00:42:10.480> i<00:42:11.200> and<00:42:11.599> problem<00:42:12.000> and<00:42:12.160> using<00:42:12.400> the + +00:42:12.550 --> 00:42:12.560 align:start position:0% +i comma x i and problem and using the + + +00:42:12.560 --> 00:42:14.150 align:start position:0% +i comma x i and problem and using the +bayes<00:42:12.880> rule<00:42:13.119> you<00:42:13.280> have<00:42:13.520> like<00:42:13.680> probability<00:42:14.000> of + +00:42:14.150 --> 00:42:14.160 align:start position:0% +bayes rule you have like probability of + + +00:42:14.160 --> 00:42:16.390 align:start position:0% +bayes rule you have like probability of +y<00:42:14.400> i<00:42:14.560> given<00:42:14.880> x<00:42:15.040> i<00:42:15.280> times<00:42:15.599> probability<00:42:16.000> of<00:42:16.160> x + +00:42:16.390 --> 00:42:16.400 align:start position:0% +y i given x i times probability of x + + +00:42:16.400 --> 00:42:18.550 align:start position:0% +y i given x i times probability of x +i<00:42:16.640> but<00:42:16.800> probability<00:42:17.200> of<00:42:17.359> x<00:42:17.599> i<00:42:18.000> doesn't<00:42:18.319> depend + +00:42:18.550 --> 00:42:18.560 align:start position:0% +i but probability of x i doesn't depend + + +00:42:18.560 --> 00:42:20.309 align:start position:0% +i but probability of x i doesn't depend +on<00:42:18.720> my<00:42:18.880> model<00:42:19.200> parameters<00:42:19.680> so<00:42:19.839> it<00:42:20.000> becomes + +00:42:20.309 --> 00:42:20.319 align:start position:0% +on my model parameters so it becomes + + +00:42:20.319 --> 00:42:23.910 align:start position:0% +on my model parameters so it becomes +constant + +00:42:23.910 --> 00:42:23.920 align:start position:0% + + + +00:42:23.920 --> 00:42:26.390 align:start position:0% + +okay<00:42:24.240> any<00:42:24.560> other<00:42:24.880> questions<00:42:25.280> there<00:42:25.440> is<00:42:25.599> a<00:42:26.079> hand + +00:42:26.390 --> 00:42:26.400 align:start position:0% +okay any other questions there is a hand + + +00:42:26.400 --> 00:42:27.109 align:start position:0% +okay any other questions there is a hand +raised + +00:42:27.109 --> 00:42:27.119 align:start position:0% +raised + + +00:42:27.119 --> 00:42:30.150 align:start position:0% +raised +alexander<00:42:28.560> please<00:42:28.880> go<00:42:28.960> ahead + +00:42:30.150 --> 00:42:30.160 align:start position:0% +alexander please go ahead + + +00:42:30.160 --> 00:42:32.550 align:start position:0% +alexander please go ahead +hi<00:42:30.480> um<00:42:30.880> if<00:42:31.040> you<00:42:31.280> have<00:42:31.599> a<00:42:31.680> situation<00:42:32.240> where<00:42:32.480> the + +00:42:32.550 --> 00:42:32.560 align:start position:0% +hi um if you have a situation where the + + +00:42:32.560 --> 00:42:34.230 align:start position:0% +hi um if you have a situation where the +labels<00:42:32.960> themselves<00:42:33.440> are<00:42:33.520> probabilities + +00:42:34.230 --> 00:42:34.240 align:start position:0% +labels themselves are probabilities + + +00:42:34.240 --> 00:42:35.990 align:start position:0% +labels themselves are probabilities +instead<00:42:34.480> of<00:42:34.640> just<00:42:34.800> a<00:42:34.880> zero<00:42:35.200> or<00:42:35.359> a<00:42:35.440> one + +00:42:35.990 --> 00:42:36.000 align:start position:0% +instead of just a zero or a one + + +00:42:36.000 --> 00:42:37.990 align:start position:0% +instead of just a zero or a one +is<00:42:36.160> there<00:42:36.319> a<00:42:36.400> way<00:42:36.640> to<00:42:37.119> have<00:42:37.359> a<00:42:37.440> loss<00:42:37.680> function + +00:42:37.990 --> 00:42:38.000 align:start position:0% +is there a way to have a loss function + + +00:42:38.000 --> 00:42:41.190 align:start position:0% +is there a way to have a loss function +that<00:42:38.079> takes<00:42:38.240> that<00:42:38.400> into<00:42:38.640> account<00:42:38.960> better + +00:42:41.190 --> 00:42:41.200 align:start position:0% +that takes that into account better + + +00:42:41.200 --> 00:42:47.510 align:start position:0% +that takes that into account better +the<00:42:41.440> labels<00:42:41.920> are<00:42:42.319> probabilities + +00:42:47.510 --> 00:42:47.520 align:start position:0% + + + +00:42:47.520 --> 00:42:49.990 align:start position:0% + +yeah<00:42:48.000> probably<00:42:48.720> so<00:42:48.880> you<00:42:49.040> can<00:42:49.359> even<00:42:49.680> in<00:42:49.839> that + +00:42:49.990 --> 00:42:50.000 align:start position:0% +yeah probably so you can even in that + + +00:42:50.000 --> 00:42:52.470 align:start position:0% +yeah probably so you can even in that +case<00:42:50.240> you<00:42:50.400> can<00:42:50.560> use<00:42:50.720> a<00:42:50.800> quadratic<00:42:51.599> loss + +00:42:52.470 --> 00:42:52.480 align:start position:0% +case you can use a quadratic loss + + +00:42:52.480 --> 00:42:55.030 align:start position:0% +case you can use a quadratic loss +uh<00:42:52.960> if<00:42:53.119> you<00:42:53.200> want<00:42:53.440> to<00:42:53.599> pre<00:42:53.920> if<00:42:54.000> your<00:42:54.240> label<00:42:54.560> is<00:42:54.880> a + +00:42:55.030 --> 00:42:55.040 align:start position:0% +uh if you want to pre if your label is a + + +00:42:55.040 --> 00:42:56.950 align:start position:0% +uh if you want to pre if your label is a +number<00:42:55.359> between<00:42:55.760> 0<00:42:56.079> and<00:42:56.240> 1 + +00:42:56.950 --> 00:42:56.960 align:start position:0% +number between 0 and 1 + + +00:42:56.960 --> 00:42:59.109 align:start position:0% +number between 0 and 1 +but<00:42:57.200> i'm<00:42:57.280> sure<00:42:57.520> you<00:42:57.680> can<00:42:57.839> even<00:42:58.160> cook<00:42:58.480> up<00:42:58.640> with + +00:42:59.109 --> 00:42:59.119 align:start position:0% +but i'm sure you can even cook up with + + +00:42:59.119 --> 00:43:01.109 align:start position:0% +but i'm sure you can even cook up with +you<00:42:59.200> know<00:42:59.440> some<00:42:59.599> better<00:42:59.920> loss<00:43:00.160> function<00:43:00.480> so + +00:43:01.109 --> 00:43:01.119 align:start position:0% +you know some better loss function so + + +00:43:01.119 --> 00:43:02.790 align:start position:0% +you know some better loss function so +uh<00:43:01.440> here<00:43:01.680> actually<00:43:02.000> i<00:43:02.079> should<00:43:02.240> mention<00:43:02.560> these + +00:43:02.790 --> 00:43:02.800 align:start position:0% +uh here actually i should mention these + + +00:43:02.800 --> 00:43:04.390 align:start position:0% +uh here actually i should mention these +loss<00:43:03.040> functions<00:43:03.680> uh + +00:43:04.390 --> 00:43:04.400 align:start position:0% +loss functions uh + + +00:43:04.400 --> 00:43:07.750 align:start position:0% +loss functions uh +probably<00:43:04.720> the<00:43:04.880> most<00:43:05.119> popular<00:43:05.680> ones<00:43:06.400> uh<00:43:06.960> but + +00:43:07.750 --> 00:43:07.760 align:start position:0% +probably the most popular ones uh but + + +00:43:07.760 --> 00:43:10.069 align:start position:0% +probably the most popular ones uh but +we<00:43:07.920> have<00:43:08.480> many<00:43:08.720> many<00:43:09.040> other<00:43:09.359> loss<00:43:09.680> functions + +00:43:10.069 --> 00:43:10.079 align:start position:0% +we have many many other loss functions + + +00:43:10.079 --> 00:43:11.109 align:start position:0% +we have many many other loss functions +so<00:43:10.560> don't + +00:43:11.109 --> 00:43:11.119 align:start position:0% +so don't + + +00:43:11.119 --> 00:43:13.510 align:start position:0% +so don't +you<00:43:11.280> know<00:43:11.440> feel<00:43:11.680> that<00:43:12.400> you're<00:43:12.720> limited<00:43:13.119> to + +00:43:13.510 --> 00:43:13.520 align:start position:0% +you know feel that you're limited to + + +00:43:13.520 --> 00:43:17.510 align:start position:0% +you know feel that you're limited to +only<00:43:13.839> using<00:43:14.160> these<00:43:14.400> loss<00:43:14.640> functions + +00:43:17.510 --> 00:43:17.520 align:start position:0% + + + +00:43:17.520 --> 00:43:22.550 align:start position:0% + +okay<00:43:18.319> very<00:43:18.560> good<00:43:18.960> um<00:43:19.680> other<00:43:20.079> questions + +00:43:22.550 --> 00:43:22.560 align:start position:0% + + + +00:43:22.560 --> 00:43:27.430 align:start position:0% + +all<00:43:22.720> right<00:43:22.880> so<00:43:23.280> let's<00:43:23.680> uh<00:43:25.119> move<00:43:25.440> on<00:43:25.839> to + +00:43:27.430 --> 00:43:27.440 align:start position:0% +all right so let's uh move on to + + +00:43:27.440 --> 00:43:35.510 align:start position:0% +all right so let's uh move on to +the<00:43:27.760> next<00:43:28.079> part + +00:43:35.510 --> 00:43:35.520 align:start position:0% + + + +00:43:35.520 --> 00:43:38.950 align:start position:0% + +okay + +00:43:38.950 --> 00:43:38.960 align:start position:0% + + + +00:43:38.960 --> 00:43:41.510 align:start position:0% + +recall<00:43:39.680> what<00:43:39.920> was<00:43:40.079> the<00:43:40.319> erm<00:43:40.800> optimization + +00:43:41.510 --> 00:43:41.520 align:start position:0% +recall what was the erm optimization + + +00:43:41.520 --> 00:43:43.430 align:start position:0% +recall what was the erm optimization +problem + +00:43:43.430 --> 00:43:43.440 align:start position:0% +problem + + +00:43:43.440 --> 00:43:46.950 align:start position:0% +problem +we<00:43:43.599> are<00:43:43.760> minimizing<00:43:45.520> over<00:43:45.839> w + +00:43:46.950 --> 00:43:46.960 align:start position:0% +we are minimizing over w + + +00:43:46.960 --> 00:43:51.829 align:start position:0% +we are minimizing over w +the<00:43:47.200> average<00:43:47.760> loss<00:43:48.640> over<00:43:48.880> my<00:43:49.200> samples + +00:43:51.829 --> 00:43:51.839 align:start position:0% + + + +00:43:51.839 --> 00:43:55.030 align:start position:0% + +plus<00:43:52.240> f<00:43:52.400> w<00:43:52.800> of<00:43:52.960> x<00:43:53.200> i<00:43:53.680> and<00:43:53.920> y<00:43:54.160> i + +00:43:55.030 --> 00:43:55.040 align:start position:0% +plus f w of x i and y i + + +00:43:55.040 --> 00:43:58.150 align:start position:0% +plus f w of x i and y i +okay<00:43:55.599> so<00:43:55.839> we<00:43:56.079> already<00:43:56.560> know + +00:43:58.150 --> 00:43:58.160 align:start position:0% +okay so we already know + + +00:43:58.160 --> 00:44:00.550 align:start position:0% +okay so we already know +uh<00:43:58.480> about<00:43:58.800> this<00:43:59.040> loss<00:43:59.359> function<00:44:00.079> so<00:44:00.240> we<00:44:00.400> have + +00:44:00.550 --> 00:44:00.560 align:start position:0% +uh about this loss function so we have + + +00:44:00.560 --> 00:44:02.230 align:start position:0% +uh about this loss function so we have +hinge<00:44:00.880> laws<00:44:01.200> cross<00:44:01.440> entropy<00:44:01.920> laws<00:44:02.160> and + +00:44:02.230 --> 00:44:02.240 align:start position:0% +hinge laws cross entropy laws and + + +00:44:02.240 --> 00:44:03.270 align:start position:0% +hinge laws cross entropy laws and +quadratic<00:44:02.800> laws + +00:44:03.270 --> 00:44:03.280 align:start position:0% +quadratic laws + + +00:44:03.280 --> 00:44:07.190 align:start position:0% +quadratic laws +now<00:44:03.440> we<00:44:03.599> want<00:44:03.760> to<00:44:03.839> see<00:44:04.000> what<00:44:04.160> is<00:44:04.400> f<00:44:04.720> w<00:44:05.200> of<00:44:05.359> x<00:44:05.920> if<00:44:06.079> w + +00:44:07.190 --> 00:44:07.200 align:start position:0% +now we want to see what is f w of x if w + + +00:44:07.200 --> 00:44:11.190 align:start position:0% +now we want to see what is f w of x if w +is<00:44:07.440> a<00:44:07.680> function<00:44:08.640> mapping<00:44:09.119> from + +00:44:11.190 --> 00:44:11.200 align:start position:0% +is a function mapping from + + +00:44:11.200 --> 00:44:14.550 align:start position:0% +is a function mapping from +a<00:44:11.359> d-dimensional<00:44:12.079> space + +00:44:14.550 --> 00:44:14.560 align:start position:0% +a d-dimensional space + + +00:44:14.560 --> 00:44:17.990 align:start position:0% +a d-dimensional space +to<00:44:15.760> in<00:44:15.920> this<00:44:16.319> in<00:44:16.560> our<00:44:16.720> case<00:44:17.040> a<00:44:17.200> one-dimensional + +00:44:17.990 --> 00:44:18.000 align:start position:0% +to in this in our case a one-dimensional + + +00:44:18.000 --> 00:44:19.430 align:start position:0% +to in this in our case a one-dimensional +space + +00:44:19.430 --> 00:44:19.440 align:start position:0% +space + + +00:44:19.440 --> 00:44:23.589 align:start position:0% +space +from<00:44:20.560> rd<00:44:20.560> rd2r<00:44:22.319> so<00:44:22.480> what<00:44:22.720> is<00:44:22.800> the<00:44:23.040> simplest + +00:44:23.589 --> 00:44:23.599 align:start position:0% +from rd rd2r so what is the simplest + + +00:44:23.599 --> 00:44:26.309 align:start position:0% +from rd rd2r so what is the simplest +function<00:44:24.000> that<00:44:24.160> you<00:44:24.319> can + +00:44:26.309 --> 00:44:26.319 align:start position:0% +function that you can + + +00:44:26.319 --> 00:44:28.390 align:start position:0% +function that you can +use<00:44:26.560> for<00:44:26.800> such<00:44:27.040> a<00:44:27.200> mapping<00:44:27.680> the<00:44:27.839> simplest + +00:44:28.390 --> 00:44:28.400 align:start position:0% +use for such a mapping the simplest + + +00:44:28.400 --> 00:44:30.390 align:start position:0% +use for such a mapping the simplest +function<00:44:28.880> would<00:44:29.119> be<00:44:29.359> a<00:44:29.520> linear + +00:44:30.390 --> 00:44:30.400 align:start position:0% +function would be a linear + + +00:44:30.400 --> 00:44:34.950 align:start position:0% +function would be a linear +function<00:44:32.720> so<00:44:32.960> in<00:44:33.040> that<00:44:33.359> case + +00:44:34.950 --> 00:44:34.960 align:start position:0% +function so in that case + + +00:44:34.960 --> 00:44:37.910 align:start position:0% +function so in that case +you<00:44:35.119> have<00:44:35.359> a<00:44:35.520> linear<00:44:36.560> or<00:44:36.800> a<00:44:36.960> fine<00:44:37.359> if<00:44:37.520> you<00:44:37.680> add + +00:44:37.910 --> 00:44:37.920 align:start position:0% +you have a linear or a fine if you add + + +00:44:37.920 --> 00:44:43.750 align:start position:0% +you have a linear or a fine if you add +the<00:44:38.160> bias<00:44:38.560> term + +00:44:43.750 --> 00:44:43.760 align:start position:0% + + + +00:44:43.760 --> 00:44:45.990 align:start position:0% + +so<00:44:44.160> what<00:44:44.400> is<00:44:44.480> the<00:44:44.640> form<00:44:44.960> of<00:44:45.040> that<00:44:45.280> function<00:44:45.839> so + +00:44:45.990 --> 00:44:46.000 align:start position:0% +so what is the form of that function so + + +00:44:46.000 --> 00:44:46.790 align:start position:0% +so what is the form of that function so +you<00:44:46.160> have<00:44:46.480> f + +00:44:46.790 --> 00:44:46.800 align:start position:0% +you have f + + +00:44:46.800 --> 00:44:49.990 align:start position:0% +you have f +w<00:44:47.760> of<00:44:48.079> x<00:44:49.520> w + +00:44:49.990 --> 00:44:50.000 align:start position:0% +w of x w + + +00:44:50.000 --> 00:44:53.270 align:start position:0% +w of x w +transpose<00:44:50.720> x<00:44:51.040> plus<00:44:51.359> b<00:44:52.160> okay<00:44:52.400> i'm<00:44:52.560> gonna<00:44:52.960> i've + +00:44:53.270 --> 00:44:53.280 align:start position:0% +transpose x plus b okay i'm gonna i've + + +00:44:53.280 --> 00:44:54.870 align:start position:0% +transpose x plus b okay i'm gonna i've +used<00:44:53.440> a<00:44:53.520> little<00:44:53.760> bit<00:44:53.920> of<00:44:54.000> notation + +00:44:54.870 --> 00:44:54.880 align:start position:0% +used a little bit of notation + + +00:44:54.880 --> 00:44:58.230 align:start position:0% +used a little bit of notation +this<00:44:55.200> w<00:44:55.760> here<00:44:56.720> is<00:44:56.960> basically<00:44:57.440> the<00:44:57.599> set<00:44:57.839> of<00:44:58.079> all + +00:44:58.230 --> 00:44:58.240 align:start position:0% +this w here is basically the set of all + + +00:44:58.240 --> 00:45:00.069 align:start position:0% +this w here is basically the set of all +of<00:44:58.319> my<00:44:58.560> model<00:44:58.880> parameters<00:44:59.520> so<00:44:59.680> it's + +00:45:00.069 --> 00:45:00.079 align:start position:0% +of my model parameters so it's + + +00:45:00.079 --> 00:45:02.630 align:start position:0% +of my model parameters so it's +you<00:45:00.240> can<00:45:00.400> think<00:45:00.560> about<00:45:00.880> the<00:45:01.119> pole<00:45:01.359> w<00:45:02.319> and<00:45:02.480> this + +00:45:02.630 --> 00:45:02.640 align:start position:0% +you can think about the pole w and this + + +00:45:02.640 --> 00:45:03.109 align:start position:0% +you can think about the pole w and this +w + +00:45:03.109 --> 00:45:03.119 align:start position:0% +w + + +00:45:03.119 --> 00:45:06.630 align:start position:0% +w +here<00:45:04.000> uh<00:45:04.400> is<00:45:04.640> the<00:45:04.800> weights<00:45:05.280> uh<00:45:06.000> part<00:45:06.319> of<00:45:06.400> my + +00:45:06.630 --> 00:45:06.640 align:start position:0% +here uh is the weights uh part of my + + +00:45:06.640 --> 00:45:08.470 align:start position:0% +here uh is the weights uh part of my +weights<00:45:07.040> and<00:45:07.200> b<00:45:07.359> is<00:45:07.520> the<00:45:07.599> bias<00:45:07.920> sample<00:45:08.079> so<00:45:08.319> this + +00:45:08.470 --> 00:45:08.480 align:start position:0% +weights and b is the bias sample so this + + +00:45:08.480 --> 00:45:08.870 align:start position:0% +weights and b is the bias sample so this +w + +00:45:08.870 --> 00:45:08.880 align:start position:0% +w + + +00:45:08.880 --> 00:45:11.270 align:start position:0% +w +has<00:45:09.040> both<00:45:09.280> of<00:45:09.440> these<00:45:09.760> uh<00:45:10.079> w<00:45:10.560> and<00:45:10.640> b's<00:45:11.119> it's + +00:45:11.270 --> 00:45:11.280 align:start position:0% +has both of these uh w and b's it's + + +00:45:11.280 --> 00:45:15.109 align:start position:0% +has both of these uh w and b's it's +abuse<00:45:11.760> of<00:45:11.920> notation + +00:45:15.109 --> 00:45:15.119 align:start position:0% + + + +00:45:15.119 --> 00:45:18.470 align:start position:0% + +all<00:45:15.280> right<00:45:15.520> so<00:45:17.520> this<00:45:17.760> is<00:45:17.920> good + +00:45:18.470 --> 00:45:18.480 align:start position:0% +all right so this is good + + +00:45:18.480 --> 00:45:21.750 align:start position:0% +all right so this is good +uh<00:45:20.079> what<00:45:20.400> why<00:45:20.720> we<00:45:21.040> like + +00:45:21.750 --> 00:45:21.760 align:start position:0% +uh what why we like + + +00:45:21.760 --> 00:45:25.589 align:start position:0% +uh what why we like +linear<00:45:22.240> models<00:45:23.200> we<00:45:23.440> like<00:45:23.760> it<00:45:23.920> because + +00:45:25.589 --> 00:45:25.599 align:start position:0% +linear models we like it because + + +00:45:25.599 --> 00:45:28.390 align:start position:0% +linear models we like it because +my<00:45:25.839> loss<00:45:26.079> is<00:45:26.240> convex<00:45:27.040> either<00:45:27.440> cross<00:45:27.760> entropy + +00:45:28.390 --> 00:45:28.400 align:start position:0% +my loss is convex either cross entropy + + +00:45:28.400 --> 00:45:29.510 align:start position:0% +my loss is convex either cross entropy +or<00:45:28.640> hinge<00:45:29.119> or + +00:45:29.510 --> 00:45:29.520 align:start position:0% +or hinge or + + +00:45:29.520 --> 00:45:32.870 align:start position:0% +or hinge or +quadratic<00:45:30.720> when<00:45:30.960> my<00:45:31.200> fw<00:45:31.760> is<00:45:31.920> linear<00:45:32.480> the<00:45:32.640> whole + +00:45:32.870 --> 00:45:32.880 align:start position:0% +quadratic when my fw is linear the whole + + +00:45:32.880 --> 00:45:33.670 align:start position:0% +quadratic when my fw is linear the whole +thing<00:45:33.119> becomes + +00:45:33.670 --> 00:45:33.680 align:start position:0% +thing becomes + + +00:45:33.680 --> 00:45:36.550 align:start position:0% +thing becomes +convex<00:45:34.160> so<00:45:34.319> i<00:45:34.560> have<00:45:34.800> a<00:45:34.960> convex<00:45:35.359> optimization + +00:45:36.550 --> 00:45:36.560 align:start position:0% +convex so i have a convex optimization + + +00:45:36.560 --> 00:45:37.510 align:start position:0% +convex so i have a convex optimization +uh + +00:45:37.510 --> 00:45:37.520 align:start position:0% +uh + + +00:45:37.520 --> 00:45:41.670 align:start position:0% +uh +problem<00:45:38.960> and<00:45:39.839> we<00:45:40.079> know<00:45:40.319> how<00:45:40.480> to + +00:45:41.670 --> 00:45:41.680 align:start position:0% +problem and we know how to + + +00:45:41.680 --> 00:45:44.069 align:start position:0% +problem and we know how to +characterize<00:45:42.560> global<00:45:43.040> optimizers<00:45:43.680> for<00:45:43.839> that + +00:45:44.069 --> 00:45:44.079 align:start position:0% +characterize global optimizers for that + + +00:45:44.079 --> 00:45:45.589 align:start position:0% +characterize global optimizers for that +optimization<00:45:44.720> problem<00:45:45.040> so + +00:45:45.589 --> 00:45:45.599 align:start position:0% +optimization problem so + + +00:45:45.599 --> 00:45:49.430 align:start position:0% +optimization problem so +the<00:45:45.839> plus<00:45:46.400> for<00:45:46.640> this<00:45:47.040> is<00:45:47.839> uh<00:45:48.240> i<00:45:48.560> basically + +00:45:49.430 --> 00:45:49.440 align:start position:0% +the plus for this is uh i basically + + +00:45:49.440 --> 00:45:52.150 align:start position:0% +the plus for this is uh i basically +get<00:45:49.680> a<00:45:49.839> convex<00:45:50.240> optimization<00:45:50.880> problem<00:45:51.440> so<00:45:51.920> i'm + +00:45:52.150 --> 00:45:52.160 align:start position:0% +get a convex optimization problem so i'm + + +00:45:52.160 --> 00:45:53.030 align:start position:0% +get a convex optimization problem so i'm +happy + +00:45:53.030 --> 00:45:53.040 align:start position:0% +happy + + +00:45:53.040 --> 00:45:56.550 align:start position:0% +happy +about<00:45:53.359> it<00:45:54.560> but<00:45:54.800> what<00:45:54.960> about<00:45:55.200> the<00:45:55.440> negative + +00:45:56.550 --> 00:45:56.560 align:start position:0% +about it but what about the negative + + +00:45:56.560 --> 00:45:59.990 align:start position:0% +about it but what about the negative +sides<00:45:56.960> for<00:45:57.520> using<00:45:57.920> a<00:45:58.000> linear<00:45:58.400> refine<00:45:59.040> function + +00:45:59.990 --> 00:46:00.000 align:start position:0% +sides for using a linear refine function + + +00:46:00.000 --> 00:46:02.870 align:start position:0% +sides for using a linear refine function +it<00:46:00.240> is<00:46:00.480> pretty<00:46:00.960> limited<00:46:01.520> in<00:46:01.680> terms<00:46:02.079> of<00:46:02.480> the + +00:46:02.870 --> 00:46:02.880 align:start position:0% +it is pretty limited in terms of the + + +00:46:02.880 --> 00:46:03.910 align:start position:0% +it is pretty limited in terms of the +representation + +00:46:03.910 --> 00:46:03.920 align:start position:0% +representation + + +00:46:03.920 --> 00:46:06.230 align:start position:0% +representation +power<00:46:04.400> so<00:46:04.720> especially<00:46:05.200> in<00:46:05.359> practice<00:46:05.839> you<00:46:06.000> may + +00:46:06.230 --> 00:46:06.240 align:start position:0% +power so especially in practice you may + + +00:46:06.240 --> 00:46:08.630 align:start position:0% +power so especially in practice you may +have<00:46:06.880> non-linear<00:46:07.680> pretty<00:46:08.079> nonlinear + +00:46:08.630 --> 00:46:08.640 align:start position:0% +have non-linear pretty nonlinear + + +00:46:08.640 --> 00:46:10.470 align:start position:0% +have non-linear pretty nonlinear +relationships<00:46:09.359> between<00:46:09.760> x<00:46:10.000> and + +00:46:10.470 --> 00:46:10.480 align:start position:0% +relationships between x and + + +00:46:10.480 --> 00:46:13.430 align:start position:0% +relationships between x and +y<00:46:10.880> so<00:46:11.680> restricting<00:46:12.319> ourselves<00:46:12.800> to<00:46:12.880> a<00:46:13.040> linear + +00:46:13.430 --> 00:46:13.440 align:start position:0% +y so restricting ourselves to a linear + + +00:46:13.440 --> 00:46:14.710 align:start position:0% +y so restricting ourselves to a linear +model + +00:46:14.710 --> 00:46:14.720 align:start position:0% +model + + +00:46:14.720 --> 00:46:17.829 align:start position:0% +model +won't<00:46:15.040> probably<00:46:15.599> be<00:46:15.920> good<00:46:16.319> in<00:46:16.480> those<00:46:16.800> cases + +00:46:17.829 --> 00:46:17.839 align:start position:0% +won't probably be good in those cases + + +00:46:17.839 --> 00:46:22.550 align:start position:0% +won't probably be good in those cases +so<00:46:18.000> we<00:46:18.240> have<00:46:18.560> here<00:46:19.119> limited<00:46:20.599> representation + +00:46:22.550 --> 00:46:22.560 align:start position:0% +so we have here limited representation + + +00:46:22.560 --> 00:46:25.670 align:start position:0% +so we have here limited representation +power<00:46:23.359> for<00:46:23.599> these<00:46:23.920> functions<00:46:24.319> so + +00:46:25.670 --> 00:46:25.680 align:start position:0% +power for these functions so + + +00:46:25.680 --> 00:46:28.390 align:start position:0% +power for these functions so +we're<00:46:26.000> a<00:46:26.079> little<00:46:26.319> bit<00:46:26.560> sad<00:46:27.680> so<00:46:27.839> in<00:46:28.000> order<00:46:28.240> to + +00:46:28.390 --> 00:46:28.400 align:start position:0% +we're a little bit sad so in order to + + +00:46:28.400 --> 00:46:29.589 align:start position:0% +we're a little bit sad so in order to +deal<00:46:28.640> with<00:46:28.880> this + +00:46:29.589 --> 00:46:29.599 align:start position:0% +deal with this + + +00:46:29.599 --> 00:46:35.589 align:start position:0% +deal with this +we<00:46:29.920> look<00:46:30.160> into<00:46:30.960> non-linear<00:46:32.319> functions + +00:46:35.589 --> 00:46:35.599 align:start position:0% + + + +00:46:35.599 --> 00:46:41.109 align:start position:0% + +from<00:46:36.880> x<00:46:37.200> to<00:46:38.000> y + +00:46:41.109 --> 00:46:41.119 align:start position:0% + + + +00:46:41.119 --> 00:46:43.829 align:start position:0% + +okay<00:46:41.359> so<00:46:41.599> how<00:46:41.839> can<00:46:42.000> i<00:46:42.160> add<00:46:42.319> non-linearity<00:46:43.280> here + +00:46:43.829 --> 00:46:43.839 align:start position:0% +okay so how can i add non-linearity here + + +00:46:43.839 --> 00:46:44.470 align:start position:0% +okay so how can i add non-linearity here +so<00:46:44.079> let's + +00:46:44.470 --> 00:46:44.480 align:start position:0% +so let's + + +00:46:44.480 --> 00:46:48.630 align:start position:0% +so let's +uh<00:46:45.119> look<00:46:45.359> at<00:46:45.440> this<00:46:45.680> linear<00:46:46.079> function<00:46:46.480> again + +00:46:48.630 --> 00:46:48.640 align:start position:0% +uh look at this linear function again + + +00:46:48.640 --> 00:46:51.430 align:start position:0% +uh look at this linear function again +i<00:46:48.800> have<00:46:49.040> w<00:46:49.359> transpose<00:46:49.920> x<00:46:50.079> plus<00:46:50.480> b<00:46:50.880> that's<00:46:51.119> my + +00:46:51.430 --> 00:46:51.440 align:start position:0% +i have w transpose x plus b that's my + + +00:46:51.440 --> 00:46:52.710 align:start position:0% +i have w transpose x plus b that's my +output + +00:46:52.710 --> 00:46:52.720 align:start position:0% +output + + +00:46:52.720 --> 00:46:55.750 align:start position:0% +output +so<00:46:53.040> in<00:46:53.200> other<00:46:53.440> words<00:46:53.760> i<00:46:53.920> have<00:46:54.960> say + +00:46:55.750 --> 00:46:55.760 align:start position:0% +so in other words i have say + + +00:46:55.760 --> 00:46:58.150 align:start position:0% +so in other words i have say +x1<00:46:56.160> to<00:46:56.319> xt<00:46:56.720> these<00:46:56.960> are<00:46:57.280> different<00:46:57.680> components + +00:46:58.150 --> 00:46:58.160 align:start position:0% +x1 to xt these are different components + + +00:46:58.160 --> 00:46:59.109 align:start position:0% +x1 to xt these are different components +of<00:46:58.319> my<00:46:58.560> x + +00:46:59.109 --> 00:46:59.119 align:start position:0% +of my x + + +00:46:59.119 --> 00:47:02.470 align:start position:0% +of my x +i<00:46:59.359> multiply<00:47:00.079> each<00:47:00.319> of<00:47:00.400> them<00:47:00.640> with<00:47:00.880> a<00:47:00.960> weight + +00:47:02.470 --> 00:47:02.480 align:start position:0% +i multiply each of them with a weight + + +00:47:02.480 --> 00:47:06.309 align:start position:0% +i multiply each of them with a weight +w1<00:47:03.680> x2<00:47:04.160> is<00:47:04.319> multiplied<00:47:04.880> by<00:47:05.200> w2 + +00:47:06.309 --> 00:47:06.319 align:start position:0% +w1 x2 is multiplied by w2 + + +00:47:06.319 --> 00:47:09.589 align:start position:0% +w1 x2 is multiplied by w2 +xd<00:47:06.720> is<00:47:06.880> multiplied<00:47:07.440> by<00:47:07.680> wd<00:47:08.960> i + +00:47:09.589 --> 00:47:09.599 align:start position:0% +xd is multiplied by wd i + + +00:47:09.599 --> 00:47:12.309 align:start position:0% +xd is multiplied by wd i +sum<00:47:09.839> them<00:47:10.079> up<00:47:10.240> so<00:47:10.400> this<00:47:10.560> is<00:47:10.720> the<00:47:10.800> w<00:47:11.359> transpose + +00:47:12.309 --> 00:47:12.319 align:start position:0% +sum them up so this is the w transpose + + +00:47:12.319 --> 00:47:13.990 align:start position:0% +sum them up so this is the w transpose +times<00:47:12.640> x<00:47:12.960> part + +00:47:13.990 --> 00:47:14.000 align:start position:0% +times x part + + +00:47:14.000 --> 00:47:17.430 align:start position:0% +times x part +and<00:47:14.240> i<00:47:14.400> add<00:47:14.800> a<00:47:14.960> bias<00:47:15.359> term<00:47:15.680> to<00:47:15.920> it<00:47:16.720> and<00:47:16.960> this<00:47:17.280> is + +00:47:17.430 --> 00:47:17.440 align:start position:0% +and i add a bias term to it and this is + + +00:47:17.440 --> 00:47:18.309 align:start position:0% +and i add a bias term to it and this is +the + +00:47:18.309 --> 00:47:18.319 align:start position:0% +the + + +00:47:18.319 --> 00:47:21.670 align:start position:0% +the +output<00:47:19.119> of<00:47:19.440> my<00:47:20.000> linear<00:47:20.480> mapping + +00:47:21.670 --> 00:47:21.680 align:start position:0% +output of my linear mapping + + +00:47:21.680 --> 00:47:23.990 align:start position:0% +output of my linear mapping +so<00:47:21.920> this<00:47:22.160> part<00:47:22.400> would<00:47:22.559> be<00:47:22.720> w<00:47:23.119> transpose<00:47:23.599> x<00:47:23.760> plus + +00:47:23.990 --> 00:47:24.000 align:start position:0% +so this part would be w transpose x plus + + +00:47:24.000 --> 00:47:25.430 align:start position:0% +so this part would be w transpose x plus +b<00:47:24.319> this<00:47:24.480> is<00:47:24.559> the + +00:47:25.430 --> 00:47:25.440 align:start position:0% +b this is the + + +00:47:25.440 --> 00:47:27.829 align:start position:0% +b this is the +linear<00:47:26.160> or<00:47:26.400> a<00:47:26.559> fine<00:47:26.880> transformations<00:47:27.599> that<00:47:27.760> i + +00:47:27.829 --> 00:47:27.839 align:start position:0% +linear or a fine transformations that i + + +00:47:27.839 --> 00:47:28.549 align:start position:0% +linear or a fine transformations that i +have + +00:47:28.549 --> 00:47:28.559 align:start position:0% +have + + +00:47:28.559 --> 00:47:30.150 align:start position:0% +have +so<00:47:28.800> in<00:47:28.880> order<00:47:29.200> to<00:47:29.359> add<00:47:29.520> a<00:47:29.599> little<00:47:29.839> bit<00:47:29.920> of<00:47:30.079> a + +00:47:30.150 --> 00:47:30.160 align:start position:0% +so in order to add a little bit of a + + +00:47:30.160 --> 00:47:32.230 align:start position:0% +so in order to add a little bit of a +non-linearity<00:47:31.280> to + +00:47:32.230 --> 00:47:32.240 align:start position:0% +non-linearity to + + +00:47:32.240 --> 00:47:36.069 align:start position:0% +non-linearity to +my<00:47:33.040> mapping<00:47:34.000> so<00:47:34.240> what<00:47:34.960> one<00:47:35.200> thing<00:47:35.359> i<00:47:35.520> can<00:47:35.760> do + +00:47:36.069 --> 00:47:36.079 align:start position:0% +my mapping so what one thing i can do + + +00:47:36.079 --> 00:47:38.230 align:start position:0% +my mapping so what one thing i can do +is<00:47:36.480> very<00:47:36.720> simple<00:47:37.119> so<00:47:37.280> i<00:47:37.359> can<00:47:37.520> just<00:47:37.760> add<00:47:38.079> a + +00:47:38.230 --> 00:47:38.240 align:start position:0% +is very simple so i can just add a + + +00:47:38.240 --> 00:47:40.470 align:start position:0% +is very simple so i can just add a +non-linear<00:47:38.880> function<00:47:39.280> here + +00:47:40.470 --> 00:47:40.480 align:start position:0% +non-linear function here + + +00:47:40.480 --> 00:47:45.589 align:start position:0% +non-linear function here +let's<00:47:40.720> say<00:47:41.040> i<00:47:41.200> have<00:47:41.520> phi<00:47:42.400> and<00:47:42.640> my<00:47:42.960> output + +00:47:45.589 --> 00:47:45.599 align:start position:0% + + + +00:47:45.599 --> 00:47:49.829 align:start position:0% + +would<00:47:45.760> be + +00:47:49.829 --> 00:47:49.839 align:start position:0% + + + +00:47:49.839 --> 00:47:53.589 align:start position:0% + +phi<00:47:50.400> of<00:47:51.280> w<00:47:51.760> transpose<00:47:52.240> x<00:47:52.480> plus<00:47:52.720> b + +00:47:53.589 --> 00:47:53.599 align:start position:0% +phi of w transpose x plus b + + +00:47:53.599 --> 00:47:56.150 align:start position:0% +phi of w transpose x plus b +and<00:47:53.760> phi<00:47:54.000> is<00:47:54.079> a<00:47:54.160> non-linear<00:47:54.800> function<00:47:55.680> we<00:47:55.920> call + +00:47:56.150 --> 00:47:56.160 align:start position:0% +and phi is a non-linear function we call + + +00:47:56.160 --> 00:48:00.309 align:start position:0% +and phi is a non-linear function we call +it<00:47:56.319> an<00:47:56.480> activation<00:47:57.200> function + +00:48:00.309 --> 00:48:00.319 align:start position:0% + + + +00:48:00.319 --> 00:48:05.349 align:start position:0% + +activation<00:48:01.040> function + +00:48:05.349 --> 00:48:05.359 align:start position:0% + + + +00:48:05.359 --> 00:48:08.230 align:start position:0% + +all<00:48:05.520> right<00:48:05.880> congratulations<00:48:07.280> this<00:48:07.520> is + +00:48:08.230 --> 00:48:08.240 align:start position:0% +all right congratulations this is + + +00:48:08.240 --> 00:48:09.829 align:start position:0% +all right congratulations this is +perhaps<00:48:08.800> the<00:48:09.040> simplest + +00:48:09.829 --> 00:48:09.839 align:start position:0% +perhaps the simplest + + +00:48:09.839 --> 00:48:12.549 align:start position:0% +perhaps the simplest +neural<00:48:10.240> network<00:48:11.280> so<00:48:11.520> this<00:48:11.680> is<00:48:11.839> a<00:48:11.920> mapping<00:48:12.319> from + +00:48:12.549 --> 00:48:12.559 align:start position:0% +neural network so this is a mapping from + + +00:48:12.559 --> 00:48:13.750 align:start position:0% +neural network so this is a mapping from +the<00:48:12.720> dimensional<00:48:13.280> space<00:48:13.599> we<00:48:13.680> have + +00:48:13.750 --> 00:48:13.760 align:start position:0% +the dimensional space we have + + +00:48:13.760 --> 00:48:16.470 align:start position:0% +the dimensional space we have +one-dimensional<00:48:14.400> space<00:48:14.720> it<00:48:14.880> is<00:48:15.040> non-linear + +00:48:16.470 --> 00:48:16.480 align:start position:0% +one-dimensional space it is non-linear + + +00:48:16.480 --> 00:48:20.309 align:start position:0% +one-dimensional space it is non-linear +so<00:48:16.640> this<00:48:16.960> uh<00:48:18.559> architecture<00:48:19.280> is<00:48:19.520> a<00:48:19.760> simple + +00:48:20.309 --> 00:48:20.319 align:start position:0% +so this uh architecture is a simple + + +00:48:20.319 --> 00:48:23.670 align:start position:0% +so this uh architecture is a simple +neural<00:48:20.720> network<00:48:21.119> architecture + +00:48:23.670 --> 00:48:23.680 align:start position:0% +neural network architecture + + +00:48:23.680 --> 00:48:26.309 align:start position:0% +neural network architecture +but<00:48:23.920> like<00:48:24.079> this<00:48:24.319> non-linearity<00:48:25.119> is<00:48:25.359> uh<00:48:25.680> not<00:48:26.240> a + +00:48:26.309 --> 00:48:26.319 align:start position:0% +but like this non-linearity is uh not a + + +00:48:26.319 --> 00:48:28.309 align:start position:0% +but like this non-linearity is uh not a +lot<00:48:26.640> right<00:48:26.880> so<00:48:27.040> we<00:48:27.200> need<00:48:27.359> to<00:48:27.520> add<00:48:27.760> a<00:48:27.839> little<00:48:28.079> bit + +00:48:28.309 --> 00:48:28.319 align:start position:0% +lot right so we need to add a little bit + + +00:48:28.319 --> 00:48:30.069 align:start position:0% +lot right so we need to add a little bit +more<00:48:28.559> non-linearity<00:48:29.280> in<00:48:29.440> order<00:48:29.680> to<00:48:29.839> have<00:48:30.000> a + +00:48:30.069 --> 00:48:30.079 align:start position:0% +more non-linearity in order to have a + + +00:48:30.079 --> 00:48:30.870 align:start position:0% +more non-linearity in order to have a +better + +00:48:30.870 --> 00:48:30.880 align:start position:0% +better + + +00:48:30.880 --> 00:48:34.150 align:start position:0% +better +uh<00:48:31.440> stronger<00:48:31.839> representation<00:48:32.720> power<00:48:33.760> so + +00:48:34.150 --> 00:48:34.160 align:start position:0% +uh stronger representation power so + + +00:48:34.160 --> 00:48:37.670 align:start position:0% +uh stronger representation power so +how<00:48:34.559> we<00:48:34.800> can<00:48:35.040> do<00:48:35.440> that<00:48:37.040> first<00:48:37.280> of<00:48:37.359> all<00:48:37.520> i'm + +00:48:37.670 --> 00:48:37.680 align:start position:0% +how we can do that first of all i'm + + +00:48:37.680 --> 00:48:39.270 align:start position:0% +how we can do that first of all i'm +going<00:48:37.760> to<00:48:37.839> summarize<00:48:38.400> this<00:48:38.559> picture<00:48:38.960> because + +00:48:39.270 --> 00:48:39.280 align:start position:0% +going to summarize this picture because + + +00:48:39.280 --> 00:48:40.470 align:start position:0% +going to summarize this picture because +i<00:48:39.359> don't<00:48:39.520> want<00:48:39.680> to<00:48:39.839> like + +00:48:40.470 --> 00:48:40.480 align:start position:0% +i don't want to like + + +00:48:40.480 --> 00:48:42.790 align:start position:0% +i don't want to like +write<00:48:40.720> down<00:48:40.960> w's<00:48:41.599> and<00:48:41.760> b's<00:48:42.160> and<00:48:42.400> phi + +00:48:42.790 --> 00:48:42.800 align:start position:0% +write down w's and b's and phi + + +00:48:42.800 --> 00:48:43.910 align:start position:0% +write down w's and b's and phi +separately + +00:48:43.910 --> 00:48:43.920 align:start position:0% +separately + + +00:48:43.920 --> 00:48:46.150 align:start position:0% +separately +so<00:48:44.079> we<00:48:44.240> are<00:48:44.400> going<00:48:44.640> to<00:48:45.200> just<00:48:45.440> summarize<00:48:45.920> this + +00:48:46.150 --> 00:48:46.160 align:start position:0% +so we are going to just summarize this + + +00:48:46.160 --> 00:48:47.430 align:start position:0% +so we are going to just summarize this +picture + +00:48:47.430 --> 00:48:47.440 align:start position:0% +picture + + +00:48:47.440 --> 00:48:51.510 align:start position:0% +picture +by<00:48:49.200> just<00:48:49.520> a<00:48:49.760> simple + +00:48:51.510 --> 00:48:51.520 align:start position:0% +by just a simple + + +00:48:51.520 --> 00:48:53.750 align:start position:0% +by just a simple +picture<00:48:52.000> like<00:48:52.160> this<00:48:52.880> this<00:48:53.119> is<00:48:53.280> called<00:48:53.520> the + +00:48:53.750 --> 00:48:53.760 align:start position:0% +picture like this this is called the + + +00:48:53.760 --> 00:48:57.750 align:start position:0% +picture like this this is called the +neuron + +00:48:57.750 --> 00:48:57.760 align:start position:0% + + + +00:48:57.760 --> 00:49:00.390 align:start position:0% + +and<00:48:58.319> implicitly<00:48:59.200> we<00:48:59.359> are<00:48:59.440> saying<00:49:00.079> all<00:49:00.319> of + +00:49:00.390 --> 00:49:00.400 align:start position:0% +and implicitly we are saying all of + + +00:49:00.400 --> 00:49:01.190 align:start position:0% +and implicitly we are saying all of +these<00:49:00.720> x's + +00:49:01.190 --> 00:49:01.200 align:start position:0% +these x's + + +00:49:01.200 --> 00:49:03.910 align:start position:0% +these x's +are<00:49:01.359> multiplied<00:49:02.000> with<00:49:02.160> the<00:49:02.319> vector<00:49:02.800> w<00:49:03.599> there's + +00:49:03.910 --> 00:49:03.920 align:start position:0% +are multiplied with the vector w there's + + +00:49:03.920 --> 00:49:05.349 align:start position:0% +are multiplied with the vector w there's +a<00:49:04.000> bias<00:49:04.400> term<00:49:04.720> here + +00:49:05.349 --> 00:49:05.359 align:start position:0% +a bias term here + + +00:49:05.359 --> 00:49:07.910 align:start position:0% +a bias term here +and<00:49:05.440> there's<00:49:05.760> an<00:49:05.920> activation<00:49:06.559> function<00:49:07.359> uh<00:49:07.839> in + +00:49:07.910 --> 00:49:07.920 align:start position:0% +and there's an activation function uh in + + +00:49:07.920 --> 00:49:09.910 align:start position:0% +and there's an activation function uh in +the<00:49:08.079> neuron<00:49:08.480> so<00:49:08.720> my<00:49:08.880> output<00:49:09.200> would<00:49:09.359> be<00:49:09.599> phi + +00:49:09.910 --> 00:49:09.920 align:start position:0% +the neuron so my output would be phi + + +00:49:09.920 --> 00:49:11.990 align:start position:0% +the neuron so my output would be phi +of<00:49:10.000> w<00:49:10.319> transpose<00:49:10.800> x<00:49:10.960> plus<00:49:11.200> p<00:49:11.440> so<00:49:11.520> it<00:49:11.680> is + +00:49:11.990 --> 00:49:12.000 align:start position:0% +of w transpose x plus p so it is + + +00:49:12.000 --> 00:49:13.750 align:start position:0% +of w transpose x plus p so it is +implicit<00:49:12.720> in<00:49:12.800> this<00:49:12.960> picture<00:49:13.280> so<00:49:13.440> we<00:49:13.599> just + +00:49:13.750 --> 00:49:13.760 align:start position:0% +implicit in this picture so we just + + +00:49:13.760 --> 00:49:14.710 align:start position:0% +implicit in this picture so we just +don't<00:49:13.920> want<00:49:14.079> to<00:49:14.240> like + +00:49:14.710 --> 00:49:14.720 align:start position:0% +don't want to like + + +00:49:14.720 --> 00:49:18.309 align:start position:0% +don't want to like +you<00:49:14.880> know<00:49:15.359> carry<00:49:15.920> all<00:49:16.079> of<00:49:16.240> these<00:49:16.559> uh + +00:49:18.309 --> 00:49:18.319 align:start position:0% +you know carry all of these uh + + +00:49:18.319 --> 00:49:21.430 align:start position:0% +you know carry all of these uh +notations<00:49:19.280> in<00:49:19.920> in<00:49:20.160> different + +00:49:21.430 --> 00:49:21.440 align:start position:0% +notations in in different + + +00:49:21.440 --> 00:49:24.950 align:start position:0% +notations in in different +networks<00:49:22.000> that<00:49:22.160> will<00:49:22.400> happen<00:49:23.760> okay<00:49:24.079> so + +00:49:24.950 --> 00:49:24.960 align:start position:0% +networks that will happen okay so + + +00:49:24.960 --> 00:49:28.150 align:start position:0% +networks that will happen okay so +one<00:49:25.200> way<00:49:25.440> to<00:49:25.599> add<00:49:25.839> non-linearity<00:49:26.800> would<00:49:27.040> be<00:49:27.839> to + +00:49:28.150 --> 00:49:28.160 align:start position:0% +one way to add non-linearity would be to + + +00:49:28.160 --> 00:49:31.589 align:start position:0% +one way to add non-linearity would be to +add<00:49:28.640> multiple<00:49:29.200> neurons<00:49:30.000> in<00:49:30.559> one<00:49:30.800> layer<00:49:31.200> or<00:49:31.440> you + +00:49:31.589 --> 00:49:31.599 align:start position:0% +add multiple neurons in one layer or you + + +00:49:31.599 --> 00:49:32.470 align:start position:0% +add multiple neurons in one layer or you +can<00:49:31.839> add + +00:49:32.470 --> 00:49:32.480 align:start position:0% +can add + + +00:49:32.480 --> 00:49:35.910 align:start position:0% +can add +them<00:49:32.880> in<00:49:33.599> cascaded<00:49:34.319> layers<00:49:35.280> so<00:49:35.520> what<00:49:35.680> do<00:49:35.839> i + +00:49:35.910 --> 00:49:35.920 align:start position:0% +them in cascaded layers so what do i + + +00:49:35.920 --> 00:49:38.150 align:start position:0% +them in cascaded layers so what do i +mean<00:49:36.079> by<00:49:36.319> that<00:49:36.559> so<00:49:36.800> you<00:49:36.960> have<00:49:37.280> x12 + +00:49:38.150 --> 00:49:38.160 align:start position:0% +mean by that so you have x12 + + +00:49:38.160 --> 00:49:43.030 align:start position:0% +mean by that so you have x12 +xt<00:49:40.319> and<00:49:40.480> then<00:49:40.720> you<00:49:40.880> have<00:49:41.119> say<00:49:41.359> one<00:49:41.599> neuron + +00:49:43.030 --> 00:49:43.040 align:start position:0% +xt and then you have say one neuron + + +00:49:43.040 --> 00:49:46.390 align:start position:0% +xt and then you have say one neuron +here<00:49:43.839> connected<00:49:44.319> to<00:49:44.640> all<00:49:44.880> of<00:49:44.960> these<00:49:45.280> inputs + +00:49:46.390 --> 00:49:46.400 align:start position:0% +here connected to all of these inputs + + +00:49:46.400 --> 00:49:52.069 align:start position:0% +here connected to all of these inputs +you<00:49:46.559> can<00:49:46.800> add<00:49:47.280> more<00:49:47.520> than<00:49:47.760> one<00:49:48.000> neuron + +00:49:52.069 --> 00:49:52.079 align:start position:0% + + + +00:49:52.079 --> 00:49:54.150 align:start position:0% + +and<00:49:52.640> then<00:49:52.880> you<00:49:53.040> can<00:49:53.280> treat<00:49:53.599> the<00:49:53.680> output<00:49:54.079> of + +00:49:54.150 --> 00:49:54.160 align:start position:0% +and then you can treat the output of + + +00:49:54.160 --> 00:49:56.150 align:start position:0% +and then you can treat the output of +these<00:49:54.400> neurons<00:49:54.800> as<00:49:54.960> your<00:49:55.280> input<00:49:55.599> to<00:49:55.760> the<00:49:55.839> next + +00:49:56.150 --> 00:49:56.160 align:start position:0% +these neurons as your input to the next + + +00:49:56.160 --> 00:49:57.190 align:start position:0% +these neurons as your input to the next +layer<00:49:56.480> so<00:49:56.640> you<00:49:56.800> can + +00:49:57.190 --> 00:49:57.200 align:start position:0% +layer so you can + + +00:49:57.200 --> 00:50:02.309 align:start position:0% +layer so you can +add<00:49:58.640> other<00:49:59.119> layers + +00:50:02.309 --> 00:50:02.319 align:start position:0% + + + +00:50:02.319 --> 00:50:06.230 align:start position:0% + +to<00:50:02.559> your<00:50:02.800> network + +00:50:06.230 --> 00:50:06.240 align:start position:0% + + + +00:50:06.240 --> 00:50:12.470 align:start position:0% + +and<00:50:06.480> at<00:50:06.559> the<00:50:06.640> end<00:50:06.880> of<00:50:06.960> the<00:50:10.839> day + +00:50:12.470 --> 00:50:12.480 align:start position:0% +and at the end of the day + + +00:50:12.480 --> 00:50:16.549 align:start position:0% +and at the end of the day +let's<00:50:12.720> say<00:50:13.040> you<00:50:13.359> have<00:50:14.800> let's<00:50:14.960> say<00:50:15.200> 10<00:50:15.440> outputs + +00:50:16.549 --> 00:50:16.559 align:start position:0% +let's say you have let's say 10 outputs + + +00:50:16.559 --> 00:50:20.230 align:start position:0% +let's say you have let's say 10 outputs +if<00:50:16.640> you<00:50:16.800> have<00:50:17.040> a<00:50:17.280> 10<00:50:19.119> class<00:50:19.520> classification + +00:50:20.230 --> 00:50:20.240 align:start position:0% +if you have a 10 class classification + + +00:50:20.240 --> 00:50:23.990 align:start position:0% +if you have a 10 class classification +probe + +00:50:23.990 --> 00:50:24.000 align:start position:0% + + + +00:50:24.000 --> 00:50:27.510 align:start position:0% + +okay<00:50:24.319> as<00:50:24.400> i<00:50:24.559> mentioned<00:50:25.680> so<00:50:26.720> these<00:50:26.960> outputs<00:50:27.440> you + +00:50:27.510 --> 00:50:27.520 align:start position:0% +okay as i mentioned so these outputs you + + +00:50:27.520 --> 00:50:29.510 align:start position:0% +okay as i mentioned so these outputs you +need<00:50:27.680> to<00:50:27.839> turn<00:50:28.079> them<00:50:28.319> into<00:50:28.640> probability<00:50:29.359> so + +00:50:29.510 --> 00:50:29.520 align:start position:0% +need to turn them into probability so + + +00:50:29.520 --> 00:50:33.270 align:start position:0% +need to turn them into probability so +you<00:50:29.680> add<00:50:29.920> a<00:50:30.079> soft + +00:50:33.270 --> 00:50:33.280 align:start position:0% + + + +00:50:33.280 --> 00:50:36.470 align:start position:0% + +soft<00:50:33.680> max<00:50:34.000> layer<00:50:34.480> here<00:50:35.440> to<00:50:35.760> turn<00:50:36.079> them + +00:50:36.470 --> 00:50:36.480 align:start position:0% +soft max layer here to turn them + + +00:50:36.480 --> 00:50:40.950 align:start position:0% +soft max layer here to turn them +into<00:50:38.839> probabilities + +00:50:40.950 --> 00:50:40.960 align:start position:0% +into probabilities + + +00:50:40.960 --> 00:50:45.510 align:start position:0% +into probabilities +for<00:50:42.160> predictions<00:50:42.720> of<00:50:42.880> each<00:50:43.200> class + +00:50:45.510 --> 00:50:45.520 align:start position:0% +for predictions of each class + + +00:50:45.520 --> 00:50:47.910 align:start position:0% +for predictions of each class +all<00:50:45.760> right<00:50:46.000> so<00:50:46.800> this<00:50:47.040> is<00:50:47.119> called<00:50:47.359> the<00:50:47.520> input + +00:50:47.910 --> 00:50:47.920 align:start position:0% +all right so this is called the input + + +00:50:47.920 --> 00:50:48.549 align:start position:0% +all right so this is called the input +layer + +00:50:48.549 --> 00:50:48.559 align:start position:0% +layer + + +00:50:48.559 --> 00:50:52.309 align:start position:0% +layer +for<00:50:48.960> obvious<00:50:49.440> reasons + +00:50:52.309 --> 00:50:52.319 align:start position:0% + + + +00:50:52.319 --> 00:50:59.430 align:start position:0% + +these<00:50:52.559> are<00:50:52.640> hidden<00:50:53.040> layers<00:50:54.000> for<00:50:54.240> my<00:50:54.480> network + +00:50:59.430 --> 00:50:59.440 align:start position:0% + + + +00:50:59.440 --> 00:51:05.349 align:start position:0% + +and<00:50:59.599> this<00:50:59.839> is<00:50:59.920> the<00:51:00.240> output<00:51:00.640> layer + +00:51:05.349 --> 00:51:05.359 align:start position:0% + + + +00:51:05.359 --> 00:51:07.589 align:start position:0% + +the<00:51:05.520> numbers<00:51:06.000> that<00:51:06.160> we<00:51:06.400> have<00:51:06.640> here<00:51:07.040> we<00:51:07.200> call<00:51:07.440> it + +00:51:07.589 --> 00:51:07.599 align:start position:0% +the numbers that we have here we call it + + +00:51:07.599 --> 00:51:10.390 align:start position:0% +the numbers that we have here we call it +logits + +00:51:10.390 --> 00:51:10.400 align:start position:0% + + + +00:51:10.400 --> 00:51:12.309 align:start position:0% + +and<00:51:10.640> after<00:51:10.960> the<00:51:11.119> softmax<00:51:11.680> these<00:51:11.920> are<00:51:12.000> the + +00:51:12.309 --> 00:51:12.319 align:start position:0% +and after the softmax these are the + + +00:51:12.319 --> 00:51:13.430 align:start position:0% +and after the softmax these are the +probabilities + +00:51:13.430 --> 00:51:13.440 align:start position:0% +probabilities + + +00:51:13.440 --> 00:51:17.670 align:start position:0% +probabilities +for<00:51:13.760> predicting<00:51:14.240> different<00:51:15.280> classes + +00:51:17.670 --> 00:51:17.680 align:start position:0% +for predicting different classes + + +00:51:17.680 --> 00:51:21.109 align:start position:0% +for predicting different classes +so<00:51:17.920> this<00:51:18.240> whole<00:51:19.040> network<00:51:20.000> is<00:51:20.319> called + +00:51:21.109 --> 00:51:21.119 align:start position:0% +so this whole network is called + + +00:51:21.119 --> 00:51:24.390 align:start position:0% +so this whole network is called +a<00:51:21.359> multi-layer<00:51:22.160> perceptron<00:51:22.880> network + +00:51:24.390 --> 00:51:24.400 align:start position:0% +a multi-layer perceptron network + + +00:51:24.400 --> 00:51:27.990 align:start position:0% +a multi-layer perceptron network +so<00:51:24.559> this<00:51:24.800> is<00:51:24.880> called<00:51:25.280> multi + +00:51:27.990 --> 00:51:28.000 align:start position:0% + + + +00:51:28.000 --> 00:51:32.790 align:start position:0% + +layer<00:51:30.240> perceptron + +00:51:32.790 --> 00:51:32.800 align:start position:0% +layer perceptron + + +00:51:32.800 --> 00:51:35.990 align:start position:0% +layer perceptron +or<00:51:33.200> mlp<00:51:34.640> or<00:51:34.960> sometimes<00:51:35.440> people<00:51:35.760> they + +00:51:35.990 --> 00:51:36.000 align:start position:0% +or mlp or sometimes people they + + +00:51:36.000 --> 00:51:38.950 align:start position:0% +or mlp or sometimes people they +refer<00:51:36.319> to<00:51:36.480> it<00:51:36.640> as<00:51:37.119> fully<00:51:37.520> connected<00:51:38.720> feed + +00:51:38.950 --> 00:51:38.960 align:start position:0% +refer to it as fully connected feed + + +00:51:38.960 --> 00:51:43.109 align:start position:0% +refer to it as fully connected feed +forward<00:51:39.440> network + +00:51:43.109 --> 00:51:43.119 align:start position:0% + + + +00:51:43.119 --> 00:51:47.670 align:start position:0% + +connected<00:51:45.280> feed + +00:51:47.670 --> 00:51:47.680 align:start position:0% + + + +00:51:47.680 --> 00:51:51.270 align:start position:0% + +forward<00:51:48.839> network + +00:51:51.270 --> 00:51:51.280 align:start position:0% +forward network + + +00:51:51.280 --> 00:51:53.270 align:start position:0% +forward network +so<00:51:51.440> the<00:51:51.599> number<00:51:51.920> of<00:51:52.079> hidden<00:51:52.480> layers<00:51:52.880> is<00:51:53.040> the + +00:51:53.270 --> 00:51:53.280 align:start position:0% +so the number of hidden layers is the + + +00:51:53.280 --> 00:51:55.349 align:start position:0% +so the number of hidden layers is the +depth<00:51:53.680> of<00:51:53.839> your + +00:51:55.349 --> 00:51:55.359 align:start position:0% +depth of your + + +00:51:55.359 --> 00:51:58.390 align:start position:0% +depth of your +your<00:51:55.599> network<00:51:57.200> this<00:51:57.359> is<00:51:57.520> the<00:51:57.680> depth + +00:51:58.390 --> 00:51:58.400 align:start position:0% +your network this is the depth + + +00:51:58.400 --> 00:52:02.069 align:start position:0% +your network this is the depth +and<00:51:58.640> the<00:51:58.960> number<00:51:59.359> of<00:51:59.920> neurons<00:52:00.640> in<00:52:00.960> each<00:52:01.280> layer + +00:52:02.069 --> 00:52:02.079 align:start position:0% +and the number of neurons in each layer + + +00:52:02.079 --> 00:52:05.030 align:start position:0% +and the number of neurons in each layer +is<00:52:02.240> the<00:52:02.480> width<00:52:02.880> of<00:52:03.040> your<00:52:03.280> network<00:52:04.720> so<00:52:04.880> there + +00:52:05.030 --> 00:52:05.040 align:start position:0% +is the width of your network so there + + +00:52:05.040 --> 00:52:05.829 align:start position:0% +is the width of your network so there +was<00:52:05.280> a<00:52:05.359> question + +00:52:05.829 --> 00:52:05.839 align:start position:0% +was a question + + +00:52:05.839 --> 00:52:09.670 align:start position:0% +was a question +earlier<00:52:06.960> why<00:52:07.200> this<00:52:07.359> is<00:52:07.520> called<00:52:07.920> deep<00:52:08.240> learning + +00:52:09.670 --> 00:52:09.680 align:start position:0% +earlier why this is called deep learning + + +00:52:09.680 --> 00:52:12.630 align:start position:0% +earlier why this is called deep learning +because<00:52:10.160> in<00:52:10.640> most<00:52:10.960> of<00:52:11.119> the<00:52:11.599> uh + +00:52:12.630 --> 00:52:12.640 align:start position:0% +because in most of the uh + + +00:52:12.640 --> 00:52:14.710 align:start position:0% +because in most of the uh +state-of-the-art<00:52:13.440> architectures + +00:52:14.710 --> 00:52:14.720 align:start position:0% +state-of-the-art architectures + + +00:52:14.720 --> 00:52:19.270 align:start position:0% +state-of-the-art architectures +we<00:52:15.040> need<00:52:16.640> depth<00:52:17.040> to<00:52:17.200> be<00:52:17.440> high<00:52:17.839> in<00:52:18.000> order<00:52:18.400> to + +00:52:19.270 --> 00:52:19.280 align:start position:0% +we need depth to be high in order to + + +00:52:19.280 --> 00:52:21.270 align:start position:0% +we need depth to be high in order to +achieve<00:52:19.760> a<00:52:19.920> good<00:52:20.160> performance<00:52:20.800> and<00:52:20.880> most<00:52:21.119> of + +00:52:21.270 --> 00:52:21.280 align:start position:0% +achieve a good performance and most of + + +00:52:21.280 --> 00:52:23.829 align:start position:0% +achieve a good performance and most of +the<00:52:21.920> architectures<00:52:22.880> uh<00:52:23.200> some<00:52:23.440> of<00:52:23.520> them<00:52:23.680> that + +00:52:23.829 --> 00:52:23.839 align:start position:0% +the architectures uh some of them that + + +00:52:23.839 --> 00:52:24.470 align:start