| 1 | |
| 00:00:00,930 --> 00:00:01,270 | |
| OK. | |
| 2 | |
| 00:00:01,290 --> 00:00:06,070 | |
| So we've just discussed all the layers that comprise of CNN. | |
| 3 | |
| 00:00:06,090 --> 00:00:14,950 | |
| So now let's talk about how we actually train our CNN's CNN's up basically a type of neural net and | |
| 4 | |
| 00:00:14,950 --> 00:00:21,760 | |
| they lend themselves perfectly to image classification and basically a spatial image features but because | |
| 5 | |
| 00:00:21,760 --> 00:00:29,260 | |
| of the max pulling and multiple filters it is very basically somewhat scale and distortion invariant. | |
| 6 | |
| 00:00:29,320 --> 00:00:36,790 | |
| So this is an illustration of the features feature maps that a CNN lens is taken from this link here. | |
| 7 | |
| 00:00:36,910 --> 00:00:38,290 | |
| Check it out. | |
| 8 | |
| 00:00:38,290 --> 00:00:43,820 | |
| So essentially you can see the beginning convolutional what it really is is basically this is input | |
| 9 | |
| 00:00:44,200 --> 00:00:47,270 | |
| and this is close to the output or the output itself. | |
| 10 | |
| 00:00:47,350 --> 00:00:51,260 | |
| So you can see the early filters Lynn basically simple. | |
| 11 | |
| 00:00:51,270 --> 00:00:54,070 | |
| Basically edge detectors are pattern recognizers. | |
| 12 | |
| 00:00:54,340 --> 00:00:59,350 | |
| However the mid-level features are a bit more abstract and a bit more complicated. | |
| 13 | |
| 00:00:59,380 --> 00:01:03,040 | |
| You see these streetwear lines here something that looks like a eyeball. | |
| 14 | |
| 00:01:03,040 --> 00:01:04,280 | |
| It might even show where it is. | |
| 15 | |
| 00:01:04,330 --> 00:01:06,670 | |
| Could be a wheel actually. | |
| 16 | |
| 00:01:06,790 --> 00:01:12,520 | |
| And then as we go for the next layer of convolutional filters remember I said you can have multiple | |
| 17 | |
| 00:01:12,520 --> 00:01:17,930 | |
| layers of them in multiple sequences of convolutional is what this is exactly an illustration of that. | |
| 18 | |
| 00:01:17,980 --> 00:01:20,580 | |
| And as you can see the patents get much more detailed. | |
| 19 | |
| 00:01:20,950 --> 00:01:24,950 | |
| And basically this is how neural nets are CNN's I should say. | |
| 20 | |
| 00:01:25,180 --> 00:01:30,610 | |
| This is how the filters actually live and this is what they pick up in images later on in about two | |
| 21 | |
| 00:01:30,610 --> 00:01:31,960 | |
| or three chapters from now. | |
| 22 | |
| 00:01:32,110 --> 00:01:39,780 | |
| We're going to do some visualizations of it when so basically just a quick review of CNN's. | |
| 23 | |
| 00:01:39,790 --> 00:01:47,610 | |
| Again an old is convolution Plus real than here and pooling than we can do convolution and again then | |
| 24 | |
| 00:01:47,630 --> 00:01:55,650 | |
| more pooling and then pass everything to fully connectedly which outputs a class probabilities. | |
| 25 | |
| 00:01:55,650 --> 00:01:58,250 | |
| So how do we train CNN's. | |
| 26 | |
| 00:01:58,290 --> 00:02:02,740 | |
| So just like in that streaming CNN is essentially the same thing. | |
| 27 | |
| 00:02:02,910 --> 00:02:03,640 | |
| OK. | |
| 28 | |
| 00:02:03,840 --> 00:02:09,750 | |
| So we have a random weird initialization of the convolutional gunnels that's all all the Reynolds we | |
| 29 | |
| 00:02:09,750 --> 00:02:13,500 | |
| described before they basically start with random values. | |
| 30 | |
| 00:02:13,500 --> 00:02:19,110 | |
| Then we follow propagate an image one sample image trail and network that goes Judaists all these layers | |
| 31 | |
| 00:02:19,110 --> 00:02:24,090 | |
| here and we get the arrow arrow being basically a class probabilities. | |
| 32 | |
| 00:02:24,090 --> 00:02:29,630 | |
| So let's say we were looking for an output of point to point four point four when the tubo probabilities | |
| 33 | |
| 00:02:29,640 --> 00:02:36,750 | |
| with this basically unclasp probabilities you wanted to be as close to this as possible way it's very | |
| 34 | |
| 00:02:36,750 --> 00:02:41,060 | |
| sure it's in one category and very little probability any other categories. | |
| 35 | |
| 00:02:42,360 --> 00:02:49,270 | |
| So then we similarly we apply a last function positron optimizer and use back propagation. | |
| 36 | |
| 00:02:49,400 --> 00:02:50,220 | |
| Ditto gradients. | |
| 37 | |
| 00:02:50,280 --> 00:02:52,890 | |
| I agree and dissent and we keep doing this. | |
| 38 | |
| 00:02:52,890 --> 00:02:58,680 | |
| We keep propagating all the images true and that work till we complete an epoch and then keep completing | |
| 39 | |
| 00:02:58,860 --> 00:03:02,050 | |
| ebox until accuracy and loss are satisfactory. | |
| 40 | |
| 00:03:03,380 --> 00:03:06,530 | |
| So now let's take a look at how we actually design and CNN's. | |