WEBVTT 00:00.360 --> 00:03.120 -: Hello and welcome back to the course on Deep Learning. 00:03.120 --> 00:06.060 So, we've learned quite a lot in this section of the course, 00:06.060 --> 00:08.550 let's summarize what we've talked about. 00:08.550 --> 00:10.080 All right, so here we go. 00:10.080 --> 00:11.790 We started with an input image 00:11.790 --> 00:15.480 to which we applied multiple different feature detectors, 00:15.480 --> 00:19.080 or also called filters, to create these feature maps, 00:19.080 --> 00:21.570 and this comprises our convolution layer. 00:21.570 --> 00:24.720 Then on top of that conventional layer we applied the RELU 00:24.720 --> 00:29.010 or Rectified Linear Unit to remove any linearity 00:29.010 --> 00:32.010 or increased non-linearity in our images. 00:32.010 --> 00:36.990 Then we applied a pooling layer to our convolution layer, 00:36.990 --> 00:40.260 so from every single feature map 00:40.260 --> 00:42.840 we created a pooled feature map. 00:42.840 --> 00:45.870 And basically the pooling layer has lots of advantages, 00:45.870 --> 00:47.790 the main purpose of the pooling layer 00:47.790 --> 00:50.940 is to make sure that we have 00:50.940 --> 00:54.720 a special in variance in our images, 00:54.720 --> 00:56.760 so basically if something tilts or twists 00:56.760 --> 01:01.230 or is a bit different to the ideal scenario, 01:01.230 --> 01:03.030 then we can still pick up that feature. 01:03.030 --> 01:07.020 Plus, pooling significantly reduces the size of our images 01:07.020 --> 01:10.890 and also pooling helps with avoiding 01:10.890 --> 01:13.560 any kind of overfitting of our data 01:13.560 --> 01:15.150 or of our model to the data, 01:15.150 --> 01:18.420 because it just simply gets rid of a lot of that data. 01:18.420 --> 01:22.140 But at the same time pooling preserves the main features 01:22.140 --> 01:24.330 that we were after, just because the way it's structured 01:24.330 --> 01:26.940 and the pooling we used was max pooling. 01:26.940 --> 01:29.520 Then we flattened all of the pooled images 01:29.520 --> 01:34.520 into one long a vector or column of all of these values 01:35.520 --> 01:38.310 and we input that into an artificial neural network, 01:38.310 --> 01:40.110 and that was step three flattening 01:40.110 --> 01:42.720 and step four is the fully connected 01:42.720 --> 01:44.040 artificial neural network 01:44.040 --> 01:46.920 where all of these features are processed 01:46.920 --> 01:47.753 through the network. 01:47.753 --> 01:50.220 And then we have this final layer, 01:50.220 --> 01:53.880 final fully connected layer, which performs the voting 01:53.880 --> 01:56.040 towards the classes that we're after. 01:56.040 --> 01:58.020 And then all of this is trained 01:58.020 --> 02:02.520 through a forward propagation and back propagation process, 02:02.520 --> 02:05.220 and lots of lots of iterations and (indistinct) 02:05.220 --> 02:09.720 and in the end we have a very well defined neural network. 02:09.720 --> 02:11.700 And another important thing is 02:11.700 --> 02:13.020 not only the weights are trained 02:13.020 --> 02:14.310 in artificial neural network part 02:14.310 --> 02:17.460 but also the feature detectors are trained 02:17.460 --> 02:21.960 and adjusted in that same gradient decent process, 02:21.960 --> 02:23.940 and that allows us to come up with the best feature maps. 02:23.940 --> 02:26.460 And in the end we get a fully trained 02:26.460 --> 02:27.690 convolutional neural network 02:27.690 --> 02:31.740 which can recognize images and classify them. 02:31.740 --> 02:32.573 So, there we go, 02:32.573 --> 02:35.700 that's how convolutional neural networks work. 02:35.700 --> 02:38.970 And now you should be totally comfortable with this concept 02:38.970 --> 02:42.300 and ready to proceed to the practical applications. 02:42.300 --> 02:44.610 If you'd like to do some additional reading, 02:44.610 --> 02:49.170 then there's a great blog by Adit Deshpande 02:49.170 --> 02:53.370 from 2016, you can see the link over there at the bottom. 02:53.370 --> 02:55.950 So, the blog is called "The Nine Deep Learning Papers 02:55.950 --> 02:59.310 You Need to Know About Understanding CNN's Part Three". 02:59.310 --> 03:01.680 And this blog actually gives you a short overview 03:01.680 --> 03:05.356 of nine different CNNs that have been created by people like 03:05.356 --> 03:07.920 Jan (indistinct) and others, 03:07.920 --> 03:10.590 which you can then go ahead and study further. 03:10.590 --> 03:13.650 So, there will be a lot of new things 03:13.650 --> 03:15.630 that will be totally new to you 03:15.630 --> 03:18.540 and that you will have to get your head around, 03:18.540 --> 03:20.610 but just keep this blog in mind 03:20.610 --> 03:22.530 or these nine papers in mind, 03:22.530 --> 03:25.170 and even if you're not ready to go through them right now 03:25.170 --> 03:26.850 maybe after the practical tutorials, 03:26.850 --> 03:30.060 maybe after you do some additional training 03:30.060 --> 03:31.530 in the space of deep learning, 03:31.530 --> 03:33.900 slowly you can then reference these works 03:33.900 --> 03:37.680 and, ideally, I think you will get a lot of value 03:37.680 --> 03:39.930 through looking through other people's neural networks 03:39.930 --> 03:43.230 and how they structured their convolution nets; 03:43.230 --> 03:46.020 and that'll help you understand what are the best practices 03:46.020 --> 03:48.420 and why people did certain things in a certain way. 03:48.420 --> 03:49.470 And that will help you 03:49.470 --> 03:51.900 with your architecture of neural networks, 03:51.900 --> 03:55.060 Because neural networks and convolutional neural networks 03:56.820 --> 03:57.990 are not an exception 03:57.990 --> 04:01.620 they are like an architecture challenge. 04:01.620 --> 04:05.400 You have to come up with a idea and then structure it, 04:05.400 --> 04:06.750 and then adjust it and tweak it 04:06.750 --> 04:08.670 to get the best possible design 04:08.670 --> 04:11.760 and the best possible and optimal performance. 04:11.760 --> 04:13.440 So, there we go, that's us for today. 04:13.440 --> 04:16.140 I hope you enjoyed today's tutorial and this whole section, 04:16.140 --> 04:17.670 and I look forward to seeing you next time. 04:17.670 --> 04:19.623 Until then, enjoy Deep Learning.