position:0% +the architectures uh some of them that +we'll + +00:52:24.470 --> 00:52:24.480 align:start position:0% +we'll + + +00:52:24.480 --> 00:52:27.829 align:start position:0% +we'll +see<00:52:24.800> today<00:52:25.760> uh<00:52:26.160> they<00:52:26.480> have<00:52:26.880> uh + +00:52:27.829 --> 00:52:27.839 align:start position:0% +see today uh they have uh + + +00:52:27.839 --> 00:52:30.230 align:start position:0% +see today uh they have uh +quite<00:52:28.079> a<00:52:28.160> high<00:52:28.480> depth<00:52:29.200> uh<00:52:29.520> and<00:52:29.680> that's<00:52:29.920> why<00:52:30.079> it + +00:52:30.230 --> 00:52:30.240 align:start position:0% +quite a high depth uh and that's why it + + +00:52:30.240 --> 00:52:32.549 align:start position:0% +quite a high depth uh and that's why it +is<00:52:30.319> called<00:52:30.559> deep<00:52:30.839> learning<00:52:31.920> but<00:52:32.160> believe<00:52:32.480> it + +00:52:32.549 --> 00:52:32.559 align:start position:0% +is called deep learning but believe it + + +00:52:32.559 --> 00:52:33.030 align:start position:0% +is called deep learning but believe it +or<00:52:32.720> not + +00:52:33.030 --> 00:52:33.040 align:start position:0% +or not + + +00:52:33.040 --> 00:52:35.510 align:start position:0% +or not +even<00:52:33.280> sometimes<00:52:33.839> you<00:52:33.920> know<00:52:34.319> people<00:52:34.640> they<00:52:34.800> have + +00:52:35.510 --> 00:52:35.520 align:start position:0% +even sometimes you know people they have + + +00:52:35.520 --> 00:52:37.349 align:start position:0% +even sometimes you know people they have +one<00:52:35.760> hidden<00:52:36.079> layer<00:52:36.480> and<00:52:36.640> they<00:52:36.800> call<00:52:37.040> it<00:52:37.119> deep + +00:52:37.349 --> 00:52:37.359 align:start position:0% +one hidden layer and they call it deep + + +00:52:37.359 --> 00:52:38.470 align:start position:0% +one hidden layer and they call it deep +learning + +00:52:38.470 --> 00:52:38.480 align:start position:0% +learning + + +00:52:38.480 --> 00:52:42.309 align:start position:0% +learning +but<00:52:38.640> that's<00:52:38.960> another<00:52:40.160> story<00:52:41.440> in<00:52:41.520> those<00:52:41.760> cases + +00:52:42.309 --> 00:52:42.319 align:start position:0% +but that's another story in those cases + + +00:52:42.319 --> 00:52:44.150 align:start position:0% +but that's another story in those cases +i<00:52:42.400> think<00:52:42.559> you<00:52:42.720> should<00:52:42.960> refer<00:52:43.440> them<00:52:43.599> to<00:52:43.839> as + +00:52:44.150 --> 00:52:44.160 align:start position:0% +i think you should refer them to as + + +00:52:44.160 --> 00:52:45.349 align:start position:0% +i think you should refer them to as +shallow<00:52:44.559> network + +00:52:45.349 --> 00:52:45.359 align:start position:0% +shallow network + + +00:52:45.359 --> 00:52:48.549 align:start position:0% +shallow network +or<00:52:45.680> you<00:52:45.920> can<00:52:46.160> call<00:52:46.480> them<00:52:46.720> shallow<00:52:47.200> learning<00:52:48.240> if + +00:52:48.549 --> 00:52:48.559 align:start position:0% +or you can call them shallow learning if + + +00:52:48.559 --> 00:52:49.589 align:start position:0% +or you can call them shallow learning if +the + +00:52:49.589 --> 00:52:49.599 align:start position:0% +the + + +00:52:49.599 --> 00:52:53.910 align:start position:0% +the +if<00:52:49.760> you're<00:52:50.000> using<00:52:50.240> those<00:52:50.480> types<00:52:50.800> of<00:52:51.359> networks + +00:52:53.910 --> 00:52:53.920 align:start position:0% +if you're using those types of networks + + +00:52:53.920 --> 00:53:01.750 align:start position:0% +if you're using those types of networks +that's<00:52:54.160> it<00:52:54.319> for<00:52:54.480> the<00:52:54.720> mlp<00:52:56.079> any<00:52:58.839> questions + +00:53:01.750 --> 00:53:01.760 align:start position:0% +that's it for the mlp any questions + + +00:53:01.760 --> 00:53:04.870 align:start position:0% +that's it for the mlp any questions +all<00:53:01.920> right<00:53:02.160> so<00:53:02.400> if<00:53:02.640> not<00:53:03.359> let's<00:53:03.680> move<00:53:03.920> on + +00:53:04.870 --> 00:53:04.880 align:start position:0% +all right so if not let's move on + + +00:53:04.880 --> 00:53:08.069 align:start position:0% +all right so if not let's move on +so<00:53:05.119> another<00:53:05.520> very<00:53:06.160> important<00:53:06.960> uh<00:53:07.359> network + +00:53:08.069 --> 00:53:08.079 align:start position:0% +so another very important uh network + + +00:53:08.079 --> 00:53:11.109 align:start position:0% +so another very important uh network +uh<00:53:08.400> architecture<00:53:09.119> is<00:53:09.520> the<00:53:09.760> convolution + +00:53:11.109 --> 00:53:11.119 align:start position:0% +uh architecture is the convolution + + +00:53:11.119 --> 00:53:14.069 align:start position:0% +uh architecture is the convolution +uh<00:53:11.440> convolutional<00:53:12.240> networks<00:53:13.200> convolutional + +00:53:14.069 --> 00:53:14.079 align:start position:0% +uh convolutional networks convolutional + + +00:53:14.079 --> 00:53:17.750 align:start position:0% +uh convolutional networks convolutional +neural + +00:53:17.750 --> 00:53:17.760 align:start position:0% + + + +00:53:17.760 --> 00:53:21.750 align:start position:0% + +networks<00:53:18.319> or<00:53:18.559> cnns<00:53:19.440> especially<00:53:20.079> on<00:53:20.319> images + +00:53:21.750 --> 00:53:21.760 align:start position:0% +networks or cnns especially on images + + +00:53:21.760 --> 00:53:24.470 align:start position:0% +networks or cnns especially on images +cnns<00:53:22.400> are<00:53:22.559> extremely<00:53:23.200> extremely<00:53:23.760> powerful + +00:53:24.470 --> 00:53:24.480 align:start position:0% +cnns are extremely extremely powerful + + +00:53:24.480 --> 00:53:26.790 align:start position:0% +cnns are extremely extremely powerful +and<00:53:24.720> most<00:53:25.040> of<00:53:25.119> the<00:53:25.359> state-of-the-art + +00:53:26.790 --> 00:53:26.800 align:start position:0% +and most of the state-of-the-art + + +00:53:26.800 --> 00:53:29.510 align:start position:0% +and most of the state-of-the-art +architectures<00:53:27.520> they<00:53:27.680> use<00:53:27.920> cns<00:53:29.040> all<00:53:29.200> right<00:53:29.359> so + +00:53:29.510 --> 00:53:29.520 align:start position:0% +architectures they use cns all right so + + +00:53:29.520 --> 00:53:30.470 align:start position:0% +architectures they use cns all right so +what<00:53:29.760> is<00:53:29.839> the<00:53:30.000> basic + +00:53:30.470 --> 00:53:30.480 align:start position:0% +what is the basic + + +00:53:30.480 --> 00:53:34.230 align:start position:0% +what is the basic +idea<00:53:31.440> for<00:53:31.680> so<00:53:31.920> in<00:53:32.079> mlp<00:53:32.960> in<00:53:33.200> one<00:53:33.440> of<00:53:33.599> the + +00:53:34.230 --> 00:53:34.240 align:start position:0% +idea for so in mlp in one of the + + +00:53:34.240 --> 00:53:36.309 align:start position:0% +idea for so in mlp in one of the +neurons<00:53:34.720> if<00:53:34.880> you<00:53:35.440> you<00:53:35.599> know<00:53:35.839> if<00:53:36.000> you<00:53:36.079> think + +00:53:36.309 --> 00:53:36.319 align:start position:0% +neurons if you you know if you think + + +00:53:36.319 --> 00:53:37.589 align:start position:0% +neurons if you you know if you think +about<00:53:36.559> this<00:53:36.720> so<00:53:36.880> you're<00:53:37.119> applying<00:53:37.440> an + +00:53:37.589 --> 00:53:37.599 align:start position:0% +about this so you're applying an + + +00:53:37.599 --> 00:53:38.950 align:start position:0% +about this so you're applying an +activation<00:53:38.160> function + +00:53:38.950 --> 00:53:38.960 align:start position:0% +activation function + + +00:53:38.960 --> 00:53:43.109 align:start position:0% +activation function +to<00:53:39.200> an<00:53:39.440> affine<00:53:39.839> transformation<00:53:40.720> of<00:53:41.920> x + +00:53:43.109 --> 00:53:43.119 align:start position:0% +to an affine transformation of x + + +00:53:43.119 --> 00:53:46.150 align:start position:0% +to an affine transformation of x +so<00:53:43.440> in<00:53:44.079> a<00:53:44.319> convolutional<00:53:45.040> network + +00:53:46.150 --> 00:53:46.160 align:start position:0% +so in a convolutional network + + +00:53:46.160 --> 00:53:48.150 align:start position:0% +so in a convolutional network +we<00:53:46.319> are<00:53:46.720> going<00:53:46.960> to<00:53:47.119> use<00:53:47.359> a<00:53:47.440> convolution + +00:53:48.150 --> 00:53:48.160 align:start position:0% +we are going to use a convolution + + +00:53:48.160 --> 00:53:49.510 align:start position:0% +we are going to use a convolution +operator + +00:53:49.510 --> 00:53:49.520 align:start position:0% +operator + + +00:53:49.520 --> 00:53:52.790 align:start position:0% +operator +in<00:53:49.599> order<00:53:50.000> to<00:53:50.400> transform<00:53:51.040> linearly<00:53:51.680> transform + +00:53:52.790 --> 00:53:52.800 align:start position:0% +in order to transform linearly transform + + +00:53:52.800 --> 00:53:56.150 align:start position:0% +in order to transform linearly transform +x<00:53:53.119> to<00:53:53.440> some<00:53:53.920> um<00:53:54.640> some<00:53:54.880> other<00:53:55.119> vectors<00:53:55.839> so<00:53:56.000> we + +00:53:56.150 --> 00:53:56.160 align:start position:0% +x to some um some other vectors so we + + +00:53:56.160 --> 00:53:57.510 align:start position:0% +x to some um some other vectors so we +are<00:53:56.240> going<00:53:56.480> to<00:53:56.640> combine + +00:53:57.510 --> 00:53:57.520 align:start position:0% +are going to combine + + +00:53:57.520 --> 00:54:00.790 align:start position:0% +are going to combine +x<00:53:58.640> uh<00:53:58.960> with<00:53:59.920> w + +00:54:00.790 --> 00:54:00.800 align:start position:0% +x uh with w + + +00:54:00.800 --> 00:54:03.190 align:start position:0% +x uh with w +and<00:54:00.880> that<00:54:01.119> will<00:54:01.280> be<00:54:01.440> the<00:54:01.680> output<00:54:02.079> of<00:54:02.800> the + +00:54:03.190 --> 00:54:03.200 align:start position:0% +and that will be the output of the + + +00:54:03.200 --> 00:54:04.950 align:start position:0% +and that will be the output of the +convolution<00:54:04.000> so<00:54:04.240> this<00:54:04.400> is<00:54:04.559> called<00:54:04.800> the + +00:54:04.950 --> 00:54:04.960 align:start position:0% +convolution so this is called the + + +00:54:04.960 --> 00:54:06.270 align:start position:0% +convolution so this is called the +convolution + +00:54:06.270 --> 00:54:06.280 align:start position:0% +convolution + + +00:54:06.280 --> 00:54:08.829 align:start position:0% +convolution +[Music] + +00:54:08.829 --> 00:54:08.839 align:start position:0% +[Music] + + +00:54:08.839 --> 00:54:11.510 align:start position:0% +[Music] +operator + +00:54:11.510 --> 00:54:11.520 align:start position:0% +operator + + +00:54:11.520 --> 00:54:14.710 align:start position:0% +operator +again<00:54:12.079> similar<00:54:12.480> to<00:54:12.640> w<00:54:12.960> transpose<00:54:13.440> x<00:54:13.680> it<00:54:13.839> is<00:54:14.319> a + +00:54:14.710 --> 00:54:14.720 align:start position:0% +again similar to w transpose x it is a + + +00:54:14.720 --> 00:54:17.829 align:start position:0% +again similar to w transpose x it is a +and<00:54:14.960> a<00:54:15.119> fine<00:54:15.680> uh<00:54:16.160> operator<00:54:17.440> so + +00:54:17.829 --> 00:54:17.839 align:start position:0% +and a fine uh operator so + + +00:54:17.839 --> 00:54:21.349 align:start position:0% +and a fine uh operator so +it<00:54:18.000> is<00:54:18.160> not<00:54:18.400> uh<00:54:19.520> adding<00:54:20.160> non-linearity<00:54:21.040> to<00:54:21.280> our + +00:54:21.349 --> 00:54:21.359 align:start position:0% +it is not uh adding non-linearity to our + + +00:54:21.359 --> 00:54:23.109 align:start position:0% +it is not uh adding non-linearity to our +system<00:54:21.680> so<00:54:21.920> the<00:54:22.079> non-linearity + +00:54:23.109 --> 00:54:23.119 align:start position:0% +system so the non-linearity + + +00:54:23.119 --> 00:54:25.030 align:start position:0% +system so the non-linearity +it's<00:54:23.280> going<00:54:23.440> to<00:54:23.599> come<00:54:23.920> using<00:54:24.240> this<00:54:24.800> phi + +00:54:25.030 --> 00:54:25.040 align:start position:0% +it's going to come using this phi + + +00:54:25.040 --> 00:54:26.390 align:start position:0% +it's going to come using this phi +transformation + +00:54:26.390 --> 00:54:26.400 align:start position:0% +transformation + + +00:54:26.400 --> 00:54:29.190 align:start position:0% +transformation +so<00:54:26.559> let's<00:54:27.280> look<00:54:27.520> at<00:54:27.680> cnns<00:54:28.160> with<00:54:28.400> a<00:54:28.480> very<00:54:28.800> simple + +00:54:29.190 --> 00:54:29.200 align:start position:0% +so let's look at cnns with a very simple + + +00:54:29.200 --> 00:54:33.670 align:start position:0% +so let's look at cnns with a very simple +example<00:54:29.839> let's<00:54:30.079> look<00:54:30.400> at<00:54:30.640> a<00:54:30.880> one-dimensional + +00:54:33.670 --> 00:54:33.680 align:start position:0% +example let's look at a one-dimensional + + +00:54:33.680 --> 00:54:38.390 align:start position:0% +example let's look at a one-dimensional +example<00:54:36.480> so<00:54:36.640> i<00:54:36.720> don't<00:54:36.880> want<00:54:37.040> to<00:54:37.119> go<00:54:37.280> into<00:54:37.599> the + +00:54:38.390 --> 00:54:38.400 align:start position:0% +example so i don't want to go into the + + +00:54:38.400 --> 00:54:40.230 align:start position:0% +example so i don't want to go into the +mathematical<00:54:39.200> notation<00:54:39.680> because<00:54:40.079> the + +00:54:40.230 --> 00:54:40.240 align:start position:0% +mathematical notation because the + + +00:54:40.240 --> 00:54:43.910 align:start position:0% +mathematical notation because the +concept<00:54:40.720> is<00:54:40.880> in<00:54:41.040> fact<00:54:41.280> very<00:54:41.680> simple + +00:54:43.910 --> 00:54:43.920 align:start position:0% +concept is in fact very simple + + +00:54:43.920 --> 00:54:47.589 align:start position:0% +concept is in fact very simple +so<00:54:44.160> here<00:54:44.400> we<00:54:44.559> have<00:54:44.880> x<00:54:46.559> let's<00:54:46.720> say<00:54:46.960> we<00:54:47.119> have<00:54:47.359> a + +00:54:47.589 --> 00:54:47.599 align:start position:0% +so here we have x let's say we have a + + +00:54:47.599 --> 00:54:49.430 align:start position:0% +so here we have x let's say we have a +four<00:54:47.920> dimensional<00:54:48.799> x + +00:54:49.430 --> 00:54:49.440 align:start position:0% +four dimensional x + + +00:54:49.440 --> 00:54:53.510 align:start position:0% +four dimensional x +with<00:54:49.680> the<00:54:49.839> numbers<00:54:50.319> 1<00:54:50.640> 3<00:54:51.599> ten<00:54:52.000> five + +00:54:53.510 --> 00:54:53.520 align:start position:0% +with the numbers 1 3 ten five + + +00:54:53.520 --> 00:54:56.230 align:start position:0% +with the numbers 1 3 ten five +these<00:54:53.760> are<00:54:53.920> my<00:54:54.079> pixel<00:54:54.480> values<00:54:55.680> now<00:54:55.920> i'm<00:54:56.000> gonna + +00:54:56.230 --> 00:54:56.240 align:start position:0% +these are my pixel values now i'm gonna + + +00:54:56.240 --> 00:54:57.349 align:start position:0% +these are my pixel values now i'm gonna +have<00:54:56.400> a<00:54:56.559> filter + +00:54:57.349 --> 00:54:57.359 align:start position:0% +have a filter + + +00:54:57.359 --> 00:55:00.390 align:start position:0% +have a filter +let's<00:54:57.599> say<00:54:58.240> my<00:54:58.480> filter<00:54:59.040> is<00:54:59.839> something<00:55:00.240> like + +00:55:00.390 --> 00:55:00.400 align:start position:0% +let's say my filter is something like + + +00:55:00.400 --> 00:55:01.589 align:start position:0% +let's say my filter is something like +this + +00:55:01.589 --> 00:55:01.599 align:start position:0% +this + + +00:55:01.599 --> 00:55:05.109 align:start position:0% +this +plus<00:55:01.920> one<00:55:02.160> minus<00:55:02.559> one<00:55:02.799> so<00:55:02.960> this<00:55:03.200> is<00:55:03.280> my<00:55:03.520> w + +00:55:05.109 --> 00:55:05.119 align:start position:0% +plus one minus one so this is my w + + +00:55:05.119 --> 00:55:07.270 align:start position:0% +plus one minus one so this is my w +these<00:55:05.359> are<00:55:05.520> my<00:55:05.760> bits<00:55:06.400> so<00:55:06.559> i'm<00:55:06.720> going<00:55:06.799> to<00:55:07.040> look + +00:55:07.270 --> 00:55:07.280 align:start position:0% +these are my bits so i'm going to look + + +00:55:07.280 --> 00:55:09.190 align:start position:0% +these are my bits so i'm going to look +at<00:55:07.359> the<00:55:07.520> inner<00:55:07.839> product<00:55:08.240> between<00:55:08.559> this<00:55:08.799> vector + +00:55:09.190 --> 00:55:09.200 align:start position:0% +at the inner product between this vector + + +00:55:09.200 --> 00:55:10.470 align:start position:0% +at the inner product between this vector +and<00:55:09.359> this<00:55:09.599> part<00:55:09.920> of<00:55:10.079> my + +00:55:10.470 --> 00:55:10.480 align:start position:0% +and this part of my + + +00:55:10.480 --> 00:55:12.950 align:start position:0% +and this part of my +image<00:55:11.440> so<00:55:11.760> what<00:55:12.000> is<00:55:12.079> the<00:55:12.319> result<00:55:12.640> of<00:55:12.720> that + +00:55:12.950 --> 00:55:12.960 align:start position:0% +image so what is the result of that + + +00:55:12.960 --> 00:55:14.870 align:start position:0% +image so what is the result of that +inner<00:55:13.200> product<00:55:13.599> 1<00:55:13.760> times<00:55:14.079> 1 + +00:55:14.870 --> 00:55:14.880 align:start position:0% +inner product 1 times 1 + + +00:55:14.880 --> 00:55:18.710 align:start position:0% +inner product 1 times 1 +plus<00:55:15.359> 3<00:55:15.599> times<00:55:15.920> -1<00:55:16.400> so<00:55:16.559> i'll<00:55:16.640> get<00:55:16.799> the<00:55:16.960> minus<00:55:17.839> 2 + +00:55:18.710 --> 00:55:18.720 align:start position:0% +plus 3 times -1 so i'll get the minus 2 + + +00:55:18.720 --> 00:55:22.630 align:start position:0% +plus 3 times -1 so i'll get the minus 2 +as<00:55:19.119> my<00:55:19.680> output<00:55:21.200> you<00:55:21.359> know<00:55:21.599> basically + +00:55:22.630 --> 00:55:22.640 align:start position:0% +as my output you know basically + + +00:55:22.640 --> 00:55:26.069 align:start position:0% +as my output you know basically +report<00:55:24.400> -2 + +00:55:26.069 --> 00:55:26.079 align:start position:0% +report -2 + + +00:55:26.079 --> 00:55:28.470 align:start position:0% +report -2 +and<00:55:26.240> then<00:55:26.640> with<00:55:26.880> the<00:55:27.200> parameter<00:55:27.839> which<00:55:28.079> is<00:55:28.160> the + +00:55:28.470 --> 00:55:28.480 align:start position:0% +and then with the parameter which is the + + +00:55:28.480 --> 00:55:29.910 align:start position:0% +and then with the parameter which is the +stride<00:55:29.119> i'm<00:55:29.359> going<00:55:29.440> to + +00:55:29.910 --> 00:55:29.920 align:start position:0% +stride i'm going to + + +00:55:29.920 --> 00:55:33.270 align:start position:0% +stride i'm going to +shift<00:55:30.319> my<00:55:30.960> weight<00:55:32.000> i'm<00:55:32.160> going<00:55:32.319> to<00:55:32.799> maybe + +00:55:33.270 --> 00:55:33.280 align:start position:0% +shift my weight i'm going to maybe + + +00:55:33.280 --> 00:55:37.589 align:start position:0% +shift my weight i'm going to maybe +shift<00:55:33.520> by<00:55:33.839> one<00:55:34.240> if<00:55:34.480> my<00:55:35.280> stride + +00:55:37.589 --> 00:55:37.599 align:start position:0% +shift by one if my stride + + +00:55:37.599 --> 00:55:39.589 align:start position:0% +shift by one if my stride +is<00:55:37.920> one<00:55:38.400> if<00:55:38.559> my<00:55:38.720> stride<00:55:39.040> is<00:55:39.119> two<00:55:39.280> i'm<00:55:39.440> gonna + +00:55:39.589 --> 00:55:39.599 align:start position:0% +is one if my stride is two i'm gonna + + +00:55:39.599 --> 00:55:40.630 align:start position:0% +is one if my stride is two i'm gonna +shift<00:55:39.839> by<00:55:40.079> two + +00:55:40.630 --> 00:55:40.640 align:start position:0% +shift by two + + +00:55:40.640 --> 00:55:42.309 align:start position:0% +shift by two +and<00:55:40.880> again<00:55:41.200> i'm<00:55:41.359> gonna<00:55:41.599> look<00:55:41.760> at<00:55:41.839> the<00:55:42.000> inner + +00:55:42.309 --> 00:55:42.319 align:start position:0% +and again i'm gonna look at the inner + + +00:55:42.319 --> 00:55:44.789 align:start position:0% +and again i'm gonna look at the inner +product<00:55:43.040> the<00:55:43.280> similarity<00:55:43.920> of<00:55:44.000> my<00:55:44.240> filter<00:55:44.559> with + +00:55:44.789 --> 00:55:44.799 align:start position:0% +product the similarity of my filter with + + +00:55:44.799 --> 00:55:46.870 align:start position:0% +product the similarity of my filter with +this<00:55:45.040> portion<00:55:45.520> of<00:55:45.760> my + +00:55:46.870 --> 00:55:46.880 align:start position:0% +this portion of my + + +00:55:46.880 --> 00:55:49.349 align:start position:0% +this portion of my +input<00:55:47.599> in<00:55:47.680> that<00:55:47.920> case<00:55:48.160> i'll<00:55:48.400> get<00:55:48.640> a<00:55:49.040> minus + +00:55:49.349 --> 00:55:49.359 align:start position:0% +input in that case i'll get a minus + + +00:55:49.359 --> 00:55:51.190 align:start position:0% +input in that case i'll get a minus +seven<00:55:49.760> for<00:55:49.920> instance + +00:55:51.190 --> 00:55:51.200 align:start position:0% +seven for instance + + +00:55:51.200 --> 00:55:54.710 align:start position:0% +seven for instance +and<00:55:51.760> keep<00:55:52.000> going<00:55:53.680> at<00:55:53.839> the<00:55:54.079> ends<00:55:54.400> you<00:55:54.559> know + +00:55:54.710 --> 00:55:54.720 align:start position:0% +and keep going at the ends you know + + +00:55:54.720 --> 00:55:55.270 align:start position:0% +and keep going at the ends you know +there<00:55:54.960> are + +00:55:55.270 --> 00:55:55.280 align:start position:0% +there are + + +00:55:55.280 --> 00:55:57.829 align:start position:0% +there are +uh<00:55:56.000> you<00:55:56.079> know<00:55:56.319> different<00:55:56.640> ways<00:55:57.040> you<00:55:57.200> can<00:55:57.520> pad + +00:55:57.829 --> 00:55:57.839 align:start position:0% +uh you know different ways you can pad + + +00:55:57.839 --> 00:56:00.309 align:start position:0% +uh you know different ways you can pad +zeros<00:55:58.319> in<00:55:58.480> order<00:55:58.799> to + +00:56:00.309 --> 00:56:00.319 align:start position:0% +zeros in order to + + +00:56:00.319 --> 00:56:02.230 align:start position:0% +zeros in order to +look<00:56:00.480> into<00:56:00.720> the<00:56:00.880> edges<00:56:01.280> or<00:56:01.440> you<00:56:01.599> can<00:56:02.000> you<00:56:02.079> know + +00:56:02.230 --> 00:56:02.240 align:start position:0% +look into the edges or you can you know + + +00:56:02.240 --> 00:56:04.309 align:start position:0% +look into the edges or you can you know +do<00:56:02.400> a<00:56:02.559> circular<00:56:03.040> convolution + +00:56:04.309 --> 00:56:04.319 align:start position:0% +do a circular convolution + + +00:56:04.319 --> 00:56:06.789 align:start position:0% +do a circular convolution +uh<00:56:05.200> those<00:56:05.440> are<00:56:05.599> like<00:56:05.920> minor<00:56:06.319> details<00:56:06.720> but + +00:56:06.789 --> 00:56:06.799 align:start position:0% +uh those are like minor details but + + +00:56:06.799 --> 00:56:08.789 align:start position:0% +uh those are like minor details but +that's<00:56:07.119> the<00:56:07.359> basic<00:56:07.839> idea<00:56:08.319> of<00:56:08.480> the + +00:56:08.789 --> 00:56:08.799 align:start position:0% +that's the basic idea of the + + +00:56:08.799 --> 00:56:11.349 align:start position:0% +that's the basic idea of the +convolution<00:56:09.440> operator<00:56:10.559> and<00:56:10.640> then<00:56:10.799> you<00:56:11.040> apply + +00:56:11.349 --> 00:56:11.359 align:start position:0% +convolution operator and then you apply + + +00:56:11.359 --> 00:56:13.030 align:start position:0% +convolution operator and then you apply +your<00:56:11.599> non-linearity + +00:56:13.030 --> 00:56:13.040 align:start position:0% +your non-linearity + + +00:56:13.040 --> 00:56:16.309 align:start position:0% +your non-linearity +on<00:56:13.520> top<00:56:13.839> of<00:56:14.000> it<00:56:14.079> so<00:56:14.400> you<00:56:14.640> have<00:56:15.760> let's<00:56:16.000> say + +00:56:16.309 --> 00:56:16.319 align:start position:0% +on top of it so you have let's say + + +00:56:16.319 --> 00:56:20.710 align:start position:0% +on top of it so you have let's say +a<00:56:16.640> relu<00:56:17.599> or<00:56:18.880> any<00:56:19.200> other<00:56:19.520> activation<00:56:20.079> functions + +00:56:20.710 --> 00:56:20.720 align:start position:0% +a relu or any other activation functions + + +00:56:20.720 --> 00:56:21.030 align:start position:0% +a relu or any other activation functions +and + +00:56:21.030 --> 00:56:21.040 align:start position:0% +and + + +00:56:21.040 --> 00:56:24.789 align:start position:0% +and +apply<00:56:21.359> your<00:56:21.599> non-linearity<00:56:22.880> on<00:56:23.680> top<00:56:24.000> of + +00:56:24.789 --> 00:56:24.799 align:start position:0% +apply your non-linearity on top of + + +00:56:24.799 --> 00:56:29.190 align:start position:0% +apply your non-linearity on top of +these<00:56:26.400> outputs<00:56:26.960> of<00:56:27.119> your<00:56:27.760> convolution + +00:56:29.190 --> 00:56:29.200 align:start position:0% +these outputs of your convolution + + +00:56:29.200 --> 00:56:32.470 align:start position:0% +these outputs of your convolution +operation + +00:56:32.470 --> 00:56:32.480 align:start position:0% + + + +00:56:32.480 --> 00:56:34.789 align:start position:0% + +okay<00:56:32.720> so<00:56:32.880> what<00:56:33.119> would<00:56:33.280> be<00:56:33.520> my<00:56:33.760> output<00:56:34.240> size + +00:56:34.789 --> 00:56:34.799 align:start position:0% +okay so what would be my output size + + +00:56:34.799 --> 00:56:36.710 align:start position:0% +okay so what would be my output size +roughly<00:56:35.200> speaking<00:56:35.599> if<00:56:35.760> my<00:56:36.000> input<00:56:36.319> let's<00:56:36.559> say + +00:56:36.710 --> 00:56:36.720 align:start position:0% +roughly speaking if my input let's say + + +00:56:36.720 --> 00:56:38.470 align:start position:0% +roughly speaking if my input let's say +is<00:56:36.880> in<00:56:37.040> d<00:56:37.200> dimension + +00:56:38.470 --> 00:56:38.480 align:start position:0% +is in d dimension + + +00:56:38.480 --> 00:56:41.510 align:start position:0% +is in d dimension +and<00:56:38.720> if<00:56:38.960> my<00:56:39.200> stride<00:56:39.680> is<00:56:39.839> a<00:56:40.000> certain<00:56:40.400> number + +00:56:41.510 --> 00:56:41.520 align:start position:0% +and if my stride is a certain number + + +00:56:41.520 --> 00:56:46.150 align:start position:0% +and if my stride is a certain number +my<00:56:41.920> output<00:56:42.799> size<00:56:43.280> here + +00:56:46.150 --> 00:56:46.160 align:start position:0% + + + +00:56:46.160 --> 00:56:48.630 align:start position:0% + +not<00:56:46.480> exactly<00:56:47.280> but<00:56:47.520> approximately<00:56:48.319> it<00:56:48.480> is + +00:56:48.630 --> 00:56:48.640 align:start position:0% +not exactly but approximately it is + + +00:56:48.640 --> 00:56:50.150 align:start position:0% +not exactly but approximately it is +going<00:56:48.720> to<00:56:48.880> be<00:56:49.040> my<00:56:49.280> input<00:56:49.680> size + +00:56:50.150 --> 00:56:50.160 align:start position:0% +going to be my input size + + +00:56:50.160 --> 00:56:53.670 align:start position:0% +going to be my input size +over<00:56:50.480> the<00:56:50.880> stripe<00:56:52.480> number + +00:56:53.670 --> 00:56:53.680 align:start position:0% +over the stripe number + + +00:56:53.680 --> 00:56:56.309 align:start position:0% +over the stripe number +there<00:56:53.920> are<00:56:54.079> some<00:56:54.400> you<00:56:54.559> know<00:56:55.440> you<00:56:55.599> know<00:56:56.079> edge + +00:56:56.309 --> 00:56:56.319 align:start position:0% +there are some you know you know edge + + +00:56:56.319 --> 00:56:58.390 align:start position:0% +there are some you know you know edge +cases<00:56:56.720> that<00:56:56.880> we<00:56:57.040> should<00:56:57.280> consider + +00:56:58.390 --> 00:56:58.400 align:start position:0% +cases that we should consider + + +00:56:58.400 --> 00:57:01.750 align:start position:0% +cases that we should consider +but<00:56:59.280> when<00:56:59.440> d<00:56:59.680> is<00:56:59.920> large<00:57:00.400> enough<00:57:00.880> so<00:57:01.119> my<00:57:01.359> output + +00:57:01.750 --> 00:57:01.760 align:start position:0% +but when d is large enough so my output + + +00:57:01.760 --> 00:57:04.230 align:start position:0% +but when d is large enough so my output +size<00:57:02.000> would<00:57:02.160> converge<00:57:02.640> to<00:57:02.880> this<00:57:03.119> number + +00:57:04.230 --> 00:57:04.240 align:start position:0% +size would converge to this number + + +00:57:04.240 --> 00:57:06.069 align:start position:0% +size would converge to this number +and<00:57:04.400> sometimes<00:57:04.880> you<00:57:04.960> know<00:57:05.359> when<00:57:05.520> your<00:57:05.680> stride + +00:57:06.069 --> 00:57:06.079 align:start position:0% +and sometimes you know when your stride + + +00:57:06.079 --> 00:57:07.750 align:start position:0% +and sometimes you know when your stride +is<00:57:06.240> one<00:57:06.640> it<00:57:06.799> means<00:57:07.040> you<00:57:07.200> are<00:57:07.280> not + +00:57:07.750 --> 00:57:07.760 align:start position:0% +is one it means you are not + + +00:57:07.760 --> 00:57:10.630 align:start position:0% +is one it means you are not +reducing<00:57:08.319> your<00:57:08.480> dimensionality<00:57:09.520> after<00:57:10.400> uh + +00:57:10.630 --> 00:57:10.640 align:start position:0% +reducing your dimensionality after uh + + +00:57:10.640 --> 00:57:13.030 align:start position:0% +reducing your dimensionality after uh +doing<00:57:10.880> this<00:57:11.040> convolutional<00:57:11.680> operation + +00:57:13.030 --> 00:57:13.040 align:start position:0% +doing this convolutional operation + + +00:57:13.040 --> 00:57:15.510 align:start position:0% +doing this convolutional operation +and<00:57:13.440> sometimes<00:57:13.920> we<00:57:14.240> actually<00:57:14.640> need<00:57:14.880> to<00:57:15.119> reduce + +00:57:15.510 --> 00:57:15.520 align:start position:0% +and sometimes we actually need to reduce + + +00:57:15.520 --> 00:57:16.789 align:start position:0% +and sometimes we actually need to reduce +the<00:57:15.599> dimensionality + +00:57:16.789 --> 00:57:16.799 align:start position:0% +the dimensionality + + +00:57:16.799 --> 00:57:19.990 align:start position:0% +the dimensionality +for<00:57:17.599> that<00:57:17.920> purpose<00:57:18.400> there<00:57:18.640> is<00:57:19.040> oftentimes<00:57:19.760> a + +00:57:19.990 --> 00:57:20.000 align:start position:0% +for that purpose there is oftentimes a + + +00:57:20.000 --> 00:57:20.630 align:start position:0% +for that purpose there is oftentimes a +pooling + +00:57:20.630 --> 00:57:20.640 align:start position:0% +pooling + + +00:57:20.640 --> 00:57:24.230 align:start position:0% +pooling +step<00:57:21.680> that<00:57:21.920> we<00:57:22.880> let's<00:57:23.040> say<00:57:23.359> looking<00:57:23.760> to + +00:57:24.230 --> 00:57:24.240 align:start position:0% +step that we let's say looking to + + +00:57:24.240 --> 00:57:27.670 align:start position:0% +step that we let's say looking to +adjacent<00:57:25.440> pixels<00:57:26.000> if<00:57:26.160> my<00:57:26.799> pulling<00:57:27.119> stride<00:57:27.520> is + +00:57:27.670 --> 00:57:27.680 align:start position:0% +adjacent pixels if my pulling stride is + + +00:57:27.680 --> 00:57:28.230 align:start position:0% +adjacent pixels if my pulling stride is +two + +00:57:28.230 --> 00:57:28.240 align:start position:0% +two + + +00:57:28.240 --> 00:57:31.510 align:start position:0% +two +and<00:57:28.400> i'm<00:57:29.040> gonna<00:57:29.359> take<00:57:29.599> the<00:57:29.839> maximum<00:57:31.040> or + +00:57:31.510 --> 00:57:31.520 align:start position:0% +and i'm gonna take the maximum or + + +00:57:31.520 --> 00:57:35.270 align:start position:0% +and i'm gonna take the maximum or +uh<00:57:32.000> sum<00:57:32.559> of<00:57:32.880> these<00:57:33.280> uh<00:57:34.160> values + +00:57:35.270 --> 00:57:35.280 align:start position:0% +uh sum of these uh values + + +00:57:35.280 --> 00:57:36.789 align:start position:0% +uh sum of these uh values +and<00:57:35.599> i'm<00:57:35.839> going<00:57:36.000> to<00:57:36.160> reduce<00:57:36.640> my + +00:57:36.789 --> 00:57:36.799 align:start position:0% +and i'm going to reduce my + + +00:57:36.799 --> 00:57:38.870 align:start position:0% +and i'm going to reduce my +dimensionality<00:57:37.680> based<00:57:38.000> on<00:57:38.160> that<00:57:38.319> so<00:57:38.559> this<00:57:38.720> is + +00:57:38.870 --> 00:57:38.880 align:start position:0% +dimensionality based on that so this is + + +00:57:38.880 --> 00:57:39.589 align:start position:0% +dimensionality based on that so this is +the + +00:57:39.589 --> 00:57:39.599 align:start position:0% +the + + +00:57:39.599 --> 00:57:42.630 align:start position:0% +the +pooling<00:57:41.040> layer + +00:57:42.630 --> 00:57:42.640 align:start position:0% +pooling layer + + +00:57:42.640 --> 00:57:46.069 align:start position:0% +pooling layer +oftentimes<00:57:43.280> it<00:57:43.440> is<00:57:43.760> added<00:57:45.280> for + +00:57:46.069 --> 00:57:46.079 align:start position:0% +oftentimes it is added for + + +00:57:46.079 --> 00:57:48.870 align:start position:0% +oftentimes it is added for +on<00:57:46.240> top<00:57:46.480> of<00:57:46.640> the<00:57:47.040> convolution<00:57:47.599> object<00:57:48.559> so<00:57:48.720> in + +00:57:48.870 --> 00:57:48.880 align:start position:0% +on top of the convolution object so in + + +00:57:48.880 --> 00:57:49.910 align:start position:0% +on top of the convolution object so in +other<00:57:49.119> words + +00:57:49.910 --> 00:57:49.920 align:start position:0% +other words + + +00:57:49.920 --> 00:57:53.589 align:start position:0% +other words +a<00:57:50.079> convolution<00:57:50.799> layer<00:57:52.559> you<00:57:52.720> can<00:57:52.960> think<00:57:53.200> about + +00:57:53.589 --> 00:57:53.599 align:start position:0% +a convolution layer you can think about + + +00:57:53.599 --> 00:57:56.630 align:start position:0% +a convolution layer you can think about +it + +00:57:56.630 --> 00:57:56.640 align:start position:0% + + + +00:57:56.640 --> 00:57:59.750 align:start position:0% + +convolution<00:57:57.280> layer<00:57:59.200> you<00:57:59.359> can + +00:57:59.750 --> 00:57:59.760 align:start position:0% +convolution layer you can + + +00:57:59.760 --> 00:58:01.430 align:start position:0% +convolution layer you can +think<00:58:00.000> about<00:58:00.319> it<00:58:00.480> as<00:58:00.559> the<00:58:00.720> following<00:58:01.119> so<00:58:01.280> you + +00:58:01.430 --> 00:58:01.440 align:start position:0% +think about it as the following so you + + +00:58:01.440 --> 00:58:03.190 align:start position:0% +think about it as the following so you +have<00:58:01.680> your<00:58:02.000> input<00:58:02.400> x + +00:58:03.190 --> 00:58:03.200 align:start position:0% +have your input x + + +00:58:03.200 --> 00:58:06.549 align:start position:0% +have your input x +then<00:58:03.359> you<00:58:03.520> have<00:58:03.760> a<00:58:04.640> convolution<00:58:05.839> operator + +00:58:06.549 --> 00:58:06.559 align:start position:0% +then you have a convolution operator + + +00:58:06.559 --> 00:58:08.870 align:start position:0% +then you have a convolution operator +this<00:58:06.799> is<00:58:06.880> an<00:58:07.119> accounting<00:58:07.520> transformation + +00:58:08.870 --> 00:58:08.880 align:start position:0% +this is an accounting transformation + + +00:58:08.880 --> 00:58:11.589 align:start position:0% +this is an accounting transformation +you<00:58:09.040> apply<00:58:09.440> your<00:58:09.599> activation<00:58:10.319> function<00:58:11.359> and + +00:58:11.589 --> 00:58:11.599 align:start position:0% +you apply your activation function and + + +00:58:11.599 --> 00:58:13.589 align:start position:0% +you apply your activation function and +there<00:58:11.839> may<00:58:12.000> be<00:58:12.240> a<00:58:12.480> pooling + +00:58:13.589 --> 00:58:13.599 align:start position:0% +there may be a pooling + + +00:58:13.599 --> 00:58:17.190 align:start position:0% +there may be a pooling +layer<00:58:14.960> you<00:58:15.119> can<00:58:15.280> do<00:58:15.440> a<00:58:15.599> max<00:58:15.920> pooling<00:58:16.480> or + +00:58:17.190 --> 00:58:17.200 align:start position:0% +layer you can do a max pooling or + + +00:58:17.200 --> 00:58:20.309 align:start position:0% +layer you can do a max pooling or +a<00:58:17.359> brush<00:58:17.680> pulling<00:58:19.040> and<00:58:19.280> this<00:58:19.520> whole<00:58:19.760> thing<00:58:20.079> is + +00:58:20.309 --> 00:58:20.319 align:start position:0% +a brush pulling and this whole thing is + + +00:58:20.319 --> 00:58:20.710 align:start position:0% +a brush pulling and this whole thing is +called + +00:58:20.710 --> 00:58:20.720 align:start position:0% +called + + +00:58:20.720 --> 00:58:24.789 align:start position:0% +called +a<00:58:20.880> convolution<00:58:21.440> layer<00:58:23.680> so<00:58:23.920> convolutions<00:58:24.559> are + +00:58:24.789 --> 00:58:24.799 align:start position:0% +a convolution layer so convolutions are + + +00:58:24.799 --> 00:58:27.109 align:start position:0% +a convolution layer so convolutions are +extremely<00:58:25.440> extremely<00:58:26.000> important + +00:58:27.109 --> 00:58:27.119 align:start position:0% +extremely extremely important + + +00:58:27.119 --> 00:58:33.190 align:start position:0% +extremely extremely important +uh<00:58:28.160> in<00:58:29.839> order<00:58:30.240> to + +00:58:33.190 --> 00:58:33.200 align:start position:0% + + + +00:58:33.200 --> 00:58:36.230 align:start position:0% + +basically<00:58:33.760> deal<00:58:34.000> with<00:58:34.240> image<00:58:34.839> applications + +00:58:36.230 --> 00:58:36.240 align:start position:0% +basically deal with image applications + + +00:58:36.240 --> 00:58:38.470 align:start position:0% +basically deal with image applications +just<00:58:36.400> to<00:58:36.559> give<00:58:36.720> you<00:58:36.880> an<00:58:37.040> example<00:58:37.680> alexnet + +00:58:38.470 --> 00:58:38.480 align:start position:0% +just to give you an example alexnet + + +00:58:38.480 --> 00:58:40.950 align:start position:0% +just to give you an example alexnet +which<00:58:38.640> is<00:58:38.880> one<00:58:39.119> of<00:58:39.280> the<00:58:39.520> very<00:58:39.920> popular + +00:58:40.950 --> 00:58:40.960 align:start position:0% +which is one of the very popular + + +00:58:40.960 --> 00:58:44.789 align:start position:0% +which is one of the very popular +architectures<00:58:42.160> on<00:58:43.680> classifications<00:58:44.559> on + +00:58:44.789 --> 00:58:44.799 align:start position:0% +architectures on classifications on + + +00:58:44.799 --> 00:58:45.910 align:start position:0% +architectures on classifications on +imagenet + +00:58:45.910 --> 00:58:45.920 align:start position:0% +imagenet + + +00:58:45.920 --> 00:58:49.030 align:start position:0% +imagenet +we<00:58:46.160> have<00:58:47.359> i<00:58:47.680> believe<00:58:48.160> five<00:58:48.480> layers<00:58:48.880> of + +00:58:49.030 --> 00:58:49.040 align:start position:0% +we have i believe five layers of + + +00:58:49.040 --> 00:58:51.829 align:start position:0% +we have i believe five layers of +convolutions + +00:58:51.829 --> 00:58:51.839 align:start position:0% +convolutions + + +00:58:51.839 --> 00:58:55.670 align:start position:0% +convolutions +first<00:58:54.720> and<00:58:55.040> then + +00:58:55.670 --> 00:58:55.680 align:start position:0% +first and then + + +00:58:55.680 --> 00:58:58.870 align:start position:0% +first and then +so<00:58:55.839> these<00:58:56.079> are<00:58:56.400> convolutions + +00:58:58.870 --> 00:58:58.880 align:start position:0% +so these are convolutions + + +00:58:58.880 --> 00:59:02.390 align:start position:0% +so these are convolutions +calm<00:58:59.280> layers<00:59:00.400> and<00:59:00.559> then<00:59:00.960> there<00:59:01.280> are<00:59:01.599> a<00:59:02.079> couple + +00:59:02.390 --> 00:59:02.400 align:start position:0% +calm layers and then there are a couple + + +00:59:02.400 --> 00:59:03.190 align:start position:0% +calm layers and then there are a couple +of + +00:59:03.190 --> 00:59:03.200 align:start position:0% +of + + +00:59:03.200 --> 00:59:09.030 align:start position:0% +of +fully<00:59:03.520> connected<00:59:04.079> layers<00:59:04.400> after<00:59:04.799> that + +00:59:09.030 --> 00:59:09.040 align:start position:0% + + + +00:59:09.040 --> 00:59:13.670 align:start position:0% + +fully<00:59:10.400> connected<00:59:10.960> layers + +00:59:13.670 --> 00:59:13.680 align:start position:0% + + + +00:59:13.680 --> 00:59:16.829 align:start position:0% + +so<00:59:13.920> convolutions<00:59:14.559> are<00:59:14.720> extremely<00:59:15.200> extremely + +00:59:16.829 --> 00:59:16.839 align:start position:0% +so convolutions are extremely extremely + + +00:59:16.839 --> 00:59:18.150 align:start position:0% +so convolutions are extremely extremely +important + +00:59:18.150 --> 00:59:18.160 align:start position:0% +important + + +00:59:18.160 --> 00:59:23.349 align:start position:0% +important +okay<00:59:18.480> so<00:59:19.040> um<00:59:20.880> there<00:59:21.040> are<00:59:21.119> some<00:59:21.440> questions<00:59:22.160> um + +00:59:23.349 --> 00:59:23.359 align:start position:0% +okay so um there are some questions um + + +00:59:23.359 --> 00:59:26.470 align:start position:0% +okay so um there are some questions um +is<00:59:23.760> pooling<00:59:24.319> uh<00:59:25.359> does<00:59:25.599> pooling<00:59:26.160> add + +00:59:26.470 --> 00:59:26.480 align:start position:0% +is pooling uh does pooling add + + +00:59:26.480 --> 00:59:28.630 align:start position:0% +is pooling uh does pooling add +non-linearity<00:59:27.280> to<00:59:27.440> the<00:59:27.599> system + +00:59:28.630 --> 00:59:28.640 align:start position:0% +non-linearity to the system + + +00:59:28.640 --> 00:59:30.630 align:start position:0% +non-linearity to the system +yes<00:59:29.040> for<00:59:29.280> instance<00:59:29.599> if<00:59:29.680> you<00:59:29.839> look<00:59:30.000> at<00:59:30.079> the<00:59:30.240> max + +00:59:30.630 --> 00:59:30.640 align:start position:0% +yes for instance if you look at the max + + +00:59:30.640 --> 00:59:33.910 align:start position:0% +yes for instance if you look at the max +pooling<00:59:31.520> the<00:59:31.839> relationship<00:59:32.559> is<00:59:32.799> a<00:59:32.880> non-linear + +00:59:33.910 --> 00:59:33.920 align:start position:0% +pooling the relationship is a non-linear + + +00:59:33.920 --> 00:59:36.470 align:start position:0% +pooling the relationship is a non-linear +from<00:59:34.079> the<00:59:34.319> input<00:59:34.559> to<00:59:34.720> the<00:59:34.880> output<00:59:36.000> another + +00:59:36.470 --> 00:59:36.480 align:start position:0% +from the input to the output another + + +00:59:36.480 --> 00:59:37.109 align:start position:0% +from the input to the output another +question + +00:59:37.109 --> 00:59:37.119 align:start position:0% +question + + +00:59:37.119 --> 00:59:39.190 align:start position:0% +question +is<00:59:37.280> this<00:59:37.599> truly<00:59:38.000> related<00:59:38.559> to<00:59:38.880> mass + +00:59:39.190 --> 00:59:39.200 align:start position:0% +is this truly related to mass + + +00:59:39.200 --> 00:59:40.230 align:start position:0% +is this truly related to mass +convolution + +00:59:40.230 --> 00:59:40.240 align:start position:0% +convolution + + +00:59:40.240 --> 00:59:43.430 align:start position:0% +convolution +uh<00:59:40.640> yes<00:59:40.960> and<00:59:41.200> no<00:59:42.720> the + +00:59:43.430 --> 00:59:43.440 align:start position:0% +uh yes and no the + + +00:59:43.440 --> 00:59:46.230 align:start position:0% +uh yes and no the +math<00:59:43.760> convolution<00:59:44.240> there<00:59:44.400> is<00:59:44.480> a<00:59:44.839> transpose + +00:59:46.230 --> 00:59:46.240 align:start position:0% +math convolution there is a transpose + + +00:59:46.240 --> 00:59:47.589 align:start position:0% +math convolution there is a transpose +but + +00:59:47.589 --> 00:59:47.599 align:start position:0% +but + + +00:59:47.599 --> 00:59:50.309 align:start position:0% +but +here<00:59:47.839> we<00:59:48.000> don't<00:59:48.160> do<00:59:48.319> the<00:59:48.559> transfer<00:59:48.960> so<00:59:49.119> in<00:59:49.200> the + +00:59:50.309 --> 00:59:50.319 align:start position:0% +here we don't do the transfer so in the + + +00:59:50.319 --> 00:59:51.349 align:start position:0% +here we don't do the transfer so in the +convolution + +00:59:51.349 --> 00:59:51.359 align:start position:0% +convolution + + +00:59:51.359 --> 00:59:54.710 align:start position:0% +convolution +the<00:59:52.000> correct<00:59:52.400> convolution<00:59:53.040> operator<00:59:54.160> so<00:59:54.400> you + +00:59:54.710 --> 00:59:54.720 align:start position:0% +the correct convolution operator so you + + +00:59:54.720 --> 00:59:55.829 align:start position:0% +the correct convolution operator so you +if<00:59:54.799> you<00:59:54.960> want<00:59:55.119> to<00:59:55.359> you<00:59:55.440> know<00:59:55.599> do<00:59:55.760> the + +00:59:55.829 --> 00:59:55.839 align:start position:0% +if you want to you know do the + + +00:59:55.839 --> 00:59:57.270 align:start position:0% +if you want to you know do the +convolution<00:59:56.319> of<00:59:56.400> this<00:59:56.559> part<00:59:56.799> and<00:59:56.880> this<00:59:57.040> part + +00:59:57.270 --> 00:59:57.280 align:start position:0% +convolution of this part and this part + + +00:59:57.280 --> 00:59:58.789 align:start position:0% +convolution of this part and this part +you've<00:59:57.440> got<00:59:57.599> to<00:59:57.760> transpose<00:59:58.319> this + +00:59:58.789 --> 00:59:58.799 align:start position:0% +you've got to transpose this + + +00:59:58.799 --> 01:00:01.349 align:start position:0% +you've got to transpose this +but<00:59:58.960> then<00:59:59.680> consider<01:00:00.160> the<01:00:00.400> transpose<01:00:00.880> value<01:00:01.280> as + +01:00:01.349 --> 01:00:01.359 align:start position:0% +but then consider the transpose value as + + +01:00:01.359 --> 01:00:02.549 align:start position:0% +but then consider the transpose value as +your<01:00:01.599> filter + +01:00:02.549 --> 01:00:02.559 align:start position:0% +your filter + + +01:00:02.559 --> 01:00:05.829 align:start position:0% +your filter +uh<01:00:03.200> and<01:00:04.240> you<01:00:04.319> know<01:00:04.559> in<01:00:04.640> that<01:00:04.960> case<01:00:05.280> that<01:00:05.440> would + +01:00:05.829 --> 01:00:05.839 align:start position:0% +uh and you know in that case that would + + +01:00:05.839 --> 01:00:07.589 align:start position:0% +uh and you know in that case that would +you<01:00:05.920> know<01:00:06.160> give<01:00:06.400> you<01:00:06.480> the<01:00:06.640> same<01:00:06.880> result<01:00:07.440> so + +01:00:07.589 --> 01:00:07.599 align:start position:0% +you know give you the same result so + + +01:00:07.599 --> 01:00:10.150 align:start position:0% +you know give you the same result so +this<01:00:07.839> is<01:00:08.000> in<01:00:08.160> fact<01:00:08.319> called<01:00:08.640> cross<01:00:08.960> correlation + +01:00:10.150 --> 01:00:10.160 align:start position:0% +this is in fact called cross correlation + + +01:00:10.160 --> 01:00:12.789 align:start position:0% +this is in fact called cross correlation +as<01:00:10.400> one<01:00:10.640> of<01:00:10.720> you<01:00:10.960> mentioned<01:00:11.760> but<01:00:11.920> it<01:00:12.160> is<01:00:12.480> very + +01:00:12.789 --> 01:00:12.799 align:start position:0% +as one of you mentioned but it is very + + +01:00:12.799 --> 01:00:15.109 align:start position:0% +as one of you mentioned but it is very +very<01:00:13.040> related<01:00:13.520> to<01:00:13.920> the<01:00:14.079> convolution<01:00:14.640> operator + +01:00:15.109 --> 01:00:15.119 align:start position:0% +very related to the convolution operator + + +01:00:15.119 --> 01:00:15.829 align:start position:0% +very related to the convolution operator +so<01:00:15.280> that's<01:00:15.440> well + +01:00:15.829 --> 01:00:15.839 align:start position:0% +so that's well + + +01:00:15.839 --> 01:00:17.349 align:start position:0% +so that's well +that's<01:00:16.079> why<01:00:16.240> these<01:00:16.559> these<01:00:16.799> are<01:00:16.880> called + +01:00:17.349 --> 01:00:17.359 align:start position:0% +that's why these these are called + + +01:00:17.359 --> 01:00:21.030 align:start position:0% +that's why these these are called +convolutional<01:00:18.000> layers + +01:00:21.030 --> 01:00:21.040 align:start position:0% + + + +01:00:21.040 --> 01:00:24.950 align:start position:0% + +okay<01:00:21.520> um<01:00:22.960> perfect + +01:00:24.950 --> 01:00:24.960 align:start position:0% +okay um perfect + + +01:00:24.960 --> 01:00:30.150 align:start position:0% +okay um perfect +uh<01:00:25.440> all<01:00:25.520> right<01:00:25.760> so<01:00:26.000> that's<01:00:26.400> it<01:00:26.640> for<01:00:27.599> the + +01:00:30.150 --> 01:00:30.160 align:start position:0% + + + +01:00:30.160 --> 01:00:37.589 align:start position:0% + +4d<01:00:30.640> models<01:00:32.240> okay<01:00:32.480> so<01:00:32.720> let's<01:00:32.960> move<01:00:33.200> on + +01:00:37.589 --> 01:00:37.599 align:start position:0% + + + +01:00:37.599 --> 01:00:44.390 align:start position:0% + +again<01:00:38.319> let's<01:00:38.720> uh<01:00:39.760> recall<01:00:40.240> what<01:00:40.480> was<01:00:40.720> my<01:00:41.040> erm + +01:00:44.390 --> 01:00:44.400 align:start position:0% + + + +01:00:44.400 --> 01:00:47.990 align:start position:0% + +i'm<01:00:44.640> minimizing<01:00:45.520> over<01:00:45.839> w<01:00:47.200> sum + +01:00:47.990 --> 01:00:48.000 align:start position:0% +i'm minimizing over w sum + + +01:00:48.000 --> 01:00:53.030 align:start position:0% +i'm minimizing over w sum +of<01:00:49.040> my<01:00:49.440> losses<01:00:50.559> between<01:00:51.040> f<01:00:51.280> w<01:00:51.680> of<01:00:51.920> x<01:00:52.160> i + +01:00:53.030 --> 01:00:53.040 align:start position:0% +of my losses between f w of x i + + +01:00:53.040 --> 01:00:57.030 align:start position:0% +of my losses between f w of x i +and<01:00:53.280> y<01:00:53.599> i<01:00:54.799> okay<01:00:55.200> so<01:00:56.000> uh + +01:00:57.030 --> 01:00:57.040 align:start position:0% +and y i okay so uh + + +01:00:57.040 --> 01:01:00.150 align:start position:0% +and y i okay so uh +so<01:00:57.280> far<01:00:57.599> we<01:00:58.000> know<01:00:58.319> what<01:00:58.640> loss<01:00:58.960> functions<01:00:59.440> are + +01:01:00.150 --> 01:01:00.160 align:start position:0% +so far we know what loss functions are + + +01:01:00.160 --> 01:01:03.510 align:start position:0% +so far we know what loss functions are +we<01:01:00.400> know<01:01:00.880> what<01:01:01.119> is<01:01:01.280> fw<01:01:02.400> uh<01:01:02.799> mlp + +01:01:03.510 --> 01:01:03.520 align:start position:0% +we know what is fw uh mlp + + +01:01:03.520 --> 01:01:05.510 align:start position:0% +we know what is fw uh mlp +convolution<01:01:04.079> or<01:01:04.240> maybe<01:01:04.559> other<01:01:04.880> types<01:01:05.200> of<01:01:05.440> you + +01:01:05.510 --> 01:01:05.520 align:start position:0% +convolution or maybe other types of you + + +01:01:05.520 --> 01:01:07.030 align:start position:0% +convolution or maybe other types of you +know<01:01:05.920> functions + +01:01:07.030 --> 01:01:07.040 align:start position:0% +know functions + + +01:01:07.040 --> 01:01:09.750 align:start position:0% +know functions +and<01:01:07.280> now<01:01:07.520> we<01:01:07.680> want<01:01:07.839> to<01:01:07.920> see<01:01:08.240> how<01:01:08.480> we<01:01:08.880> can<01:01:09.440> solve + +01:01:09.750 --> 01:01:09.760 align:start position:0% +and now we want to see how we can solve + + +01:01:09.760 --> 01:01:11.510 align:start position:0% +and now we want to see how we can solve +this<01:01:10.000> optimization + +01:01:11.510 --> 01:01:11.520 align:start position:0% +this optimization + + +01:01:11.520 --> 01:01:14.870 align:start position:0% +this optimization +problem<01:01:12.960> if<01:01:13.520> if<01:01:13.920> w + +01:01:14.870 --> 01:01:14.880 align:start position:0% +problem if if w + + +01:01:14.880 --> 01:01:19.190 align:start position:0% +problem if if w +basically<01:01:15.280> the<01:01:15.760> basic<01:01:16.160> question<01:01:16.640> here<01:01:16.960> is + +01:01:19.190 --> 01:01:19.200 align:start position:0% +basically the basic question here is + + +01:01:19.200 --> 01:01:23.510 align:start position:0% +basically the basic question here is +how<01:01:19.520> to<01:01:21.280> solve + +01:01:23.510 --> 01:01:23.520 align:start position:0% +how to solve + + +01:01:23.520 --> 01:01:29.030 align:start position:0% +how to solve +the<01:01:23.839> optimization<01:01:24.559> problem + +01:01:29.030 --> 01:01:29.040 align:start position:0% + + + +01:01:29.040 --> 01:01:33.190 align:start position:0% + +okay<01:01:29.359> so<01:01:29.920> if<01:01:31.839> my<01:01:32.240> objective + +01:01:33.190 --> 01:01:33.200 align:start position:0% +okay so if my objective + + +01:01:33.200 --> 01:01:37.430 align:start position:0% +okay so if my objective +is<01:01:33.359> convex<01:01:33.920> let's<01:01:34.160> call<01:01:34.400> the<01:01:34.559> whole<01:01:34.799> thing + +01:01:37.430 --> 01:01:37.440 align:start position:0% + + + +01:01:37.440 --> 01:01:41.270 align:start position:0% + +l<01:01:37.760> of<01:01:37.920> w<01:01:38.400> so<01:01:38.559> this<01:01:38.720> is<01:01:38.880> my<01:01:39.200> objective<01:01:39.760> function + +01:01:41.270 --> 01:01:41.280 align:start position:0% +l of w so this is my objective function + + +01:01:41.280 --> 01:01:47.349 align:start position:0% +l of w so this is my objective function +if<01:01:42.559> l<01:01:43.040> of<01:01:43.280> w<01:01:44.000> is<01:01:44.160> convex + +01:01:47.349 --> 01:01:47.359 align:start position:0% + + + +01:01:47.359 --> 01:01:50.150 align:start position:0% + +uh<01:01:47.680> that<01:01:47.920> happens<01:01:48.799> because<01:01:49.119> my<01:01:49.359> loss<01:01:49.920> is + +01:01:50.150 --> 01:01:50.160 align:start position:0% +uh that happens because my loss is + + +01:01:50.160 --> 01:01:51.589 align:start position:0% +uh that happens because my loss is +convex<01:01:50.640> and<01:01:50.799> if<01:01:50.960> my<01:01:51.200> f + +01:01:51.589 --> 01:01:51.599 align:start position:0% +convex and if my f + + +01:01:51.599 --> 01:01:54.150 align:start position:0% +convex and if my f +of<01:01:51.839> w<01:01:52.480> is<01:01:52.720> an<01:01:52.880> affine<01:01:53.359> function<01:01:53.760> i'll<01:01:53.920> get<01:01:54.079> the + +01:01:54.150 --> 01:01:54.160 align:start position:0% +of w is an affine function i'll get the + + +01:01:54.160 --> 01:01:56.069 align:start position:0% +of w is an affine function i'll get the +convex<01:01:54.559> optimization<01:01:55.200> at<01:01:55.280> the<01:01:55.440> end + +01:01:56.069 --> 01:01:56.079 align:start position:0% +convex optimization at the end + + +01:01:56.079 --> 01:01:59.109 align:start position:0% +convex optimization at the end +i'm<01:01:56.319> very<01:01:56.640> happy<01:01:57.119> because<01:01:58.319> i<01:01:58.720> know + +01:01:59.109 --> 01:01:59.119 align:start position:0% +i'm very happy because i know + + +01:01:59.119 --> 01:02:09.349 align:start position:0% +i'm very happy because i know +gradient<01:01:59.599> descent<01:02:00.480> or<01:02:04.839> stochastic + +01:02:09.349 --> 01:02:09.359 align:start position:0% + + + +01:02:09.359 --> 01:02:12.390 align:start position:0% + +so<01:02:09.760> what<01:02:10.000> is<01:02:10.160> these<01:02:10.400> notifications + +01:02:12.390 --> 01:02:12.400 align:start position:0% +so what is these notifications + + +01:02:12.400 --> 01:02:15.670 align:start position:0% +so what is these notifications +or<01:02:13.680> the<01:02:13.839> stochastic<01:02:14.400> gradient<01:02:14.880> descent + +01:02:15.670 --> 01:02:15.680 align:start position:0% +or the stochastic gradient descent + + +01:02:15.680 --> 01:02:20.069 align:start position:0% +or the stochastic gradient descent +can<01:02:17.119> be<01:02:17.359> used<01:02:18.559> gradient<01:02:19.039> descent + +01:02:20.069 --> 01:02:20.079 align:start position:0% +can be used gradient descent + + +01:02:20.079 --> 01:02:23.109 align:start position:0% +can be used gradient descent +can<01:02:20.319> be<01:02:20.559> used<01:02:22.720> to + +01:02:23.109 --> 01:02:23.119 align:start position:0% +can be used to + + +01:02:23.119 --> 01:02:26.630 align:start position:0% +can be used to +find<01:02:24.160> w<01:02:24.640> star<01:02:24.960> which<01:02:25.200> is<01:02:25.359> my<01:02:25.520> global<01:02:26.079> optimizer + +01:02:26.630 --> 01:02:26.640 align:start position:0% +find w star which is my global optimizer + + +01:02:26.640 --> 01:02:27.670 align:start position:0% +find w star which is my global optimizer +so<01:02:26.880> i'll<01:02:27.039> have + +01:02:27.670 --> 01:02:27.680 align:start position:0% +so i'll have + + +01:02:27.680 --> 01:02:30.470 align:start position:0% +so i'll have +l<01:02:28.000> of<01:02:28.160> w<01:02:28.559> in<01:02:28.640> that<01:02:28.880> case<01:02:29.680> something<01:02:30.079> like<01:02:30.240> this + +01:02:30.470 --> 01:02:30.480 align:start position:0% +l of w in that case something like this + + +01:02:30.480 --> 01:02:31.829 align:start position:0% +l of w in that case something like this +it<01:02:30.640> is<01:02:30.799> convex + +01:02:31.829 --> 01:02:31.839 align:start position:0% +it is convex + + +01:02:31.839 --> 01:02:35.109 align:start position:0% +it is convex +so<01:02:32.000> this<01:02:32.240> is<01:02:32.400> my<01:02:32.640> w<01:02:33.039> this<01:02:33.280> is<01:02:33.359> my<01:02:33.520> w<01:02:33.920> star + +01:02:35.109 --> 01:02:35.119 align:start position:0% +so this is my w this is my w star + + +01:02:35.119 --> 01:02:37.910 align:start position:0% +so this is my w this is my w star +so<01:02:36.079> the<01:02:36.240> way<01:02:36.480> that<01:02:36.720> we<01:02:37.119> will<01:02:37.359> solve<01:02:37.599> this + +01:02:37.910 --> 01:02:37.920 align:start position:0% +so the way that we will solve this + + +01:02:37.920 --> 01:02:39.349 align:start position:0% +so the way that we will solve this +problem<01:02:38.319> we'll<01:02:38.640> start<01:02:38.880> with<01:02:39.039> some + +01:02:39.349 --> 01:02:39.359 align:start position:0% +problem we'll start with some + + +01:02:39.359 --> 01:02:42.549 align:start position:0% +problem we'll start with some +initial<01:02:40.160> value<01:02:40.480> for<01:02:40.640> w<01:02:41.039> let's<01:02:41.200> say<01:02:41.440> w0<01:02:42.319> and<01:02:42.480> i'm + +01:02:42.549 --> 01:02:42.559 align:start position:0% +initial value for w let's say w0 and i'm + + +01:02:42.559 --> 01:02:44.710 align:start position:0% +initial value for w let's say w0 and i'm +going<01:02:42.720> to<01:02:42.880> move<01:02:43.200> in<01:02:43.359> the<01:02:43.599> opposite<01:02:44.000> direction + +01:02:44.710 --> 01:02:44.720 align:start position:0% +going to move in the opposite direction + + +01:02:44.720 --> 01:02:46.829 align:start position:0% +going to move in the opposite direction +of<01:02:45.039> my + +01:02:46.829 --> 01:02:46.839 align:start position:0% +of my + + +01:02:46.839 --> 01:02:50.309 align:start position:0% +of my +gradient<01:02:47.599> and<01:02:48.160> if<01:02:48.640> my<01:02:48.960> learning<01:02:49.440> rate + +01:02:50.309 --> 01:02:50.319 align:start position:0% +gradient and if my learning rate + + +01:02:50.319 --> 01:02:52.549 align:start position:0% +gradient and if my learning rate +my<01:02:50.480> step<01:02:50.799> sizes<01:02:51.280> are<01:02:51.599> properly<01:02:52.079> said<01:02:52.400> i'm + +01:02:52.549 --> 01:02:52.559 align:start position:0% +my step sizes are properly said i'm + + +01:02:52.559 --> 01:02:53.430 align:start position:0% +my step sizes are properly said i'm +going<01:02:52.640> to<01:02:52.799> convert + +01:02:53.430 --> 01:02:53.440 align:start position:0% +going to convert + + +01:02:53.440 --> 01:02:56.630 align:start position:0% +going to convert +to<01:02:54.480> w<01:02:54.880> star<01:02:55.200> at<01:02:55.280> the<01:02:55.440> end<01:02:55.599> of<01:02:55.680> the<01:02:55.839> day + +01:02:56.630 --> 01:02:56.640 align:start position:0% +to w star at the end of the day + + +01:02:56.640 --> 01:03:06.470 align:start position:0% +to w star at the end of the day +so<01:02:56.880> that's<01:02:57.200> the<01:02:58.079> gradient<01:02:58.480> descent<01:03:00.839> algorithm + +01:03:06.470 --> 01:03:06.480 align:start position:0% + + + +01:03:06.480 --> 01:03:09.910 align:start position:0% + +so<01:03:07.039> basically<01:03:07.680> we<01:03:08.160> start<01:03:08.880> with<01:03:09.119> some + +01:03:09.910 --> 01:03:09.920 align:start position:0% +so basically we start with some + + +01:03:09.920 --> 01:03:12.950 align:start position:0% +so basically we start with some +initialization<01:03:10.880> for<01:03:11.200> my<01:03:11.839> weights<01:03:12.640> random + +01:03:12.950 --> 01:03:12.960 align:start position:0% +initialization for my weights random + + +01:03:12.960 --> 01:03:14.470 align:start position:0% +initialization for my weights random +initializations<01:03:13.760> you<01:03:13.839> can<01:03:14.000> use<01:03:14.319> like + +01:03:14.470 --> 01:03:14.480 align:start position:0% +initializations you can use like + + +01:03:14.480 --> 01:03:15.670 align:start position:0% +initializations you can use like +gaussians + +01:03:15.670 --> 01:03:15.680 align:start position:0% +gaussians + + +01:03:15.680 --> 01:03:18.789 align:start position:0% +gaussians +and<01:03:15.839> then<01:03:16.640> weights<01:03:17.039> at<01:03:17.359> iteration<01:03:18.000> t<01:03:18.240> plus<01:03:18.480> 1 + +01:03:18.789 --> 01:03:18.799 align:start position:0% +and then weights at iteration t plus 1 + + +01:03:18.799 --> 01:03:19.910 align:start position:0% +and then weights at iteration t plus 1 +will<01:03:19.039> be + +01:03:19.910 --> 01:03:19.920 align:start position:0% +will be + + +01:03:19.920 --> 01:03:23.430 align:start position:0% +will be +my<01:03:20.160> base<01:03:20.400> at<01:03:20.559> iteration<01:03:21.119> t<01:03:21.760> minus<01:03:22.720> eta<01:03:23.200> which + +01:03:23.430 --> 01:03:23.440 align:start position:0% +my base at iteration t minus eta which + + +01:03:23.440 --> 01:03:24.789 align:start position:0% +my base at iteration t minus eta which +is<01:03:23.520> my<01:03:23.760> learning<01:03:24.160> rate + +01:03:24.789 --> 01:03:24.799 align:start position:0% +is my learning rate + + +01:03:24.799 --> 01:03:27.990 align:start position:0% +is my learning rate +times<01:03:25.200> the<01:03:25.359> gradient<01:03:26.000> of<01:03:26.160> my<01:03:26.480> loss<01:03:27.200> at<01:03:27.440> the + +01:03:27.990 --> 01:03:28.000 align:start position:0% +times the gradient of my loss at the + + +01:03:28.000 --> 01:03:31.270 align:start position:0% +times the gradient of my loss at the +previous<01:03:28.480> iteration<01:03:29.520> so<01:03:29.920> this<01:03:30.559> data<01:03:31.039> called + +01:03:31.270 --> 01:03:31.280 align:start position:0% +previous iteration so this data called + + +01:03:31.280 --> 01:03:34.390 align:start position:0% +previous iteration so this data called +learning<01:03:31.760> rate + +01:03:34.390 --> 01:03:34.400 align:start position:0% + + + +01:03:34.400 --> 01:03:36.549 align:start position:0% + +and<01:03:34.640> here<01:03:34.960> we<01:03:35.119> are<01:03:35.280> using<01:03:35.599> the<01:03:35.760> fixed<01:03:36.160> learning + +01:03:36.549 --> 01:03:36.559 align:start position:0% +and here we are using the fixed learning + + +01:03:36.559 --> 01:03:37.990 align:start position:0% +and here we are using the fixed learning +right + +01:03:37.990 --> 01:03:38.000 align:start position:0% +right + + +01:03:38.000 --> 01:03:41.829 align:start position:0% +right +okay<01:03:38.480> so<01:03:39.599> if<01:03:39.839> you<01:03:40.160> think<01:03:40.400> about<01:03:40.799> this + +01:03:41.829 --> 01:03:41.839 align:start position:0% +okay so if you think about this + + +01:03:41.839 --> 01:03:45.510 align:start position:0% +okay so if you think about this +gradient<01:03:42.880> of<01:03:43.280> the<01:03:43.680> loss<01:03:44.720> my<01:03:45.039> loss + +01:03:45.510 --> 01:03:45.520 align:start position:0% +gradient of the loss my loss + + +01:03:45.520 --> 01:03:48.069 align:start position:0% +gradient of the loss my loss +is<01:03:45.920> sum<01:03:46.480> of<01:03:46.640> the<01:03:46.880> losses<01:03:47.280> of<01:03:47.440> individual + +01:03:48.069 --> 01:03:48.079 align:start position:0% +is sum of the losses of individual + + +01:03:48.079 --> 01:03:49.349 align:start position:0% +is sum of the losses of individual +samples + +01:03:49.349 --> 01:03:49.359 align:start position:0% +samples + + +01:03:49.359 --> 01:03:51.990 align:start position:0% +samples +in<01:03:49.440> that<01:03:49.760> case<01:03:50.240> my<01:03:50.480> gradient<01:03:51.359> is<01:03:51.520> going<01:03:51.680> to<01:03:51.839> be + +01:03:51.990 --> 01:03:52.000 align:start position:0% +in that case my gradient is going to be + + +01:03:52.000 --> 01:03:53.510 align:start position:0% +in that case my gradient is going to be +the<01:03:52.240> sum + +01:03:53.510 --> 01:03:53.520 align:start position:0% +the sum + + +01:03:53.520 --> 01:03:57.589 align:start position:0% +the sum +of<01:03:53.760> the<01:03:53.920> gradients<01:03:54.799> that<01:03:55.039> i'll<01:03:55.599> observe + +01:03:57.589 --> 01:03:57.599 align:start position:0% +of the gradients that i'll observe + + +01:03:57.599 --> 01:04:00.789 align:start position:0% +of the gradients that i'll observe +over<01:03:58.000> individual<01:03:58.640> samples<01:03:59.119> so<01:03:59.440> i<01:03:59.680> have<01:04:00.160> 1<01:04:00.400> over + +01:04:00.789 --> 01:04:00.799 align:start position:0% +over individual samples so i have 1 over + + +01:04:00.799 --> 01:04:04.470 align:start position:0% +over individual samples so i have 1 over +n<01:04:01.520> the<01:04:01.760> gradient<01:04:02.400> of + +01:04:04.470 --> 01:04:04.480 align:start position:0% +n the gradient of + + +01:04:04.480 --> 01:04:08.309 align:start position:0% +n the gradient of +loss<01:04:05.520> on<01:04:05.920> individual<01:04:06.559> samples + +01:04:08.309 --> 01:04:08.319 align:start position:0% +loss on individual samples + + +01:04:08.319 --> 01:04:10.230 align:start position:0% +loss on individual samples +and<01:04:08.480> this<01:04:08.720> computation<01:04:09.440> can<01:04:09.680> be<01:04:09.920> very + +01:04:10.230 --> 01:04:10.240 align:start position:0% +and this computation can be very + + +01:04:10.240 --> 01:04:11.670 align:start position:0% +and this computation can be very +expensive<01:04:10.799> because + +01:04:11.670 --> 01:04:11.680 align:start position:0% +expensive because + + +01:04:11.680 --> 01:04:13.670 align:start position:0% +expensive because +we<01:04:11.839> need<01:04:12.079> to<01:04:12.240> compute<01:04:12.720> gradient<01:04:13.200> at<01:04:13.359> each<01:04:13.599> of + +01:04:13.670 --> 01:04:13.680 align:start position:0% +we need to compute gradient at each of + + +01:04:13.680 --> 01:04:14.710 align:start position:0% +we need to compute gradient at each of +the<01:04:13.839> samples + +01:04:14.710 --> 01:04:14.720 align:start position:0% +the samples + + +01:04:14.720 --> 01:04:17.589 align:start position:0% +the samples +for<01:04:15.039> n<01:04:15.280> samples<01:04:16.000> in<01:04:16.160> each<01:04:16.400> iteration<01:04:17.200> so<01:04:17.359> it<01:04:17.520> is + +01:04:17.589 --> 01:04:17.599 align:start position:0% +for n samples in each iteration so it is + + +01:04:17.599 --> 01:04:18.069 align:start position:0% +for n samples in each iteration so it is +a + +01:04:18.069 --> 01:04:18.079 align:start position:0% +a + + +01:04:18.079 --> 01:04:22.390 align:start position:0% +a +it's<01:04:18.319> an<01:04:18.480> expensive<01:04:19.599> uh<01:04:20.079> step<01:04:21.039> so<01:04:21.200> in<01:04:21.359> order<01:04:21.680> to + +01:04:22.390 --> 01:04:22.400 align:start position:0% +it's an expensive uh step so in order to + + +01:04:22.400 --> 01:04:25.190 align:start position:0% +it's an expensive uh step so in order to +reduce<01:04:22.799> the<01:04:22.880> computational<01:04:23.680> cost<01:04:24.000> we<01:04:24.160> can + +01:04:25.190 --> 01:04:25.200 align:start position:0% +reduce the computational cost we can + + +01:04:25.200 --> 01:04:25.589 align:start position:0% +reduce the computational cost we can +take + +01:04:25.589 --> 01:04:25.599 align:start position:0% +take + + +01:04:25.599 --> 01:04:28.630 align:start position:0% +take +samples<01:04:26.319> random<01:04:26.799> samples<01:04:27.760> from<01:04:28.000> our<01:04:28.240> training + +01:04:28.630 --> 01:04:28.640 align:start position:0% +samples random samples from our training + + +01:04:28.640 --> 01:04:30.950 align:start position:0% +samples random samples from our training +set<01:04:28.880> in<01:04:29.039> order<01:04:29.280> to<01:04:29.440> approximate<01:04:30.160> this + +01:04:30.950 --> 01:04:30.960 align:start position:0% +set in order to approximate this + + +01:04:30.960 --> 01:04:34.069 align:start position:0% +set in order to approximate this +gradient<01:04:31.920> so<01:04:32.079> that's<01:04:32.400> the<01:04:32.720> key<01:04:33.039> idea<01:04:33.599> of + +01:04:34.069 --> 01:04:34.079 align:start position:0% +gradient so that's the key idea of + + +01:04:34.079 --> 01:04:39.670 align:start position:0% +gradient so that's the key idea of +stochastic<01:04:34.640> gradient<01:04:35.119> descent + +01:04:39.670 --> 01:04:39.680 align:start position:0% + + + +01:04:39.680 --> 01:04:42.870 align:start position:0% + +so<01:04:39.839> in<01:04:39.920> the<01:04:40.160> stochastic<01:04:40.799> gradient<01:04:41.200> descent + +01:04:42.870 --> 01:04:42.880 align:start position:0% +so in the stochastic gradient descent + + +01:04:42.880 --> 01:04:45.589 align:start position:0% +so in the stochastic gradient descent +uh<01:04:43.119> it<01:04:43.280> is<01:04:43.440> very<01:04:43.680> similar<01:04:44.160> to<01:04:44.400> the<01:04:45.200> gradient + +01:04:45.589 --> 01:04:45.599 align:start position:0% +uh it is very similar to the gradient + + +01:04:45.599 --> 01:04:47.190 align:start position:0% +uh it is very similar to the gradient +descent<01:04:46.000> the<01:04:46.079> only<01:04:46.319> difference<01:04:46.720> here<01:04:46.880> is<01:04:47.039> that + +01:04:47.190 --> 01:04:47.200 align:start position:0% +descent the only difference here is that + + +01:04:47.200 --> 01:04:48.630 align:start position:0% +descent the only difference here is that +instead<01:04:47.440> of<01:04:47.599> looking<01:04:47.920> at<01:04:48.079> the + +01:04:48.630 --> 01:04:48.640 align:start position:0% +instead of looking at the + + +01:04:48.640 --> 01:04:51.109 align:start position:0% +instead of looking at the +uh<01:04:49.280> average<01:04:49.760> gradient<01:04:50.240> over<01:04:50.480> all<01:04:50.640> samples<01:04:50.960> we + +01:04:51.109 --> 01:04:51.119 align:start position:0% +uh average gradient over all samples we + + +01:04:51.119 --> 01:04:52.789 align:start position:0% +uh average gradient over all samples we +are<01:04:51.200> going<01:04:51.359> to<01:04:51.440> look<01:04:51.599> at<01:04:51.680> the<01:04:51.839> average<01:04:52.319> on<01:04:52.559> a + +01:04:52.789 --> 01:04:52.799 align:start position:0% +are going to look at the average on a + + +01:04:52.799 --> 01:04:54.150 align:start position:0% +are going to look at the average on a +mini<01:04:53.119> batch + +01:04:54.150 --> 01:04:54.160 align:start position:0% +mini batch + + +01:04:54.160 --> 01:04:57.750 align:start position:0% +mini batch +so<01:04:54.319> we<01:04:54.480> have<01:04:55.760> w<01:04:56.319> t<01:04:56.559> plus<01:04:56.799> one + +01:04:57.750 --> 01:04:57.760 align:start position:0% +so we have w t plus one + + +01:04:57.760 --> 01:05:00.870 align:start position:0% +so we have w t plus one +would<01:04:57.920> be<01:04:58.319> w<01:04:58.720> of<01:04:58.880> t<01:04:59.359> minus<01:04:59.760> eta + +01:05:00.870 --> 01:05:00.880 align:start position:0% +would be w of t minus eta + + +01:05:00.880 --> 01:05:03.750 align:start position:0% +would be w of t minus eta +let's<01:05:01.039> say<01:05:01.280> my<01:05:01.520> mini<01:05:01.839> batch<01:05:02.160> size<01:05:02.480> is<01:05:02.720> b<01:05:03.599> and + +01:05:03.750 --> 01:05:03.760 align:start position:0% +let's say my mini batch size is b and + + +01:05:03.760 --> 01:05:05.349 align:start position:0% +let's say my mini batch size is b and +i'm<01:05:03.839> going<01:05:04.000> to<01:05:04.079> look<01:05:04.400> at + +01:05:05.349 --> 01:05:05.359 align:start position:0% +i'm going to look at + + +01:05:05.359 --> 01:05:08.230 align:start position:0% +i'm going to look at +a<01:05:05.680> set<01:05:05.920> of<01:05:06.079> samples<01:05:06.640> in<01:05:06.720> this<01:05:06.960> mini<01:05:07.280> batch + +01:05:08.230 --> 01:05:08.240 align:start position:0% +a set of samples in this mini batch + + +01:05:08.240 --> 01:05:09.029 align:start position:0% +a set of samples in this mini batch +usually<01:05:08.720> much + +01:05:09.029 --> 01:05:09.039 align:start position:0% +usually much + + +01:05:09.039 --> 01:05:11.850 align:start position:0% +usually much +smaller<01:05:10.079> than + +01:05:11.850 --> 01:05:11.860 align:start position:0% +smaller than + + +01:05:11.860 --> 01:05:14.150 align:start position:0% +smaller than +[Music] + +01:05:14.150 --> 01:05:14.160 align:start position:0% +[Music] + + +01:05:14.160 --> 01:05:18.950 align:start position:0% +[Music] +what<01:05:15.839> we<01:05:16.000> have<01:05:16.240> in<01:05:16.400> our<01:05:16.640> training<01:05:16.960> set + +01:05:18.950 --> 01:05:18.960 align:start position:0% +what we have in our training set + + +01:05:18.960 --> 01:05:23.750 align:start position:0% +what we have in our training set +w<01:05:19.280> of<01:05:19.520> x<01:05:19.760> i<01:05:21.119> and<01:05:21.760> why<01:05:22.079> i + +01:05:23.750 --> 01:05:23.760 align:start position:0% +w of x i and why i + + +01:05:23.760 --> 01:05:27.029 align:start position:0% +w of x i and why i +so<01:05:24.000> this<01:05:24.240> is<01:05:24.400> called<01:05:24.880> my + +01:05:27.029 --> 01:05:27.039 align:start position:0% +so this is called my + + +01:05:27.039 --> 01:05:31.190 align:start position:0% +so this is called my +mini<01:05:29.039> batch + +01:05:31.190 --> 01:05:31.200 align:start position:0% +mini batch + + +01:05:31.200 --> 01:05:34.390 align:start position:0% +mini batch +okay<01:05:31.599> so<01:05:32.000> this<01:05:32.240> is<01:05:32.400> the<01:05:33.119> sgd + +01:05:34.390 --> 01:05:34.400 align:start position:0% +okay so this is the sgd + + +01:05:34.400 --> 01:05:37.829 align:start position:0% +okay so this is the sgd +and<01:05:35.039> in<01:05:35.200> fact<01:05:35.599> sgd<01:05:36.319> works<01:05:36.720> very<01:05:37.039> well<01:05:37.440> in + +01:05:37.829 --> 01:05:37.839 align:start position:0% +and in fact sgd works very well in + + +01:05:37.839 --> 01:05:41.270 align:start position:0% +and in fact sgd works very well in +uh<01:05:38.319> in<01:05:38.480> practice<01:05:39.119> and<01:05:39.280> it<01:05:39.440> is<01:05:40.079> um + +01:05:41.270 --> 01:05:41.280 align:start position:0% +uh in practice and it is um + + +01:05:41.280 --> 01:05:42.789 align:start position:0% +uh in practice and it is um +you<01:05:41.440> know<01:05:41.680> computationally<01:05:42.559> also + +01:05:42.789 --> 01:05:42.799 align:start position:0% +you know computationally also + + +01:05:42.799 --> 01:05:44.390 align:start position:0% +you know computationally also +inexpensive + +01:05:44.390 --> 01:05:44.400 align:start position:0% +inexpensive + + +01:05:44.400 --> 01:05:47.430 align:start position:0% +inexpensive +uh<01:05:44.960> for<01:05:45.200> the<01:05:46.000> convex<01:05:47.039> uh + +01:05:47.430 --> 01:05:47.440 align:start position:0% +uh for the convex uh + + +01:05:47.440 --> 01:05:49.829 align:start position:0% +uh for the convex uh +functions<01:05:48.160> there<01:05:48.400> are<01:05:48.640> some<01:05:49.280> proofs<01:05:49.680> that + +01:05:49.829 --> 01:05:49.839 align:start position:0% +functions there are some proofs that + + +01:05:49.839 --> 01:05:51.349 align:start position:0% +functions there are some proofs that +gradient<01:05:50.240> descent + +01:05:51.349 --> 01:05:51.359 align:start position:0% +gradient descent + + +01:05:51.359 --> 01:05:53.990 align:start position:0% +gradient descent +converges<01:05:51.920> with<01:05:52.160> an<01:05:52.319> exponential<01:05:53.039> rate<01:05:53.680> uh + +01:05:53.990 --> 01:05:54.000 align:start position:0% +converges with an exponential rate uh + + +01:05:54.000 --> 01:05:55.510 align:start position:0% +converges with an exponential rate uh +you<01:05:54.079> should<01:05:54.240> take<01:05:54.480> a<01:05:54.559> look<01:05:54.799> into<01:05:55.200> some<01:05:55.440> of + +01:05:55.510 --> 01:05:55.520 align:start position:0% +you should take a look into some of + + +01:05:55.520 --> 01:05:56.390 align:start position:0% +you should take a look into some of +those<01:05:55.839> uh + +01:05:56.390 --> 01:05:56.400 align:start position:0% +those uh + + +01:05:56.400 --> 01:05:57.910 align:start position:0% +those uh +some<01:05:56.640> of<01:05:56.720> those<01:05:56.960> proofs<01:05:57.280> i'm<01:05:57.359> gonna<01:05:57.599> review + +01:05:57.910 --> 01:05:57.920 align:start position:0% +some of those proofs i'm gonna review + + +01:05:57.920 --> 01:05:59.829 align:start position:0% +some of those proofs i'm gonna review +some<01:05:58.160> of<01:05:58.240> them<01:05:58.480> in<01:05:58.559> the<01:05:58.640> next<01:05:58.880> lecture + +01:05:59.829 --> 01:05:59.839 align:start position:0% +some of them in the next lecture + + +01:05:59.839 --> 01:06:03.190 align:start position:0% +some of them in the next lecture +uh<01:06:00.480> for<01:06:00.720> sgd<01:06:01.599> if<01:06:01.760> your<01:06:02.240> learning<01:06:02.720> rate<01:06:02.960> is + +01:06:03.190 --> 01:06:03.200 align:start position:0% +uh for sgd if your learning rate is + + +01:06:03.200 --> 01:06:03.829 align:start position:0% +uh for sgd if your learning rate is +adjusted + +01:06:03.829 --> 01:06:03.839 align:start position:0% +adjusted + + +01:06:03.839 --> 01:06:07.990 align:start position:0% +adjusted +properly<01:06:05.119> for<01:06:05.359> convex<01:06:05.839> optimization<01:06:06.559> it<01:06:06.799> also + +01:06:07.990 --> 01:06:08.000 align:start position:0% +properly for convex optimization it also + + +01:06:08.000 --> 01:06:12.390 align:start position:0% +properly for convex optimization it also +converges<01:06:08.640> to<01:06:08.880> your<01:06:09.119> global<01:06:10.319> optimizer + +01:06:12.390 --> 01:06:12.400 align:start position:0% +converges to your global optimizer + + +01:06:12.400 --> 01:06:16.950 align:start position:0% +converges to your global optimizer +okay<01:06:13.359> but<01:06:13.760> now<01:06:14.240> in<01:06:14.400> deep<01:06:14.720> learning + +01:06:16.950 --> 01:06:16.960 align:start position:0% +okay but now in deep learning + + +01:06:16.960 --> 01:06:18.710 align:start position:0% +okay but now in deep learning +again<01:06:17.200> there's<01:06:17.520> a<01:06:17.599> tradeoff<01:06:18.000> right<01:06:18.160> so<01:06:18.400> the + +01:06:18.710 --> 01:06:18.720 align:start position:0% +again there's a tradeoff right so the + + +01:06:18.720 --> 01:06:20.710 align:start position:0% +again there's a tradeoff right so the +linear<01:06:19.119> function<01:06:19.520> was<01:06:19.839> good + +01:06:20.710 --> 01:06:20.720 align:start position:0% +linear function was good + + +01:06:20.720 --> 01:06:24.549 align:start position:0% +linear function was good +but<01:06:21.520> uh<01:06:22.000> we<01:06:22.640> have<01:06:22.960> a<01:06:23.200> limited<01:06:23.760> representation + +01:06:24.549 --> 01:06:24.559 align:start position:0% +but uh we have a limited representation + + +01:06:24.559 --> 01:06:25.190 align:start position:0% +but uh we have a limited representation +power + +01:06:25.190 --> 01:06:25.200 align:start position:0% +power + + +01:06:25.200 --> 01:06:26.710 align:start position:0% +power +but<01:06:25.359> in<01:06:25.520> deep<01:06:25.760> learning<01:06:26.160> we<01:06:26.319> have<01:06:26.559> a + +01:06:26.710 --> 01:06:26.720 align:start position:0% +but in deep learning we have a + + +01:06:26.720 --> 01:06:30.150 align:start position:0% +but in deep learning we have a +non-linear<01:06:27.440> function + +01:06:30.150 --> 01:06:30.160 align:start position:0% + + + +01:06:30.160 --> 01:06:33.990 align:start position:0% + +uh<01:06:30.720> and<01:06:31.119> it<01:06:31.440> may<01:06:31.760> we<01:06:31.920> may<01:06:32.160> end<01:06:32.480> up<01:06:32.799> getting + +01:06:33.990 --> 01:06:34.000 align:start position:0% +uh and it may we may end up getting + + +01:06:34.000 --> 01:06:36.630 align:start position:0% +uh and it may we may end up getting +a<01:06:34.160> non-convex<01:06:35.359> optimization<01:06:36.079> problem<01:06:36.400> at<01:06:36.559> the + +01:06:36.630 --> 01:06:36.640 align:start position:0% +a non-convex optimization problem at the + + +01:06:36.640 --> 01:06:37.750 align:start position:0% +a non-convex optimization problem at the +end + +01:06:37.750 --> 01:06:37.760 align:start position:0% +end + + +01:06:37.760 --> 01:06:41.750 align:start position:0% +end +so<01:06:39.799> non-convex + +01:06:41.750 --> 01:06:41.760 align:start position:0% +so non-convex + + +01:06:41.760 --> 01:06:45.190 align:start position:0% +so non-convex +optimization<01:06:42.559> problem<01:06:43.200> so<01:06:43.520> how<01:06:44.480> can<01:06:44.720> we<01:06:44.960> deal + +01:06:45.190 --> 01:06:45.200 align:start position:0% +optimization problem so how can we deal + + +01:06:45.200 --> 01:06:46.870 align:start position:0% +optimization problem so how can we deal +with<01:06:45.359> this<01:06:45.599> problem + +01:06:46.870 --> 01:06:46.880 align:start position:0% +with this problem + + +01:06:46.880 --> 01:06:50.150 align:start position:0% +with this problem +so<01:06:47.760> even<01:06:48.160> in<01:06:48.240> that<01:06:48.559> case<01:06:48.960> we<01:06:49.119> are<01:06:49.359> using<01:06:49.920> the + +01:06:50.150 --> 01:06:50.160 align:start position:0% +so even in that case we are using the + + +01:06:50.160 --> 01:06:52.309 align:start position:0% +so even in that case we are using the +same<01:06:50.480> algorithms<01:06:51.200> we<01:06:51.280> are<01:06:51.440> using<01:06:51.760> gradient + +01:06:52.309 --> 01:06:52.319 align:start position:0% +same algorithms we are using gradient + + +01:06:52.319 --> 01:06:53.029 align:start position:0% +same algorithms we are using gradient +descent + +01:06:53.029 --> 01:06:53.039 align:start position:0% +descent + + +01:06:53.039 --> 01:06:55.910 align:start position:0% +descent +or<01:06:53.359> stochastic<01:06:53.920> gradient<01:06:54.480> descent<01:06:55.359> in<01:06:55.520> order + +01:06:55.910 --> 01:06:55.920 align:start position:0% +or stochastic gradient descent in order + + +01:06:55.920 --> 01:06:56.870 align:start position:0% +or stochastic gradient descent in order +to + +01:06:56.870 --> 01:06:56.880 align:start position:0% +to + + +01:06:56.880 --> 01:06:59.109 align:start position:0% +to +solve<01:06:57.520> the<01:06:57.760> non-convex<01:06:58.480> optimization + +01:06:59.109 --> 01:06:59.119 align:start position:0% +solve the non-convex optimization + + +01:06:59.119 --> 01:07:00.230 align:start position:0% +solve the non-convex optimization +problem + +01:07:00.230 --> 01:07:00.240 align:start position:0% +problem + + +01:07:00.240 --> 01:07:03.029 align:start position:0% +problem +so<01:07:00.559> one<01:07:00.799> of<01:07:00.880> the<01:07:01.039> mysteries<01:07:02.000> that<01:07:02.640> we<01:07:02.799> have + +01:07:03.029 --> 01:07:03.039 align:start position:0% +so one of the mysteries that we have + + +01:07:03.039 --> 01:07:04.230 align:start position:0% +so one of the mysteries that we have +been<01:07:03.280> thinking<01:07:03.599> about<01:07:03.920> is + +01:07:04.230 --> 01:07:04.240 align:start position:0% +been thinking about is + + +01:07:04.240 --> 01:07:09.109 align:start position:0% +been thinking about is +why<01:07:05.520> in<01:07:05.680> fact<01:07:06.079> sgd<01:07:07.200> or<01:07:07.440> gradient<01:07:07.920> descent + +01:07:09.109 --> 01:07:09.119 align:start position:0% +why in fact sgd or gradient descent + + +01:07:09.119 --> 01:07:14.230 align:start position:0% +why in fact sgd or gradient descent +works<01:07:10.400> well<01:07:11.359> in<01:07:12.079> many<01:07:12.839> cases + +01:07:14.230 --> 01:07:14.240 align:start position:0% +works well in many cases + + +01:07:14.240 --> 01:07:19.349 align:start position:0% +works well in many cases +in<01:07:14.480> many<01:07:16.240> non-convex + +01:07:19.349 --> 01:07:19.359 align:start position:0% + + + +01:07:19.359 --> 01:07:21.910 align:start position:0% + +deep<01:07:19.680> learning<01:07:20.559> optimization<01:07:21.280> problems<01:07:21.680> not + +01:07:21.910 --> 01:07:21.920 align:start position:0% +deep learning optimization problems not + + +01:07:21.920 --> 01:07:24.069 align:start position:0% +deep learning optimization problems not +in<01:07:22.079> all<01:07:22.240> of<01:07:22.400> them<01:07:22.559> but<01:07:22.799> in<01:07:22.960> many + +01:07:24.069 --> 01:07:24.079 align:start position:0% +in all of them but in many + + +01:07:24.079 --> 01:07:25.829 align:start position:0% +in all of them but in many +optimization<01:07:24.720> problems<01:07:25.119> they<01:07:25.280> work<01:07:25.520> pretty + +01:07:25.829 --> 01:07:25.839 align:start position:0% +optimization problems they work pretty + + +01:07:25.839 --> 01:07:28.470 align:start position:0% +optimization problems they work pretty +well<01:07:26.079> so<01:07:26.480> why<01:07:26.720> is<01:07:26.880> that<01:07:27.119> the<01:07:27.280> case + +01:07:28.470 --> 01:07:28.480 align:start position:0% +well so why is that the case + + +01:07:28.480 --> 01:07:32.470 align:start position:0% +well so why is that the case +so<01:07:30.400> that's<01:07:30.799> basically<01:07:31.440> going<01:07:31.680> to<01:07:31.920> be<01:07:32.160> the + +01:07:32.470 --> 01:07:32.480 align:start position:0% +so that's basically going to be the + + +01:07:32.480 --> 01:07:33.109 align:start position:0% +so that's basically going to be the +topic + +01:07:33.109 --> 01:07:33.119 align:start position:0% +topic + + +01:07:33.119 --> 01:07:36.630 align:start position:0% +topic +for<01:07:34.000> uh<01:07:34.720> or<01:07:34.960> next<01:07:35.359> lecture<01:07:35.839> to<01:07:36.079> see + +01:07:36.630 --> 01:07:36.640 align:start position:0% +for uh or next lecture to see + + +01:07:36.640 --> 01:07:38.950 align:start position:0% +for uh or next lecture to see +what<01:07:36.880> is<01:07:37.119> especially<01:07:37.760> about<01:07:38.079> these<01:07:38.640> deep + +01:07:38.950 --> 01:07:38.960 align:start position:0% +what is especially about these deep + + +01:07:38.960 --> 01:07:40.870 align:start position:0% +what is especially about these deep +learning<01:07:39.440> architectures<01:07:40.319> that + +01:07:40.870 --> 01:07:40.880 align:start position:0% +learning architectures that + + +01:07:40.880 --> 01:07:42.789 align:start position:0% +learning architectures that +even<01:07:41.119> though<01:07:41.599> you<01:07:41.760> have<01:07:42.000> a<01:07:42.079> non-convex + +01:07:42.789 --> 01:07:42.799 align:start position:0% +even though you have a non-convex + + +01:07:42.799 --> 01:07:44.230 align:start position:0% +even though you have a non-convex +optimization<01:07:43.760> these + +01:07:44.230 --> 01:07:44.240 align:start position:0% +optimization these + + +01:07:44.240 --> 01:07:45.990 align:start position:0% +optimization these +local<01:07:44.640> search<01:07:44.960> methods<01:07:45.359> they<01:07:45.520> work<01:07:45.760> with + +01:07:45.990 --> 01:07:46.000 align:start position:0% +local search methods they work with + + +01:07:46.000 --> 01:07:47.990 align:start position:0% +local search methods they work with +either<01:07:46.400> do<01:07:46.559> we<01:07:46.720> get<01:07:46.880> the<01:07:47.039> loss<01:07:47.359> function<01:07:47.760> like + +01:07:47.990 --> 01:07:48.000 align:start position:0% +either do we get the loss function like + + +01:07:48.000 --> 01:07:48.710 align:start position:0% +either do we get the loss function like +this + +01:07:48.710 --> 01:07:48.720 align:start position:0% +this + + +01:07:48.720 --> 01:07:50.470 align:start position:0% +this +because<01:07:49.119> if<01:07:49.280> i<01:07:49.359> get<01:07:49.520> the<01:07:49.680> loss<01:07:49.920> function<01:07:50.319> like + +01:07:50.470 --> 01:07:50.480 align:start position:0% +because if i get the loss function like + + +01:07:50.480 --> 01:07:51.750 align:start position:0% +because if i get the loss function like +this<01:07:50.880> and<01:07:51.039> if<01:07:51.200> i + +01:07:51.750 --> 01:07:51.760 align:start position:0% +this and if i + + +01:07:51.760 --> 01:07:55.589 align:start position:0% +this and if i +start<01:07:52.319> my<01:07:52.720> sgd<01:07:53.839> here + +01:07:55.589 --> 01:07:55.599 align:start position:0% +start my sgd here + + +01:07:55.599 --> 01:07:57.750 align:start position:0% +start my sgd here +then<01:07:55.839> i'm<01:07:56.000> going<01:07:56.079> to<01:07:56.319> end<01:07:56.480> up<01:07:56.640> in<01:07:56.799> this<01:07:57.359> point + +01:07:57.750 --> 01:07:57.760 align:start position:0% +then i'm going to end up in this point + + +01:07:57.760 --> 01:07:59.510 align:start position:0% +then i'm going to end up in this point +which<01:07:58.079> is<01:07:58.240> a<01:07:58.559> bad + +01:07:59.510 --> 01:07:59.520 align:start position:0% +which is a bad + + +01:07:59.520 --> 01:08:03.750 align:start position:0% +which is a bad +and<01:07:59.839> poor<01:08:00.319> local<01:08:01.599> local<01:08:02.000> minima + +01:08:03.750 --> 01:08:03.760 align:start position:0% +and poor local local minima + + +01:08:03.760 --> 01:08:05.750 align:start position:0% +and poor local local minima +compared<01:08:04.160> to<01:08:04.640> this<01:08:04.880> point<01:08:05.119> which<01:08:05.359> is<01:08:05.520> my + +01:08:05.750 --> 01:08:05.760 align:start position:0% +compared to this point which is my + + +01:08:05.760 --> 01:08:07.430 align:start position:0% +compared to this point which is my +global + +01:08:07.430 --> 01:08:07.440 align:start position:0% +global + + +01:08:07.440 --> 01:08:10.150 align:start position:0% +global +minimum<01:08:08.400> is<01:08:08.720> this<01:08:09.039> the<01:08:09.359> picture<01:08:09.760> that<01:08:10.000> we + +01:08:10.150 --> 01:08:10.160 align:start position:0% +minimum is this the picture that we + + +01:08:10.160 --> 01:08:11.510 align:start position:0% +minimum is this the picture that we +should<01:08:10.400> have<01:08:10.640> in<01:08:10.720> mind + +01:08:11.510 --> 01:08:11.520 align:start position:0% +should have in mind + + +01:08:11.520 --> 01:08:14.950 align:start position:0% +should have in mind +or<01:08:11.839> maybe<01:08:12.160> we<01:08:12.960> don't<01:08:13.359> get<01:08:13.680> in<01:08:13.960> non-convexity + +01:08:14.950 --> 01:08:14.960 align:start position:0% +or maybe we don't get in non-convexity + + +01:08:14.960 --> 01:08:15.990 align:start position:0% +or maybe we don't get in non-convexity +like<01:08:15.280> this + +01:08:15.990 --> 01:08:16.000 align:start position:0% +like this + + +01:08:16.000 --> 01:08:19.110 align:start position:0% +like this +and<01:08:16.159> that's<01:08:16.480> why<01:08:16.880> sgd<01:08:17.920> uh<01:08:18.319> works<01:08:18.719> well + +01:08:19.110 --> 01:08:19.120 align:start position:0% +and that's why sgd uh works well + + +01:08:19.120 --> 01:08:22.950 align:start position:0% +and that's why sgd uh works well +in<01:08:19.759> practice<01:08:21.920> uh + +01:08:22.950 --> 01:08:22.960 align:start position:0% +in practice uh + + +01:08:22.960 --> 01:08:25.990 align:start position:0% +in practice uh +okay<01:08:23.440> so<01:08:23.839> uh<01:08:24.239> there<01:08:24.480> are<01:08:24.719> other<01:08:25.040> variations + +01:08:25.990 --> 01:08:26.000 align:start position:0% +okay so uh there are other variations + + +01:08:26.000 --> 01:08:28.870 align:start position:0% +okay so uh there are other variations +of<01:08:26.400> uh<01:08:27.040> these<01:08:27.359> local<01:08:27.839> search<01:08:28.080> methods<01:08:28.560> like + +01:08:28.870 --> 01:08:28.880 align:start position:0% +of uh these local search methods like + + +01:08:28.880 --> 01:08:30.349 align:start position:0% +of uh these local search methods like +momentum + +01:08:30.349 --> 01:08:30.359 align:start position:0% +momentum + + +01:08:30.359 --> 01:08:32.950 align:start position:0% +momentum +[Music] + +01:08:32.950 --> 01:08:32.960 align:start position:0% +[Music] + + +01:08:32.960 --> 01:08:38.470 align:start position:0% +[Music] +momentum<01:08:35.120> you<01:08:35.279> can<01:08:35.520> also<01:08:35.920> have<01:08:36.839> momentum + +01:08:38.470 --> 01:08:38.480 align:start position:0% +momentum you can also have momentum + + +01:08:38.480 --> 01:08:41.189 align:start position:0% +momentum you can also have momentum +plus<01:08:38.799> sgd<01:08:39.759> so<01:08:40.000> here<01:08:40.239> you<01:08:40.400> are<01:08:40.480> not<01:08:40.719> using<01:08:41.040> the + +01:08:41.189 --> 01:08:41.199 align:start position:0% +plus sgd so here you are not using the + + +01:08:41.199 --> 01:08:42.709 align:start position:0% +plus sgd so here you are not using the +gradient<01:08:41.839> itself<01:08:42.400> but + +01:08:42.709 --> 01:08:42.719 align:start position:0% +gradient itself but + + +01:08:42.719 --> 01:08:45.269 align:start position:0% +gradient itself but +you<01:08:42.880> are<01:08:43.359> keeping<01:08:43.839> some<01:08:44.080> updates<01:08:44.640> and<01:08:44.880> combine + +01:08:45.269 --> 01:08:45.279 align:start position:0% +you are keeping some updates and combine + + +01:08:45.279 --> 01:08:46.709 align:start position:0% +you are keeping some updates and combine +it<01:08:45.440> with<01:08:45.600> your<01:08:45.839> gradient + +01:08:46.709 --> 01:08:46.719 align:start position:0% +it with your gradient + + +01:08:46.719 --> 01:08:48.550 align:start position:0% +it with your gradient +in<01:08:46.880> case<01:08:47.199> your<01:08:47.440> curvature<01:08:47.920> is<01:08:48.080> high<01:08:48.319> so<01:08:48.480> you + +01:08:48.550 --> 01:08:48.560 align:start position:0% +in case your curvature is high so you + + +01:08:48.560 --> 01:08:50.630 align:start position:0% +in case your curvature is high so you +don't<01:08:48.719> want<01:08:48.960> to<01:08:49.600> change<01:08:49.920> the<01:08:50.239> update + +01:08:50.630 --> 01:08:50.640 align:start position:0% +don't want to change the update + + +01:08:50.640 --> 01:08:52.070 align:start position:0% +don't want to change the update +direction<01:08:51.199> suddenly + +01:08:52.070 --> 01:08:52.080 align:start position:0% +direction suddenly + + +01:08:52.080 --> 01:08:55.030 align:start position:0% +direction suddenly +so<01:08:52.319> your<01:08:53.199> updates<01:08:53.679> will<01:08:53.920> have<01:08:54.239> some<01:08:54.480> momentum + +01:08:55.030 --> 01:08:55.040 align:start position:0% +so your updates will have some momentum + + +01:08:55.040 --> 01:08:55.669 align:start position:0% +so your updates will have some momentum +from<01:08:55.199> the + +01:08:55.669 --> 01:08:55.679 align:start position:0% +from the + + +01:08:55.679 --> 01:08:58.070 align:start position:0% +from the +previous<01:08:56.080> parts<01:08:56.719> you<01:08:56.880> can<01:08:57.040> combine<01:08:57.520> that<01:08:57.759> with + +01:08:58.070 --> 01:08:58.080 align:start position:0% +previous parts you can combine that with + + +01:08:58.080 --> 01:08:59.189 align:start position:0% +previous parts you can combine that with +sgd + +01:08:59.189 --> 01:08:59.199 align:start position:0% +sgd + + +01:08:59.199 --> 01:09:03.430 align:start position:0% +sgd +and<01:08:59.440> there<01:08:59.759> are<01:08:59.920> some<01:09:02.159> solvers<01:09:02.799> that<01:09:03.040> use + +01:09:03.430 --> 01:09:03.440 align:start position:0% +and there are some solvers that use + + +01:09:03.440 --> 01:09:07.189 align:start position:0% +and there are some solvers that use +adaptive<01:09:04.080> learning<01:09:05.839> adaptive + +01:09:07.189 --> 01:09:07.199 align:start position:0% +adaptive learning adaptive + + +01:09:07.199 --> 01:09:10.630 align:start position:0% +adaptive learning adaptive +learning<01:09:07.679> rates + +01:09:10.630 --> 01:09:10.640 align:start position:0% + + + +01:09:10.640 --> 01:09:13.510 align:start position:0% + +so<01:09:10.880> the<01:09:12.080> main<01:09:12.480> idea<01:09:12.719> is<01:09:12.880> that<01:09:13.040> if<01:09:13.199> your + +01:09:13.510 --> 01:09:13.520 align:start position:0% +so the main idea is that if your + + +01:09:13.520 --> 01:09:14.470 align:start position:0% +so the main idea is that if your +gradient + +01:09:14.470 --> 01:09:14.480 align:start position:0% +gradient + + +01:09:14.480 --> 01:09:17.510 align:start position:0% +gradient +if<01:09:14.560> your<01:09:14.880> gradient<01:09:15.839> is<01:09:16.159> large<01:09:16.960> in<01:09:17.120> terms<01:09:17.359> of + +01:09:17.510 --> 01:09:17.520 align:start position:0% +if your gradient is large in terms of + + +01:09:17.520 --> 01:09:18.470 align:start position:0% +if your gradient is large in terms of +the<01:09:17.679> norm + +01:09:18.470 --> 01:09:18.480 align:start position:0% +the norm + + +01:09:18.480 --> 01:09:21.910 align:start position:0% +the norm +it<01:09:18.719> means<01:09:18.960> you<01:09:19.359> uh<01:09:19.679> if<01:09:19.920> you<01:09:20.400> use<01:09:20.640> a<01:09:20.799> fixed<01:09:21.279> rate + +01:09:21.910 --> 01:09:21.920 align:start position:0% +it means you uh if you use a fixed rate + + +01:09:21.920 --> 01:09:24.470 align:start position:0% +it means you uh if you use a fixed rate +then<01:09:22.319> the<01:09:22.480> parameter<01:09:23.199> assigned<01:09:23.520> to<01:09:23.679> that<01:09:23.920> will + +01:09:24.470 --> 01:09:24.480 align:start position:0% +then the parameter assigned to that will + + +01:09:24.480 --> 01:09:26.630 align:start position:0% +then the parameter assigned to that will +change<01:09:24.799> significantly<01:09:25.759> so<01:09:25.920> you<01:09:26.159> want<01:09:26.400> your + +01:09:26.630 --> 01:09:26.640 align:start position:0% +change significantly so you want your + + +01:09:26.640 --> 01:09:27.910 align:start position:0% +change significantly so you want your +learning<01:09:26.960> rate<01:09:27.279> to<01:09:27.520> be + +01:09:27.910 --> 01:09:27.920 align:start position:0% +learning rate to be + + +01:09:27.920 --> 01:09:31.430 align:start position:0% +learning rate to be +small<01:09:28.799> and<01:09:28.960> if<01:09:29.120> your<01:09:29.359> gradient<01:09:29.920> is<01:09:30.400> small + +01:09:31.430 --> 01:09:31.440 align:start position:0% +small and if your gradient is small + + +01:09:31.440 --> 01:09:33.829 align:start position:0% +small and if your gradient is small +then<01:09:31.679> you<01:09:32.400> want<01:09:32.719> your<01:09:32.960> learning<01:09:33.279> rate<01:09:33.520> to<01:09:33.600> be + +01:09:33.829 --> 01:09:33.839 align:start position:0% +then you want your learning rate to be + + +01:09:33.839 --> 01:09:35.430 align:start position:0% +then you want your learning rate to be +larger<01:09:34.239> because<01:09:34.480> otherwise<01:09:34.960> you<01:09:35.120> are<01:09:35.279> not + +01:09:35.430 --> 01:09:35.440 align:start position:0% +larger because otherwise you are not + + +01:09:35.440 --> 01:09:35.910 align:start position:0% +larger because otherwise you are not +going<01:09:35.600> to + +01:09:35.910 --> 01:09:35.920 align:start position:0% +going to + + +01:09:35.920 --> 01:09:39.430 align:start position:0% +going to +significantly<01:09:36.719> change<01:09:36.960> your<01:09:37.440> parameters<01:09:38.080> so + +01:09:39.430 --> 01:09:39.440 align:start position:0% +significantly change your parameters so + + +01:09:39.440 --> 01:09:42.390 align:start position:0% +significantly change your parameters so +some<01:09:39.679> of<01:09:39.839> the<01:09:40.159> interesting<01:09:40.640> algorithms<01:09:41.359> here + +01:09:42.390 --> 01:09:42.400 align:start position:0% +some of the interesting algorithms here + + +01:09:42.400 --> 01:09:42.789 align:start position:0% +some of the interesting algorithms here +are + +01:09:42.789 --> 01:09:42.799 align:start position:0% +are + + +01:09:42.799 --> 01:09:46.950 align:start position:0% +are +rms<01:09:43.440> prop<01:09:44.960> and<01:09:45.600> adam + +01:09:46.950 --> 01:09:46.960 align:start position:0% +rms prop and adam + + +01:09:46.960 --> 01:09:49.349 align:start position:0% +rms prop and adam +which<01:09:47.199> is<01:09:47.359> basically<01:09:47.920> the<01:09:48.640> combination<01:09:49.199> of + +01:09:49.349 --> 01:09:49.359 align:start position:0% +which is basically the combination of + + +01:09:49.359 --> 01:09:50.470 align:start position:0% +which is basically the combination of +rms<01:09:49.839> prop + +01:09:50.470 --> 01:09:50.480 align:start position:0% +rms prop + + +01:09:50.480 --> 01:09:53.910 align:start position:0% +rms prop +with<01:09:50.960> momentum<01:09:52.000> with<01:09:52.239> some<01:09:52.560> uh<01:09:52.960> with<01:09:53.199> some<01:09:53.520> uh + +01:09:53.910 --> 01:09:53.920 align:start position:0% +with momentum with some uh with some uh + + +01:09:53.920 --> 01:09:56.470 align:start position:0% +with momentum with some uh with some uh +changes<01:09:54.960> so<01:09:55.199> unfortunately<01:09:55.840> we<01:09:56.000> don't<01:09:56.159> have + +01:09:56.470 --> 01:09:56.480 align:start position:0% +changes so unfortunately we don't have + + +01:09:56.480 --> 01:09:58.630 align:start position:0% +changes so unfortunately we don't have +uh<01:09:56.800> time<01:09:57.120> to<01:09:57.360> go<01:09:57.520> into<01:09:57.760> the<01:09:57.920> details<01:09:58.320> so<01:09:58.480> you + +01:09:58.630 --> 01:09:58.640 align:start position:0% +uh time to go into the details so you + + +01:09:58.640 --> 01:10:00.229 align:start position:0% +uh time to go into the details so you +can<01:09:58.960> you<01:09:59.120> know<01:09:59.360> take<01:09:59.520> a<01:09:59.600> look<01:09:59.840> at<01:09:59.920> some<01:10:00.159> of + +01:10:00.229 --> 01:10:00.239 align:start position:0% +can you know take a look at some of + + +01:10:00.239 --> 01:10:01.510 align:start position:0% +can you know take a look at some of +these<01:10:00.480> concepts<01:10:00.960> these<01:10:01.199> are + +01:10:01.510 --> 01:10:01.520 align:start position:0% +these concepts these are + + +01:10:01.520 --> 01:10:04.630 align:start position:0% +these concepts these are +you<01:10:01.679> know<01:10:01.840> very<01:10:02.159> practical<01:10:02.880> uh<01:10:03.600> solvers + +01:10:04.630 --> 01:10:04.640 align:start position:0% +you know very practical uh solvers + + +01:10:04.640 --> 01:10:08.630 align:start position:0% +you know very practical uh solvers +uh<01:10:05.120> in<01:10:05.440> for<01:10:05.600> deep<01:10:05.840> learning<01:10:06.640> optimizations + +01:10:08.630 --> 01:10:08.640 align:start position:0% +uh in for deep learning optimizations + + +01:10:08.640 --> 01:10:10.229 align:start position:0% +uh in for deep learning optimizations +right<01:10:08.880> so<01:10:09.120> let<01:10:09.280> me<01:10:09.440> see<01:10:09.600> if<01:10:09.760> there<01:10:09.920> are<01:10:10.000> some + +01:10:10.229 --> 01:10:10.239 align:start position:0% +right so let me see if there are some + + +01:10:10.239 --> 01:10:15.270 align:start position:0% +right so let me see if there are some +questions + +01:10:15.270 --> 01:10:15.280 align:start position:0% + + + +01:10:15.280 --> 01:10:18.790 align:start position:0% + +um<01:10:16.480> yeah<01:10:16.719> so<01:10:17.040> the<01:10:17.199> derivatives<01:10:18.000> of<01:10:18.320> the + +01:10:18.790 --> 01:10:18.800 align:start position:0% +um yeah so the derivatives of the + + +01:10:18.800 --> 01:10:21.590 align:start position:0% +um yeah so the derivatives of the +pooling<01:10:19.360> or<01:10:19.520> the<01:10:19.760> whole<01:10:20.000> network<01:10:20.960> uh<01:10:21.280> you<01:10:21.440> can + +01:10:21.590 --> 01:10:21.600 align:start position:0% +pooling or the whole network uh you can + + +01:10:21.600 --> 01:10:23.030 align:start position:0% +pooling or the whole network uh you can +use<01:10:21.840> the<01:10:22.000> chain<01:10:22.320> rule + +01:10:23.030 --> 01:10:23.040 align:start position:0% +use the chain rule + + +01:10:23.040 --> 01:10:25.110 align:start position:0% +use the chain rule +uh<01:10:23.280> in<01:10:23.440> order<01:10:23.760> to<01:10:23.920> compute<01:10:24.239> it<01:10:24.400> there's<01:10:24.640> a + +01:10:25.110 --> 01:10:25.120 align:start position:0% +uh in order to compute it there's a + + +01:10:25.120 --> 01:10:26.390 align:start position:0% +uh in order to compute it there's a +there's<01:10:25.360> a<01:10:25.440> very<01:10:25.760> efficient + +01:10:26.390 --> 01:10:26.400 align:start position:0% +there's a very efficient + + +01:10:26.400 --> 01:10:28.709 align:start position:0% +there's a very efficient +algorithm<01:10:26.880> called<01:10:27.199> back<01:10:27.520> prop<01:10:28.159> in<01:10:28.320> order<01:10:28.560> to + +01:10:28.709 --> 01:10:28.719 align:start position:0% +algorithm called back prop in order to + + +01:10:28.719 --> 01:10:30.390 align:start position:0% +algorithm called back prop in order to +compute<01:10:29.040> the<01:10:29.199> gradient<01:10:29.760> so<01:10:29.920> you<01:10:30.000> should<01:10:30.159> take + +01:10:30.390 --> 01:10:30.400 align:start position:0% +compute the gradient so you should take + + +01:10:30.400 --> 01:10:32.870 align:start position:0% +compute the gradient so you should take +a<01:10:30.640> look<01:10:30.800> into<01:10:31.040> that + +01:10:32.870 --> 01:10:32.880 align:start position:0% +a look into that + + +01:10:32.880 --> 01:10:34.870 align:start position:0% +a look into that +uh<01:10:33.120> there's<01:10:33.360> another<01:10:33.679> question<01:10:34.080> artist<01:10:34.560> works + +01:10:34.870 --> 01:10:34.880 align:start position:0% +uh there's another question artist works + + +01:10:34.880 --> 01:10:36.790 align:start position:0% +uh there's another question artist works +on<01:10:35.040> analyzing<01:10:35.600> sgd + +01:10:36.790 --> 01:10:36.800 align:start position:0% +on analyzing sgd + + +01:10:36.800 --> 01:10:39.030 align:start position:0% +on analyzing sgd +with<01:10:37.040> random<01:10:37.440> box<01:10:37.760> on<01:10:37.920> the<01:10:38.159> lost<01:10:38.400> surface<01:10:38.880> due + +01:10:39.030 --> 01:10:39.040 align:start position:0% +with random box on the lost surface due + + +01:10:39.040 --> 01:10:40.630 align:start position:0% +with random box on the lost surface due +to<01:10:39.199> the<01:10:39.360> noise<01:10:39.600> of<01:10:39.760> the<01:10:39.840> gradient + +01:10:40.630 --> 01:10:40.640 align:start position:0% +to the noise of the gradient + + +01:10:40.640 --> 01:10:43.350 align:start position:0% +to the noise of the gradient +yes<01:10:40.880> there<01:10:41.199> are<01:10:41.440> some<01:10:42.080> works<01:10:42.480> on<01:10:42.719> analyzing + +01:10:43.350 --> 01:10:43.360 align:start position:0% +yes there are some works on analyzing + + +01:10:43.360 --> 01:10:44.630 align:start position:0% +yes there are some works on analyzing +sgd + +01:10:44.630 --> 01:10:44.640 align:start position:0% +sgd + + +01:10:44.640 --> 01:10:47.270 align:start position:0% +sgd +uh<01:10:45.120> so<01:10:45.360> next<01:10:45.679> lecture<01:10:46.080> we<01:10:46.239> are<01:10:46.320> going<01:10:46.560> to<01:10:46.880> uh + +01:10:47.270 --> 01:10:47.280 align:start position:0% +uh so next lecture we are going to uh + + +01:10:47.280 --> 01:10:49.189 align:start position:0% +uh so next lecture we are going to uh +read<01:10:47.520> some<01:10:47.679> of<01:10:47.840> these<01:10:48.080> papers + +01:10:49.189 --> 01:10:49.199 align:start position:0% +read some of these papers + + +01:10:49.199 --> 01:10:51.430 align:start position:0% +read some of these papers +uh<01:10:49.840> first<01:10:50.159> we'll<01:10:50.400> look<01:10:50.560> into<01:10:50.880> the<01:10:51.040> gradient + +01:10:51.430 --> 01:10:51.440 align:start position:0% +uh first we'll look into the gradient + + +01:10:51.440 --> 01:10:53.189 align:start position:0% +uh first we'll look into the gradient +descent<01:10:51.920> full<01:10:52.080> batch<01:10:52.400> gradient<01:10:52.719> descent<01:10:53.120> and + +01:10:53.189 --> 01:10:53.199 align:start position:0% +descent full batch gradient descent and + + +01:10:53.199 --> 01:10:54.870 align:start position:0% +descent full batch gradient descent and +then<01:10:53.360> we'll<01:10:53.600> look<01:10:53.840> into + +01:10:54.870 --> 01:10:54.880 align:start position:0% +then we'll look into + + +01:10:54.880 --> 01:11:05.510 align:start position:0% +then we'll look into +understanding<01:10:56.840> sgt<01:10:57.920> so<01:10:58.159> any<01:10:58.480> other<01:10:58.719> questions + +01:11:05.510 --> 01:11:05.520 align:start position:0% + + + +01:11:05.520 --> 01:11:07.669 align:start position:0% + +so<01:11:05.679> just<01:11:06.000> a<01:11:06.239> note<01:11:06.560> if<01:11:06.719> there<01:11:06.880> are<01:11:07.199> no<01:11:07.440> more + +01:11:07.669 --> 01:11:07.679 align:start position:0% +so just a note if there are no more + + +01:11:07.679 --> 01:11:09.350 align:start position:0% +so just a note if there are no more +questions + +01:11:09.350 --> 01:11:09.360 align:start position:0% +questions + + +01:11:09.360 --> 01:11:12.790 align:start position:0% +questions +make<01:11:09.679> sure<01:11:10.080> to<01:11:10.560> read<01:11:10.800> the<01:11:11.120> papers<01:11:12.320> uh + +01:11:12.790 --> 01:11:12.800 align:start position:0% +make sure to read the papers uh + + +01:11:12.800 --> 01:11:16.149 align:start position:0% +make sure to read the papers uh +before<01:11:14.480> before<01:11:14.880> the<01:11:15.040> next<01:11:15.360> lecture + +01:11:16.149 --> 01:11:16.159 align:start position:0% +before before the next lecture + + +01:11:16.159 --> 01:11:18.390 align:start position:0% +before before the next lecture +so<01:11:16.400> actually<01:11:16.719> i<01:11:16.800> see<01:11:17.040> one<01:11:17.679> hand<01:11:18.000> is<01:11:18.159> being + +01:11:18.390 --> 01:11:18.400 align:start position:0% +so actually i see one hand is being + + +01:11:18.400 --> 01:11:21.510 align:start position:0% +so actually i see one hand is being +raised<01:11:18.800> uh<01:11:19.120> pocket<01:11:19.600> go<01:11:19.760> ahead + +01:11:21.510 --> 01:11:21.520 align:start position:0% +raised uh pocket go ahead + + +01:11:21.520 --> 01:11:24.630 align:start position:0% +raised uh pocket go ahead +yes<01:11:21.920> so<01:11:22.239> i<01:11:22.400> just<01:11:22.560> have<01:11:22.719> one<01:11:22.960> question<01:11:23.360> that + +01:11:24.630 --> 01:11:24.640 align:start position:0% +yes so i just have one question that + + +01:11:24.640 --> 01:11:27.189 align:start position:0% +yes so i just have one question that +how<01:11:24.800> do<01:11:24.880> you<01:11:25.520> go<01:11:25.679> about<01:11:26.080> choosing<01:11:26.640> which + +01:11:27.189 --> 01:11:27.199 align:start position:0% +how do you go about choosing which + + +01:11:27.199 --> 01:11:29.590 align:start position:0% +how do you go about choosing which +optimizer<01:11:27.840> to<01:11:28.000> use<01:11:28.320> whether<01:11:28.640> to<01:11:28.719> use<01:11:29.040> r<01:11:29.280> but + +01:11:29.590 --> 01:11:29.600 align:start position:0% +optimizer to use whether to use r but + + +01:11:29.600 --> 01:11:32.790 align:start position:0% +optimizer to use whether to use r but +like<01:11:30.320> on<01:11:30.560> a<01:11:30.640> stop<01:11:30.960> atom<01:11:31.440> or<01:11:31.679> sgd<01:11:32.320> is<01:11:32.480> there<01:11:32.640> a + +01:11:32.790 --> 01:11:32.800 align:start position:0% +like on a stop atom or sgd is there a + + +01:11:32.800 --> 01:11:35.350 align:start position:0% +like on a stop atom or sgd is there a +trick + +01:11:35.350 --> 01:11:35.360 align:start position:0% + + + +01:11:35.360 --> 01:11:38.790 align:start position:0% + +excellent<01:11:35.920> question<01:11:36.560> uh<01:11:36.960> the<01:11:37.840> answer<01:11:38.400> is<01:11:38.640> we + +01:11:38.790 --> 01:11:38.800 align:start position:0% +excellent question uh the answer is we + + +01:11:38.800 --> 01:11:39.990 align:start position:0% +excellent question uh the answer is we +don't<01:11:39.199> know<01:11:39.520> so + +01:11:39.990 --> 01:11:40.000 align:start position:0% +don't know so + + +01:11:40.000 --> 01:11:43.669 align:start position:0% +don't know so +it's<01:11:40.159> very<01:11:40.480> empirical<01:11:41.760> and<01:11:42.640> or<01:11:42.880> understanding + +01:11:43.669 --> 01:11:43.679 align:start position:0% +it's very empirical and or understanding + + +01:11:43.679 --> 01:11:47.030 align:start position:0% +it's very empirical and or understanding +of<01:11:43.920> these<01:11:44.239> optimizers<01:11:45.120> are<01:11:45.360> not + +01:11:47.030 --> 01:11:47.040 align:start position:0% +of these optimizers are not + + +01:11:47.040 --> 01:11:50.229 align:start position:0% +of these optimizers are not +deep<01:11:47.280> enough<01:11:47.679> so<01:11:47.920> in<01:11:48.080> practice<01:11:48.960> uh<01:11:49.520> oftentimes + +01:11:50.229 --> 01:11:50.239 align:start position:0% +deep enough so in practice uh oftentimes + + +01:11:50.239 --> 01:11:51.830 align:start position:0% +deep enough so in practice uh oftentimes +you<01:11:50.480> try<01:11:50.880> some<01:11:51.120> of<01:11:51.280> these + +01:11:51.830 --> 01:11:51.840 align:start position:0% +you try some of these + + +01:11:51.840 --> 01:11:55.189 align:start position:0% +you try some of these +different<01:11:52.239> optimization<01:11:52.960> problems<01:11:53.920> and<01:11:54.159> then + +01:11:55.189 --> 01:11:55.199 align:start position:0% +different optimization problems and then + + +01:11:55.199 --> 01:11:58.470 align:start position:0% +different optimization problems and then +see<01:11:55.440> which<01:11:55.679> one<01:11:56.400> works<01:11:56.719> the<01:11:56.800> best + +01:11:58.470 --> 01:11:58.480 align:start position:0% +see which one works the best + + +01:11:58.480 --> 01:12:01.590 align:start position:0% +see which one works the best +there<01:11:58.640> is<01:11:58.719> another<01:11:59.120> hand<01:11:59.360> raised<01:12:00.239> go<01:12:00.400> ahead + +01:12:01.590 --> 01:12:01.600 align:start position:0% +there is another hand raised go ahead + + +01:12:01.600 --> 01:12:04.870 align:start position:0% +there is another hand raised go ahead +hi<01:12:01.920> professor<01:12:04.239> with<01:12:04.480> the + +01:12:04.870 --> 01:12:04.880 align:start position:0% +hi professor with the + + +01:12:04.880 --> 01:12:07.590 align:start position:0% +hi professor with the +local<01:12:05.280> minimum<01:12:05.760> and<01:12:05.920> the<01:12:06.080> global<01:12:06.480> minimum + +01:12:07.590 --> 01:12:07.600 align:start position:0% +local minimum and the global minimum + + +01:12:07.600 --> 01:12:08.070 align:start position:0% +local minimum and the global minimum +diagram + +01:12:08.070 --> 01:12:08.080 align:start position:0% +diagram + + +01:12:08.080 --> 01:12:10.709 align:start position:0% +diagram +that<01:12:08.159> you<01:12:08.400> showed<01:12:09.120> is<01:12:09.280> it<01:12:09.600> isn't<01:12:10.000> it<01:12:10.400> more + +01:12:10.709 --> 01:12:10.719 align:start position:0% +that you showed is it isn't it more + + +01:12:10.719 --> 01:12:11.910 align:start position:0% +that you showed is it isn't it more +probable<01:12:11.360> that<01:12:11.520> we'll + +01:12:11.910 --> 01:12:11.920 align:start position:0% +probable that we'll + + +01:12:11.920 --> 01:12:15.030 align:start position:0% +probable that we'll +get<01:12:12.239> saddle<01:12:12.640> points<01:12:13.360> instead + +01:12:15.030 --> 01:12:15.040 align:start position:0% +get saddle points instead + + +01:12:15.040 --> 01:12:19.189 align:start position:0% +get saddle points instead +uh<01:12:15.520> yes<01:12:17.040> yes<01:12:17.440> there<01:12:17.679> is<01:12:18.000> also<01:12:18.480> like<01:12:18.719> a<01:12:18.800> saddle + +01:12:19.189 --> 01:12:19.199 align:start position:0% +uh yes yes there is also like a saddle + + +01:12:19.199 --> 01:12:21.590 align:start position:0% +uh yes yes there is also like a saddle +point<01:12:19.600> problem<01:12:20.080> when<01:12:20.320> you<01:12:20.560> have + +01:12:21.590 --> 01:12:21.600 align:start position:0% +point problem when you have + + +01:12:21.600 --> 01:12:23.590 align:start position:0% +point problem when you have +some<01:12:21.920> eigenvalues<01:12:22.640> of<01:12:22.719> the<01:12:22.880> haitians<01:12:23.360> that + +01:12:23.590 --> 01:12:23.600 align:start position:0% +some eigenvalues of the haitians that + + +01:12:23.600 --> 01:12:24.709 align:start position:0% +some eigenvalues of the haitians that +are<01:12:24.159> uh + +01:12:24.709 --> 01:12:24.719 align:start position:0% +are uh + + +01:12:24.719 --> 01:12:26.470 align:start position:0% +are uh +positive<01:12:25.280> and<01:12:25.440> some<01:12:25.679> eigenvalues<01:12:26.320> are + +01:12:26.470 --> 01:12:26.480 align:start position:0% +positive and some eigenvalues are + + +01:12:26.480 --> 01:12:29.189 align:start position:0% +positive and some eigenvalues are +negative<01:12:27.600> so<01:12:27.840> i'm<01:12:27.920> gonna<01:12:28.159> talk<01:12:28.400> about<01:12:28.719> that + +01:12:29.189 --> 01:12:29.199 align:start position:0% +negative so i'm gonna talk about that + + +01:12:29.199 --> 01:12:32.470 align:start position:0% +negative so i'm gonna talk about that +in<01:12:29.360> fact<01:12:29.920> having<01:12:30.640> sgd<01:12:31.600> is<01:12:31.760> helpful + +01:12:32.470 --> 01:12:32.480 align:start position:0% +in fact having sgd is helpful + + +01:12:32.480 --> 01:12:35.110 align:start position:0% +in fact having sgd is helpful +in<01:12:32.960> skipping<01:12:33.440> those<01:12:33.679> saddle<01:12:34.080> points<01:12:34.480> because + +01:12:35.110 --> 01:12:35.120 align:start position:0% +in skipping those saddle points because + + +01:12:35.120 --> 01:12:37.030 align:start position:0% +in skipping those saddle points because +you<01:12:35.280> add<01:12:35.440> a<01:12:35.520> little<01:12:35.760> bit<01:12:35.920> of<01:12:36.080> randomness + +01:12:37.030 --> 01:12:37.040 align:start position:0% +you add a little bit of randomness + + +01:12:37.040 --> 01:12:39.110 align:start position:0% +you add a little bit of randomness +so<01:12:37.280> in<01:12:37.440> other<01:12:37.679> words<01:12:38.000> even<01:12:38.159> though<01:12:38.400> sgd<01:12:38.960> is + +01:12:39.110 --> 01:12:39.120 align:start position:0% +so in other words even though sgd is + + +01:12:39.120 --> 01:12:40.950 align:start position:0% +so in other words even though sgd is +very<01:12:39.360> computationally<01:12:40.239> efficient + +01:12:40.950 --> 01:12:40.960 align:start position:0% +very computationally efficient + + +01:12:40.960 --> 01:12:45.350 align:start position:0% +very computationally efficient +it<01:12:41.199> also<01:12:41.520> helps<01:12:41.840> us<01:12:42.080> to<01:12:42.320> not<01:12:42.880> uh<01:12:43.920> basically + +01:12:45.350 --> 01:12:45.360 align:start position:0% +it also helps us to not uh basically + + +01:12:45.360 --> 01:12:48.550 align:start position:0% +it also helps us to not uh basically +trap<01:12:45.840> in<01:12:45.920> those<01:12:46.159> saddle<01:12:46.560> points<01:12:47.600> but<01:12:48.080> in<01:12:48.320> fact + +01:12:48.550 --> 01:12:48.560 align:start position:0% +trap in those saddle points but in fact + + +01:12:48.560 --> 01:12:50.149 align:start position:0% +trap in those saddle points but in fact +that<01:12:48.880> picture<01:12:49.199> that<01:12:49.440> i<01:12:49.600> draw + +01:12:50.149 --> 01:12:50.159 align:start position:0% +that picture that i draw + + +01:12:50.159 --> 01:12:52.790 align:start position:0% +that picture that i draw +is<01:12:50.480> not<01:12:51.040> very<01:12:51.440> accurate<01:12:51.920> so<01:12:52.080> that's<01:12:52.320> why<01:12:52.480> i<01:12:52.560> put + +01:12:52.790 --> 01:12:52.800 align:start position:0% +is not very accurate so that's why i put + + +01:12:52.800 --> 01:12:54.149 align:start position:0% +is not very accurate so that's why i put +a<01:12:52.960> question<01:12:53.360> mark + +01:12:54.149 --> 01:12:54.159 align:start position:0% +a question mark + + +01:12:54.159 --> 01:12:56.870 align:start position:0% +a question mark +and<01:12:54.640> uh<01:12:55.199> that's<01:12:55.679> basically<01:12:56.320> something<01:12:56.640> that + +01:12:56.870 --> 01:12:56.880 align:start position:0% +and uh that's basically something that + + +01:12:56.880 --> 01:12:58.229 align:start position:0% +and uh that's basically something that +will<01:12:57.199> uh + +01:12:58.229 --> 01:12:58.239 align:start position:0% +will uh + + +01:12:58.239 --> 01:13:02.830 align:start position:0% +will uh +dig<01:12:58.480> a<01:12:58.560> little<01:12:58.800> bit<01:12:58.960> deeper<01:13:00.080> uh<01:13:01.040> in<01:13:01.199> the<01:13:01.360> next + +01:13:02.830 --> 01:13:02.840 align:start position:0% +dig a little bit deeper uh in the next + + +01:13:02.840 --> 01:13:06.790 align:start position:0% +dig a little bit deeper uh in the next +lecture + +01:13:06.790 --> 01:13:06.800 align:start position:0% + + + +01:13:06.800 --> 01:13:11.830 align:start position:0% + +okay<01:13:07.040> any<01:13:08.239> other<01:13:09.040> questions + +01:13:11.830 --> 01:13:11.840 align:start position:0% + + + +01:13:11.840 --> 01:13:13.910 align:start position:0% + +so<01:13:12.080> my<01:13:12.320> question<01:13:12.719> is<01:13:12.880> along<01:13:13.199> the<01:13:13.280> same<01:13:13.520> lines + +01:13:13.910 --> 01:13:13.920 align:start position:0% +so my question is along the same lines + + +01:13:13.920 --> 01:13:15.030 align:start position:0% +so my question is along the same lines +as<01:13:14.080> gold<01:13:14.400> kids + +01:13:15.030 --> 01:13:15.040 align:start position:0% +as gold kids + + +01:13:15.040 --> 01:13:17.990 align:start position:0% +as gold kids +is<01:13:15.120> there<01:13:15.280> an<01:13:15.440> intuition<01:13:15.920> behind<01:13:16.560> what<01:13:17.760> to + +01:13:17.990 --> 01:13:18.000 align:start position:0% +is there an intuition behind what to + + +01:13:18.000 --> 01:13:19.270 align:start position:0% +is there an intuition behind what to +choose<01:13:18.320> when<01:13:18.480> we<01:13:18.640> are + +01:13:19.270 --> 01:13:19.280 align:start position:0% +choose when we are + + +01:13:19.280 --> 01:13:23.669 align:start position:0% +choose when we are +designing<01:13:19.840> a<01:13:19.920> convolution<01:13:21.040> cnn<01:13:22.080> basically + +01:13:23.669 --> 01:13:23.679 align:start position:0% +designing a convolution cnn basically + + +01:13:23.679 --> 01:13:26.310 align:start position:0% +designing a convolution cnn basically +yeah<01:13:23.920> there<01:13:24.159> are<01:13:24.320> some<01:13:24.560> intuitions<01:13:25.440> uh<01:13:26.159> for + +01:13:26.310 --> 01:13:26.320 align:start position:0% +yeah there are some intuitions uh for + + +01:13:26.320 --> 01:13:27.270 align:start position:0% +yeah there are some intuitions uh for +instance<01:13:26.800> uh + +01:13:27.270 --> 01:13:27.280 align:start position:0% +instance uh + + +01:13:27.280 --> 01:13:30.830 align:start position:0% +instance uh +one<01:13:27.920> main<01:13:28.239> thing<01:13:28.480> is<01:13:28.640> the<01:13:29.760> power<01:13:30.159> of<01:13:30.320> over + +01:13:30.830 --> 01:13:30.840 align:start position:0% +one main thing is the power of over + + +01:13:30.840 --> 01:13:32.709 align:start position:0% +one main thing is the power of over +parameterizations<01:13:31.920> in<01:13:32.159> the + +01:13:32.709 --> 01:13:32.719 align:start position:0% +parameterizations in the + + +01:13:32.719 --> 01:13:35.189 align:start position:0% +parameterizations in the +deep<01:13:32.960> learning<01:13:33.920> so<01:13:34.159> in<01:13:34.239> fact<01:13:34.560> we'll<01:13:34.800> see<01:13:35.120> in + +01:13:35.189 --> 01:13:35.199 align:start position:0% +deep learning so in fact we'll see in + + +01:13:35.199 --> 01:13:36.470 align:start position:0% +deep learning so in fact we'll see in +the<01:13:35.440> next<01:13:35.760> couple<01:13:36.080> of + +01:13:36.470 --> 01:13:36.480 align:start position:0% +the next couple of + + +01:13:36.480 --> 01:13:38.709 align:start position:0% +the next couple of +lectures<01:13:37.040> that<01:13:37.760> when<01:13:38.000> your<01:13:38.320> number<01:13:38.640> of + +01:13:38.709 --> 01:13:38.719 align:start position:0% +lectures that when your number of + + +01:13:38.719 --> 01:13:39.669 align:start position:0% +lectures that when your number of +parameters + +01:13:39.669 --> 01:13:39.679 align:start position:0% +parameters + + +01:13:39.679 --> 01:13:41.990 align:start position:0% +parameters +are<01:13:40.480> larger<01:13:40.960> in<01:13:41.040> fact<01:13:41.280> than<01:13:41.440> the<01:13:41.600> number<01:13:41.840> of + +01:13:41.990 --> 01:13:42.000 align:start position:0% +are larger in fact than the number of + + +01:13:42.000 --> 01:13:43.669 align:start position:0% +are larger in fact than the number of +training<01:13:42.400> samples + +01:13:43.669 --> 01:13:43.679 align:start position:0% +training samples + + +01:13:43.679 --> 01:13:45.910 align:start position:0% +training samples +very<01:13:44.000> interesting<01:13:44.560> behaviors<01:13:45.199> are<01:13:45.360> happening + +01:13:45.910 --> 01:13:45.920 align:start position:0% +very interesting behaviors are happening + + +01:13:45.920 --> 01:13:49.510 align:start position:0% +very interesting behaviors are happening +in<01:13:46.480> non-convex<01:13:47.120> optimization<01:13:47.840> problems + +01:13:49.510 --> 01:13:49.520 align:start position:0% +in non-convex optimization problems + + +01:13:49.520 --> 01:13:51.910 align:start position:0% +in non-convex optimization problems +and<01:13:50.239> you<01:13:50.400> know<01:13:50.560> that's<01:13:50.880> one<01:13:51.199> one<01:13:51.440> of<01:13:51.520> them<01:13:51.760> but + +01:13:51.910 --> 01:13:51.920 align:start position:0% +and you know that's one one of them but + + +01:13:51.920 --> 01:13:53.669 align:start position:0% +and you know that's one one of them but +there<01:13:52.159> are<01:13:52.400> a<01:13:52.480> lot<01:13:52.719> of + +01:13:53.669 --> 01:13:53.679 align:start position:0% +there are a lot of + + +01:13:53.679 --> 01:13:55.830 align:start position:0% +there are a lot of +empirical<01:13:54.320> observations<01:13:54.880> that<01:13:55.040> we<01:13:55.199> do<01:13:55.440> not + +01:13:55.830 --> 01:13:55.840 align:start position:0% +empirical observations that we do not + + +01:13:55.840 --> 01:13:58.709 align:start position:0% +empirical observations that we do not +have<01:13:56.560> good<01:13:56.880> theoretical<01:13:57.600> foundations + +01:13:58.709 --> 01:13:58.719 align:start position:0% +have good theoretical foundations + + +01:13:58.719 --> 01:14:02.149 align:start position:0% +have good theoretical foundations +about<01:13:59.040> them<01:13:59.280> yet<01:14:00.080> and<01:14:00.719> uh<01:14:01.199> potentially<01:14:01.840> if<01:14:02.000> i + +01:14:02.149 --> 01:14:02.159 align:start position:0% +about them yet and uh potentially if i + + +01:14:02.159 --> 01:14:03.350 align:start position:0% +about them yet and uh potentially if i +teach<01:14:02.400> this<01:14:02.640> course<01:14:02.960> in + +01:14:03.350 --> 01:14:03.360 align:start position:0% +teach this course in + + +01:14:03.360 --> 01:14:07.830 align:start position:0% +teach this course in +uh<01:14:04.600> 2025<01:14:06.000> you<01:14:06.159> know<01:14:06.400> maybe<01:14:06.800> we'll<01:14:07.120> include + +01:14:07.830 --> 01:14:07.840 align:start position:0% +uh 2025 you know maybe we'll include + + +01:14:07.840 --> 01:14:11.030 align:start position:0% +uh 2025 you know maybe we'll include +you<01:14:08.000> know<01:14:08.239> some<01:14:09.280> other<01:14:10.080> topics + +01:14:11.030 --> 01:14:11.040 align:start position:0% +you know some other topics + + +01:14:11.040 --> 01:14:14.550 align:start position:0% +you know some other topics +to<01:14:11.360> justify<01:14:12.000> those<01:14:12.320> empirical<01:14:12.960> observations + +01:14:14.550 --> 01:14:14.560 align:start position:0% +to justify those empirical observations + + +01:14:14.560 --> 01:14:16.070 align:start position:0% +to justify those empirical observations +yes<01:14:15.040> you<01:14:15.120> know<01:14:15.280> the<01:14:15.440> answer<01:14:15.760> is<01:14:15.920> like + +01:14:16.070 --> 01:14:16.080 align:start position:0% +yes you know the answer is like + + +01:14:16.080 --> 01:14:18.310 align:start position:0% +yes you know the answer is like +partially<01:14:16.560> yes<01:14:16.880> we'll<01:14:17.440> you<01:14:17.520> know<01:14:17.760> within<01:14:18.080> this + +01:14:18.310 --> 01:14:18.320 align:start position:0% +partially yes we'll you know within this + + +01:14:18.320 --> 01:14:19.270 align:start position:0% +partially yes we'll you know within this +course<01:14:18.560> we'll + +01:14:19.270 --> 01:14:19.280 align:start position:0% +course we'll + + +01:14:19.280 --> 01:14:21.270 align:start position:0% +course we'll +develop<01:14:19.760> some<01:14:20.000> intuitions<01:14:20.560> about<01:14:20.880> some<01:14:21.120> of + +01:14:21.270 --> 01:14:21.280 align:start position:0% +develop some intuitions about some of + + +01:14:21.280 --> 01:14:23.590 align:start position:0% +develop some intuitions about some of +the<01:14:21.600> empirically<01:14:22.320> relevant + +01:14:23.590 --> 01:14:23.600 align:start position:0% +the empirically relevant + + +01:14:23.600 --> 01:14:25.669 align:start position:0% +the empirically relevant +deep<01:14:23.840> learning<01:14:24.239> optimization<01:14:24.960> problems<01:14:25.360> and + +01:14:25.669 --> 01:14:25.679 align:start position:0% +deep learning optimization problems and + + +01:14:25.679 --> 01:14:26.709 align:start position:0% +deep learning optimization problems and +applications + +01:14:26.709 --> 01:14:26.719 align:start position:0% +applications + + +01:14:26.719 --> 01:14:28.790 align:start position:0% +applications +but<01:14:27.280> the<01:14:27.520> picture<01:14:28.000> is<01:14:28.159> not<01:14:28.320> going<01:14:28.480> to<01:14:28.560> be + +01:14:28.790 --> 01:14:28.800 align:start position:0% +but the picture is not going to be + + +01:14:28.800 --> 01:14:29.750 align:start position:0% +but the picture is not going to be +complete + +01:14:29.750 --> 01:14:29.760 align:start position:0% +complete + + +01:14:29.760 --> 01:14:32.390 align:start position:0% +complete +so<01:14:30.480> this<01:14:30.719> is<01:14:30.880> a<01:14:30.960> very<01:14:31.360> active<01:14:32.000> area<01:14:32.239> of + +01:14:32.390 --> 01:14:32.400 align:start position:0% +so this is a very active area of + + +01:14:32.400 --> 01:14:34.709 align:start position:0% +so this is a very active area of +research + +01:14:34.709 --> 01:14:34.719 align:start position:0% +research + + +01:14:34.719 --> 01:14:37.110 align:start position:0% +research +all<01:14:34.800> right<01:14:35.040> so<01:14:35.600> i<01:14:35.840> see<01:14:36.159> some<01:14:36.400> other<01:14:36.640> hands<01:14:36.960> are + +01:14:37.110 --> 01:14:37.120 align:start position:0% +all right so i see some other hands are + + +01:14:37.120 --> 01:14:41.270 align:start position:0% +all right so i see some other hands are +raised<01:14:37.520> go<01:14:37.679> ahead + +01:14:41.270 --> 01:14:41.280 align:start position:0% + + + +01:14:41.280 --> 01:14:42.790 align:start position:0% + +all<01:14:41.360> right<01:14:41.520> for<01:14:41.679> the<01:14:41.840> group<01:14:42.080> project<01:14:42.480> will<01:14:42.640> we + +01:14:42.790 --> 01:14:42.800 align:start position:0% +all right for the group project will we + + +01:14:42.800 --> 01:14:44.470 align:start position:0% +all right for the group project will we +have<01:14:42.960> access<01:14:43.360> to<01:14:43.440> any<01:14:43.679> additional<01:14:44.080> computing + +01:14:44.470 --> 01:14:44.480 align:start position:0% +have access to any additional computing + + +01:14:44.480 --> 01:14:45.830 align:start position:0% +have access to any additional computing +resources + +01:14:45.830 --> 01:14:45.840 align:start position:0% +resources + + +01:14:45.840 --> 01:14:54.830 align:start position:0% +resources +uh<01:14:46.320> yes<01:14:46.800> so<01:14:46.960> i'll<01:14:47.520> make<01:14:47.679> a<01:14:47.840> post<01:14:48.159> about<01:14:48.400> that + +01:14:54.830 --> 01:14:54.840 align:start position:0% + + + +01:14:54.840 --> 01:14:56.950 align:start position:0% + +shortly + +01:14:56.950 --> 01:14:56.960 align:start position:0% +shortly + + +01:14:56.960 --> 01:15:01.030 align:start position:0% +shortly +any<01:14:57.199> other<01:14:57.520> questions + +01:15:01.030 --> 01:15:01.040 align:start position:0% + + + +01:15:01.040 --> 01:15:04.630 align:start position:0% + +all<01:15:01.199> right<01:15:01.920> if<01:15:02.159> not<01:15:03.600> thank<01:15:03.760> you<01:15:03.920> everyone<01:15:04.400> have + +01:15:04.630 --> 01:15:04.640 align:start position:0% +all right if not thank you everyone have + + +01:15:04.640 --> 01:15:13.110 align:start position:0% +all right if not thank you everyone have +a<01:15:04.719> great<01:15:04.960> day<01:15:05.199> and<01:15:05.360> i'll<01:15:05.520> see<01:15:05.679> you<01:15:05.760> on<01:15:08.840> thursday + +01:15:13.110 --> 01:15:13.120 align:start position:0% + + + +01:15:13.120 --> 01:15:15.199 align:start position:0% + +you + diff --git "a/MLwG/Lecture 1 - Deep Learning Foundations\357\274\232 review of basic DL models and optimization solvers [6Nctj1GCFVo].webp" "b/MLwG/Lecture 1 - Deep Learning Foundations\357\274\232 review of basic DL models and optimization solvers [6Nctj1GCFVo].webp" new file mode 100644 index 0000000000000000000000000000000000000000..8c66db999118bb5869985c3daedc028c2d18084d --- /dev/null +++ "b/MLwG/Lecture 1 - Deep Learning Foundations\357\274\232 review of basic DL models and optimization solvers [6Nctj1GCFVo].webp" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad08a576ac570c396b8c88876c3a5388fb9cecf110f6e6f8e9c0f549c295c3f6 +size 31266 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 10.1-Heterogeneous & Knowledge Graph Embedding [Rfkntma6ZUI].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 10.1-Heterogeneous & Knowledge Graph Embedding [Rfkntma6ZUI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..12012eede06782e773454613fcb7280832a9d43f --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 10.1-Heterogeneous & Knowledge Graph Embedding [Rfkntma6ZUI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66842e14843280df5cee69a21a8910453a560b9ada1d2b691736f23c220ed952 +size 151861402 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 10.3 - Knowledge Graph Completion Algorithms [Xm5VrxZYhu4].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 10.3 - Knowledge Graph Completion Algorithms [Xm5VrxZYhu4].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..dd0db58171b2e5ee0ebbf7e38245bfd3a455ef87 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 10.3 - Knowledge Graph Completion Algorithms [Xm5VrxZYhu4].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:315c20fe26a8ca4c4db5cd4c42472d52425974ad9ec781e1e068811397a07613 +size 153381160 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 12.1-Fast Neural Subgraph Matching & Counting [lRCDpfJoMiE].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 12.1-Fast Neural Subgraph Matching & Counting [lRCDpfJoMiE].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..9ddab0779f875093444891fd97a1a51dda60707f --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 12.1-Fast Neural Subgraph Matching & Counting [lRCDpfJoMiE].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37b4ff7bd31af6cf167bf09a873b677c027102fc287e2825140805682d955093 +size 150351013 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 13.1 - Community Detection in Networks [KXi4ha79o3s].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 13.1 - Community Detection in Networks [KXi4ha79o3s].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..2863db4712cc81ae8584a5c3c009922e4767c8d3 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 13.1 - Community Detection in Networks [KXi4ha79o3s].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e742d5683a30398778e7476736c3b7f283a10341ae1736d77607c49fe99a0ee5 +size 114513199 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 13.4 - Detecting Overlapping Communities [SHcVHrsgj8w].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 13.4 - Detecting Overlapping Communities [SHcVHrsgj8w].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..83372748c3e021ff8cbe916245833d951b169e62 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 13.4 - Detecting Overlapping Communities [SHcVHrsgj8w].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2385ea379b1399c9eb56f1afe001d9c74b0c53935700a2dccfef39102171ff51 +size 103692071 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.1 - Deep Generative Models for Graphs [IMpkHvQ0LA4].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.1 - Deep Generative Models for Graphs [IMpkHvQ0LA4].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..69bdaa41b5ff3bd3d8182a6e432f5fa1967522ec --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.1 - Deep Generative Models for Graphs [IMpkHvQ0LA4].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7241c0c2b7d067b2bab9d3d8e1e2cbd7ed3491e87ee517e0d7e6a5b61cd1c45 +size 72956879 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.2 - Graph RNN\357\274\232 Generating Realistic Graphs [xGDUYQGvRac].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.2 - Graph RNN\357\274\232 Generating Realistic Graphs [xGDUYQGvRac].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..9835b7403cd736de5ccdc26e256a9a4dd6d2054e --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.2 - Graph RNN\357\274\232 Generating Realistic Graphs [xGDUYQGvRac].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4626b36f3f0cc19ddcc44e653e89c73ab2399b6114208a63656d07f8fd56c01a +size 114785325 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.3 - Scaling Up & Evaluating Graph Gen [kDu5YY9d1TM].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.3 - Scaling Up & Evaluating Graph Gen [kDu5YY9d1TM].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..5924eff20f130c167fcbe5b498f9fd0a2afe74fd --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.3 - Scaling Up & Evaluating Graph Gen [kDu5YY9d1TM].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0728e5553fc45d0e590bb7fd6d476d5588e3818d6c5896c2d66a5e96774086b7 +size 75247646 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.4 - Applications of Deep Graph Generation [SdBk8fMmwUU].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.4 - Applications of Deep Graph Generation [SdBk8fMmwUU].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..b605d45330a435ee8c6fda70bd7c11e2c618f146 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 15.4 - Applications of Deep Graph Generation [SdBk8fMmwUU].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af8bf586f069c94fe6ea194d157f7653340d748aaf853c29705671a51187c34b +size 68769864 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.1 - Limitations of Graph Neural Networks [yT5PziOXxQg].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.1 - Limitations of Graph Neural Networks [yT5PziOXxQg].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..e9aa35c88ea38ab0ca0788bd9c643d993877fdfd --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.1 - Limitations of Graph Neural Networks [yT5PziOXxQg].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a51e843f958a8685494974f015d12e437febc9cd473ee1c6a9d61a605322636d +size 49348610 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.2 - Position-Aware Graph Neural Networks [6ZFvToZUjGA].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.2 - Position-Aware Graph Neural Networks [6ZFvToZUjGA].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..feaa9fa20df288c04e587807e04f5df81f97d06a --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.2 - Position-Aware Graph Neural Networks [6ZFvToZUjGA].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:254438739ba30cd1463a5fd15361f64cac2622ba78ccfe62fed3fcad94bbc8ab +size 64777958 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.3 - Identity-Aware Graph Neural Networks [SJqWQGh__N8].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.3 - Identity-Aware Graph Neural Networks [SJqWQGh__N8].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..36ed801e16dd7ef598e586c9617492bb353d6af5 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.3 - Identity-Aware Graph Neural Networks [SJqWQGh__N8].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7a0dad09650c60ab55ca339afce3fe8d03e7c9c0c3e870c1d7d778db69ba015 +size 81849013 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.4 - Robustness of Graph Neural Networks [uBmtTL3EitI].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.4 - Robustness of Graph Neural Networks [uBmtTL3EitI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..b420cf04f8891a0519272234afb6e2afe7ec758e --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 16.4 - Robustness of Graph Neural Networks [uBmtTL3EitI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ace7bf023891dfac8e47b0ba7cb2844852dbcfa65527305ed89386166c81dae8 +size 109622155 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 17.1 - Scaling up Graph Neural Networks [2nPCw3yHlnI].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 17.1 - Scaling up Graph Neural Networks [2nPCw3yHlnI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..341f8ffa1c565a262fdc79e566ae8f2ff8b54b13 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 17.1 - Scaling up Graph Neural Networks [2nPCw3yHlnI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fddb2ba53b96a358587c64595905d16961eacb10ebe0ba01a276bf5247018c4 +size 82768459 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 19.1 - Pre-Training Graph Neural Networks [JDW82csukhE].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 19.1 - Pre-Training Graph Neural Networks [JDW82csukhE].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..192370707e8dccc1a46adc0e9f92909d21c112bf --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 19.1 - Pre-Training Graph Neural Networks [JDW82csukhE].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a2d2a9c6498e9a5b3e9b7f7ffed8d7f95c964ea97ec69e37ee69a98b1ec51d6 +size 53138105 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 19.3 - Design Space of Graph Neural Networks [8OhnwzT9ypg].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 19.3 - Design Space of Graph Neural Networks [8OhnwzT9ypg].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..ce37c915c1e15c7d93c02c84052d247672d8f9e9 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 19.3 - Design Space of Graph Neural Networks [8OhnwzT9ypg].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:203ef0a6f0d319bd989394914729ea7af86d7a9d04993ccd4eba7684fefb0519 +size 63898903 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.1 - Traditional Feature-based Methods\357\274\232 Node [3IS7UhNMQ3U].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.1 - Traditional Feature-based Methods\357\274\232 Node [3IS7UhNMQ3U].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..00974368c53b51584f268198687a4b0eca6d024c --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.1 - Traditional Feature-based Methods\357\274\232 Node [3IS7UhNMQ3U].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f86645e824399798a4f09169417d09f3582c63d89e14bce812411ccf48a5f301 +size 158803212 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.2 - Traditional Feature-based Methods\357\274\232 Link [4dVwlE9jYxY].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.2 - Traditional Feature-based Methods\357\274\232 Link [4dVwlE9jYxY].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..36f9d59410a956950719357c86c954df01d0d69e --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.2 - Traditional Feature-based Methods\357\274\232 Link [4dVwlE9jYxY].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c69b21566e67e8d5cedc3e00669a5d3f85a73691216cfd91f4fc097dda88ebaa +size 106426191 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.3 - Traditional Feature-based Methods\357\274\232 Graph [buzsHTa4Hgs].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.3 - Traditional Feature-based Methods\357\274\232 Graph [buzsHTa4Hgs].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..5f7f96c9e04ab772536a50cf7f404990bcd6e36d --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 2.3 - Traditional Feature-based Methods\357\274\232 Graph [buzsHTa4Hgs].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bc4c2a82a43e751388fe3f9ea44ef8b04e48c446346d3cba6880bf6339f9885 +size 116920863 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 3.2-Random Walk Approaches for Node Embeddings [Xv0wRy66Big].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 3.2-Random Walk Approaches for Node Embeddings [Xv0wRy66Big].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..23954c508c23a7de54d52c1ed0e674ec3a376e74 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 3.2-Random Walk Approaches for Node Embeddings [Xv0wRy66Big].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97de385636908293bfe8b314a44bf3bda83cd354730e5a9cea620fcaf2f82eca +size 178225145 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 4.4 - Matrix Factorization and Node Embeddings [r12qJZZVtqc].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 4.4 - Matrix Factorization and Node Embeddings [r12qJZZVtqc].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..6cd8091aec6156fad33a460c9f3327b3aa6ec133 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 4.4 - Matrix Factorization and Node Embeddings [r12qJZZVtqc].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cce3ff4f2ca0bb6905968f6376c4ac54908ec16375138ae81591e08b8ed89d6b +size 81040656 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 5.1 - Message passing and Node Classification [6g9vtxUmfwM].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 5.1 - Message passing and Node Classification [6g9vtxUmfwM].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..99603ce2a7fb9225f1106b2dd9a29f12b34dc1e5 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 5.1 - Message passing and Node Classification [6g9vtxUmfwM].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32ebe2a7ddccc682d9c8eff7a5eb02f3840cf054e21cb3dd2e214eef570aa65a +size 110526219 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 5.2 - Relational and Iterative Classification [QUO-HQ44EDc].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 5.2 - Relational and Iterative Classification [QUO-HQ44EDc].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..e4899d32305487260bc0b5023886fcd4140af7ef --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 5.2 - Relational and Iterative Classification [QUO-HQ44EDc].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:192c50ed2f6e7a4824b96fed64336473da03856905ac2fdf7fcd8fbe9a5e9308 +size 177870539 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 6.1 - Introduction to Graph Neural Networks [F3PgltDzllc].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 6.1 - Introduction to Graph Neural Networks [F3PgltDzllc].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..42d847be50f2d16742173eb6607e052d2bdc4913 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 6.1 - Introduction to Graph Neural Networks [F3PgltDzllc].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5985c2876963c628a3c765909cc5e37d9e789070a14d9aeddaa1954f7b56da17 +size 58626160 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 9.1 - How Expressive are Graph Neural Networks [5vMEgYbka0A].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 9.1 - How Expressive are Graph Neural Networks [5vMEgYbka0A].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..bf64c1441ae5381d922298d9c43ba455994348ae --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 9.1 - How Expressive are Graph Neural Networks [5vMEgYbka0A].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:141bc9579366c65640dd5b9a0116bf1df40ba72b93e1cdd28f006a79f29e10c5 +size 104897360 diff --git "a/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 9.2 - Designing the Most Powerful GNNs [B5y47gWt3co].mp4" "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 9.2 - Designing the Most Powerful GNNs [B5y47gWt3co].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..931b4d2bb1090aba54ae6e1c2c26d53a01b18d93 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 ML with Graphs \357\275\234 2021 \357\275\234 Lecture 9.2 - Designing the Most Powerful GNNs [B5y47gWt3co].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cea5b9330d4640d39f404bc500fb1b7031226935dd06c870342fd3040d5e872 +size 126816811 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.1 - Why Graphs [JAB_plj2rbA].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.1 - Why Graphs [JAB_plj2rbA].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..f7e6d3ac433895aca58a4cefdd50d95755311822 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.1 - Why Graphs [JAB_plj2rbA].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f01fde82be95cb3aa8f1ff2bad84541f5bbc6aa7dd54b52ccaf29c39a2441bd2 +size 76482020 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.2 - Applications of Graph ML [aBHC6xzx9YI].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.2 - Applications of Graph ML [aBHC6xzx9YI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..fdeb9833e35a4c04815b03c92a593bc6715bd149 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.2 - Applications of Graph ML [aBHC6xzx9YI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5774fa5170b54b9db65ce300ffa7fe787636b540a918e4bf515cfce2dca81e42 +size 142215741 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.3 - Choice of Graph Representation\342\200\213 [P-m1Qv6-8cI].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.3 - Choice of Graph Representation\342\200\213 [P-m1Qv6-8cI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..40951633263f51fe35941310bcc50ce3f01826cf --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 1.3 - Choice of Graph Representation\342\200\213 [P-m1Qv6-8cI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ea8b0f164a3eb4b770140476bf868e77e4dec56e6a8485c6b56dd536b17e829 +size 97544814 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 10.2 - Knowledge Graph Completion [xop5tC9T5xM].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 10.2 - Knowledge Graph Completion [xop5tC9T5xM].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..f4d948421bbe087f858263fc491e9ce6fcce6daf --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 10.2 - Knowledge Graph Completion [xop5tC9T5xM].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62379a950975ae0af8834f1ae5aa57d4363e39521a2c4856b4abb94b0074c591 +size 32302532 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.1 - Reasoning in Knowledge Graphs [X9yl0pTP9fY].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.1 - Reasoning in Knowledge Graphs [X9yl0pTP9fY].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..9705171b1cf56b0ddec8d068d9cbb30dcd5373a5 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.1 - Reasoning in Knowledge Graphs [X9yl0pTP9fY].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92775d4180b43c944cc5c9e2994aa8b3bb42b04e4117ad1c5827c42c92345861 +size 72692441 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.2 - Answering Predictive Queries [qaRIBNE-4Ho].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.2 - Answering Predictive Queries [qaRIBNE-4Ho].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..4b2146dc31ea1c82c965e464499b548fef3bd2c7 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.2 - Answering Predictive Queries [qaRIBNE-4Ho].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9d063230583b5fbac7ded253021956d669d7944830c5affca20eb8508ba2829 +size 56461150 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.3 - Query2box\357\274\232 Reasoning over KGs [Nt66M2OsbCw].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.3 - Query2box\357\274\232 Reasoning over KGs [Nt66M2OsbCw].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..bef29d80d4683f424f94624b7281e7dd8c7d6896 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 11.3 - Query2box\357\274\232 Reasoning over KGs [Nt66M2OsbCw].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d00f1640be91519dca99e53795b1b833cd82272fbca5f7726822f320b232d952 +size 152732796 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 12.2 - Neural Subgraph Matching [4Ia5QzQ_QNI].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 12.2 - Neural Subgraph Matching [4Ia5QzQ_QNI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..7ac03c9689955a868c8c89f6ebf94537f0c95ca7 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 12.2 - Neural Subgraph Matching [4Ia5QzQ_QNI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:345b0c27bb70bc1695df1f158a3f605fab231023678cf76b4ed56c88a5e438e4 +size 105636886 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 12.3 - Finding Frequent Subgraphs [kUv4gY5e0hg].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 12.3 - Finding Frequent Subgraphs [kUv4gY5e0hg].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..013179c18f9b58cb0255bf8e73a9aa7945fe7c85 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 12.3 - Finding Frequent Subgraphs [kUv4gY5e0hg].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c364d227459ddba35cddff42d1648d0caad65128808deabe1477ca2546474b5d +size 118479164 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 13.2 - Network Communities [mJQrtXZT5pw].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 13.2 - Network Communities [mJQrtXZT5pw].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..4f58aaac7f90cb3715543bff5efa86a1c72e77af --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 13.2 - Network Communities [mJQrtXZT5pw].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbd030ba3ec31bb3df63e0495b8a51d6e8642d18ddd5bc58905fc4e9213a4b56 +size 78507429 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 13.3 - Louvain Algorithm [0zuiLBOIcsw].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 13.3 - Louvain Algorithm [0zuiLBOIcsw].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..806c69b474a85e3a0e855c4d76cd0ce487993088 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 13.3 - Louvain Algorithm [0zuiLBOIcsw].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d4e4d5ad5016d681b54cbdbce37bbf26e920aec74b07557375549515ff3e6d3 +size 65262163 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.1 - Generative Models for Graphs [OhjXpYIo9h4].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.1 - Generative Models for Graphs [OhjXpYIo9h4].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..fc05c19e60b86b4bc52906b693eda34a1880acb0 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.1 - Generative Models for Graphs [OhjXpYIo9h4].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:173e41e984e11bfdfb2c4ecac2eda32d775a589e6fa9564271da6120a5650231 +size 88133299 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.2 - Erdos Renyi Random Graphs [VC43S6Thwg0].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.2 - Erdos Renyi Random Graphs [VC43S6Thwg0].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..89ddc6c955d56ef394f1a9dab39ef2a67292993d --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.2 - Erdos Renyi Random Graphs [VC43S6Thwg0].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cb13dabe0a0b37cfcce198c8f34700173eeff21a40954a5a965b97ffdf32b46 +size 101064813 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.4 - Kronecker Graph Model [Xnpt8US31cQ].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.4 - Kronecker Graph Model [Xnpt8US31cQ].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..6eaa75776eaf5ccd6010e828346d789cd876cb8d --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 14.4 - Kronecker Graph Model [Xnpt8US31cQ].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85d6ad3c4baa98477559b90c412e5a672864ee1f0d32e5f106f9f2fae2da212d +size 75205906 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.2 - GraphSAGE Neighbor Sampling [LLUxwHc7O4A].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.2 - GraphSAGE Neighbor Sampling [LLUxwHc7O4A].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..0cb9d2c256ca7aa3bc13b6c04ab28044531ec55a --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.2 - GraphSAGE Neighbor Sampling [LLUxwHc7O4A].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e2611da7f6f1e823d5e7c63e58045696646b2e839f200a35d96194a7c6f6750 +size 93175880 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.3 - Cluster GCN\357\274\232 Scaling up GNNs [RJkR8Ig6dXI].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.3 - Cluster GCN\357\274\232 Scaling up GNNs [RJkR8Ig6dXI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..c8aa191074cb4b95c3e748a40bd653684074d156 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.3 - Cluster GCN\357\274\232 Scaling up GNNs [RJkR8Ig6dXI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa6d4af0c7fa779fc00cb2b2842b8b908253b833ca7c66dbd464bc7ac5d6b704 +size 98641528 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.4 - Scaling up by Simplifying GNNs [iTRW9Gh7yKI].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.4 - Scaling up by Simplifying GNNs [iTRW9Gh7yKI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..2d1eb289fb2dc1985812c17c2ad057891501474d --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 17.4 - Scaling up by Simplifying GNNs [iTRW9Gh7yKI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a24d208c922fc2f8290c04d118d5574614ce9a14b3765452810a5a2e7863bd7a +size 79363594 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 18 - GNNs in Computational Biology [_hy9AgZXhbQ].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 18 - GNNs in Computational Biology [_hy9AgZXhbQ].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..b61178c06c3e6e944394ebecb28b06a48b84624a --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 18 - GNNs in Computational Biology [_hy9AgZXhbQ].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:211179fec7c36aeae08d0787d92f26bc671ee2c965bef44776ca550c705faea0 +size 310755926 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 19.2 - Hyperbolic Graph Embeddings [m2zoddmgvd0].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 19.2 - Hyperbolic Graph Embeddings [m2zoddmgvd0].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..88297886ebd197e241a261e10f5804dfb65923f4 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 19.2 - Hyperbolic Graph Embeddings [m2zoddmgvd0].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d905043689a62baf3251d302ae65b76da3c24cab80b9c5aa97132e6c5c420c8 +size 74410850 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 3.1 - Node Embeddings [rMq21iY61SE].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 3.1 - Node Embeddings [rMq21iY61SE].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..3bf2ee17f6a32e7a70aec9dc24d69655d06e755c --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 3.1 - Node Embeddings [rMq21iY61SE].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d615193f7ecbd4c23e9d964147a16981eb32eee61a2d72540f1acfdd3a5466ff +size 62221081 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 3.3 - Embedding Entire Graphs [eliMLfJeu7A].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 3.3 - Embedding Entire Graphs [eliMLfJeu7A].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..446e2209789d84ec1160aeb4a1b5b1ecc279442c --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 3.3 - Embedding Entire Graphs [eliMLfJeu7A].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:737d1fb59ee3f61c1758c861b482b905ebf6a67e33fdc685b5e578c830cf9414 +size 131483181 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.1 - PageRank [TU0ankRcHmo].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.1 - PageRank [TU0ankRcHmo].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..0b95555017c7d6bf442543861319005fbf315f6c --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.1 - PageRank [TU0ankRcHmo].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb91b550620ca02dba04d1d46214c4382c042cdaf3b379f91735e5af6d0e4e3e +size 139754905 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.2 - PageRank\357\274\232 How to Solve\357\274\237 [rK2ZBmQHVVs].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.2 - PageRank\357\274\232 How to Solve\357\274\237 [rK2ZBmQHVVs].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..f416c95cf2c2460d15b5635512edb171faeed5b8 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.2 - PageRank\357\274\232 How to Solve\357\274\237 [rK2ZBmQHVVs].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd95d003658e0f035d3870d49113fd9a5c606664a977b0306ba97b8a204bdaae +size 94587994 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.3 - Random Walk with Restarts [HbzQzUaJ_9I].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.3 - Random Walk with Restarts [HbzQzUaJ_9I].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..59ffec3f5d811cf6fa9b2772894867c112e0417d --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 4.3 - Random Walk with Restarts [HbzQzUaJ_9I].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c136c357c277b90487cbdc53e5e4709f667820138e18061bc517087166494bac +size 69787152 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 5.3 - Collective Classification [kh3I_UTtUOo].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 5.3 - Collective Classification [kh3I_UTtUOo].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..2367016e42755ef80663db9bfa5709928e6ff4b4 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 5.3 - Collective Classification [kh3I_UTtUOo].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de23c557a4d47866381d4012156b292be8302ca67452869a59a13b8b4e7a5436 +size 130827401 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 6.2 - Basics of Deep Learning [tutlI9YzJ2g].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 6.2 - Basics of Deep Learning [tutlI9YzJ2g].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..088667cfa43e73579f875bd3f5ebd37586d4b7f7 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 6.2 - Basics of Deep Learning [tutlI9YzJ2g].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ee919e147c18184c09e574f9b0aff4d39e51f55dbf10dcd6abd81cd2692c709 +size 118351929 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 6.3 - Deep Learning for Graphs [MH4yvtgAR-4].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 6.3 - Deep Learning for Graphs [MH4yvtgAR-4].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..e8fa1dd1b142f7acd786743ae74427ed1f3b385e --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 6.3 - Deep Learning for Graphs [MH4yvtgAR-4].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfea586483fc8fa4bb9815a9d04ecaa7ad49b85fe6336a8ba79dd8c4060e3458 +size 134361858 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.1 - A general Perspective on GNNs [RU9uTa_-ZOw].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.1 - A general Perspective on GNNs [RU9uTa_-ZOw].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..3769fdf52dc90e0aeca11579db7f9329888b2acc --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.1 - A general Perspective on GNNs [RU9uTa_-ZOw].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d312489c6036b8b73760a3182b55411655b12065ce5dcb8c1d4de1adcda517b4 +size 24338666 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.2 - A Single Layer of a GNN [247Mkqj_wRM].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.2 - A Single Layer of a GNN [247Mkqj_wRM].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..cf86d89a0c4ebb5b3cdbaca91f321d33e749cf32 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.2 - A Single Layer of a GNN [247Mkqj_wRM].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6c5d17b211c63c26cb6ce11c963bc93234ddc891f945c12d4be6a4c19c86f51 +size 156947212 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.3 - Stacking layers of a GNN [ew1cnUjRgl4].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.3 - Stacking layers of a GNN [ew1cnUjRgl4].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..42d60b16fb2f61e43ce28c9234f61fe0e4867285 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 7.3 - Stacking layers of a GNN [ew1cnUjRgl4].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:027e7046dc2fc6d0e24ce98c98d19debfbdce01a2156eb76623f7477240863d1 +size 91003175 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.1 - Graph Augmentation for GNNs [1A6VoEkQnhQ].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.1 - Graph Augmentation for GNNs [1A6VoEkQnhQ].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..e1f12104492787d812c6f159e9882f7bd7da17e8 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.1 - Graph Augmentation for GNNs [1A6VoEkQnhQ].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:960db5512012f899bd1274ef5af4109ecd5322b9b900c72af75fd2651a8e7bd4 +size 117442241 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.2 - Training Graph Neural Networks [eXIIH8YVxKI].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.2 - Training Graph Neural Networks [eXIIH8YVxKI].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..c6f6a97944c52cf413b026dc7ae27a757bcb58b6 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.2 - Training Graph Neural Networks [eXIIH8YVxKI].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e9b72a00a6843c55087d67d4a61815b773850b08eedfb640c1c8fc0b1ca2599 +size 167292899 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.3 - Setting up GNN Prediction Tasks [ewEW_EMzRuo].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.3 - Setting up GNN Prediction Tasks [ewEW_EMzRuo].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..c0e207edaf51f6fd7a66749b0d1e50792b23b343 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning with Graphs \357\275\234 2021 \357\275\234 Lecture 8.3 - Setting up GNN Prediction Tasks [ewEW_EMzRuo].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c3e1aa92ca2dad5b4d566b1c9fe35ccbece749d1b0fb417d32f83dbc578bd0a +size 81307347 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Geometric Graph Learning, Minkai Xu [hDnlfTx4dZs].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Geometric Graph Learning, Minkai Xu [hDnlfTx4dZs].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..f71ebfa2e5da5b26bb35245d9eedd7a977ad2d78 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Geometric Graph Learning, Minkai Xu [hDnlfTx4dZs].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d36c4e4ba655f001216982436aa593b7fde06e89e0cb50e676117741e407025 +size 360560449 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Graph Neural Networks [ZfK4FDk9uy8].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Graph Neural Networks [ZfK4FDk9uy8].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..b74e9450ff79d99f9b314d3998dbaec1fc076134 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Graph Neural Networks [ZfK4FDk9uy8].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a48772419c8eec303855d1b064861c0e845a919ef3f38355a077362236fac6c5 +size 366480013 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Label Propagation on Graphs [1E8fie5KSl0].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Label Propagation on Graphs [1E8fie5KSl0].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..aeb1ae161fac2e3ab44e0bf57cf36ad83c8c0b81 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Label Propagation on Graphs [1E8fie5KSl0].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c918f284a7ae2cbcacc6c0f18e68d6eef5f19fb14f20659094cfddd05bef8f3a +size 376931775 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Machine Learning with Heterogeneous Graphs [uvrlKxj8HVU].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Machine Learning with Heterogeneous Graphs [uvrlKxj8HVU].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..1d22bd8dc507b1ce16fa131b8c927a6fe59491d1 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Machine Learning with Heterogeneous Graphs [uvrlKxj8HVU].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b90dd17fa3439226c016447c39063ea2d31a7a1880528ae3abf0917db53376aa +size 383902634 diff --git "a/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Trustworthy Graph AI, Rex Ying [O6eTvxWy07A].mp4" "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Trustworthy Graph AI, Rex Ying [O6eTvxWy07A].mp4" new file mode 100644 index 0000000000000000000000000000000000000000..79083be141468149302301930af3b3512548c630 --- /dev/null +++ "b/MLwG/Stanford CS224W\357\274\232 Machine Learning w\342\247\270 Graphs I 2023 I Trustworthy Graph AI, Rex Ying [O6eTvxWy07A].mp4" @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41c2728b426c681444aed77a6ca649f56e8473098caa58a6262da4841b53cd99 +size 245381508