1 00:00:00,450 --> 00:00:01,680 Hi and welcome back. 2 00:00:02,040 --> 00:00:08,250 In this section, we'll take a look at a very important topic, which is understanding how do we visualize 3 00:00:08,250 --> 00:00:10,710 water filters in our convolutional networks? 4 00:00:11,640 --> 00:00:17,670 So this will help you gain a better intuition and an understanding of how your CNN's actually learning 5 00:00:17,670 --> 00:00:18,630 what it's learning. 6 00:00:18,960 --> 00:00:23,400 And hopefully, this helps you overall in understanding the entire deep learning process. 7 00:00:24,270 --> 00:00:26,400 So let's take a look at what CNN's Lynn. 8 00:00:26,520 --> 00:00:31,500 Well, firstly, we know learning involves adjusting the weights and parameters that lead to the lowest 9 00:00:31,500 --> 00:00:31,920 loss. 10 00:00:32,310 --> 00:00:34,530 That's that much we know we've covered that before. 11 00:00:35,250 --> 00:00:39,450 So let's take a look at this CNN here and look at the parameters. 12 00:00:39,660 --> 00:00:41,280 These are the parameters we're learning. 13 00:00:41,850 --> 00:00:43,740 So what does that mean? 14 00:00:44,040 --> 00:00:44,630 What do you mean? 15 00:00:44,640 --> 00:00:46,680 What do we mean by learning these these parameters? 16 00:00:46,680 --> 00:00:48,300 And what do we actually learn here? 17 00:00:48,900 --> 00:00:52,410 Well, let's take a look at an untrained filter. 18 00:00:52,710 --> 00:00:58,200 So let's take a look at a model like this, and let's look at a little convolutional little one right 19 00:00:58,200 --> 00:00:58,530 here. 20 00:00:59,100 --> 00:01:05,640 Here we have 22 filters, each of which each has a size of tree by tree by one, and would have been 21 00:01:05,640 --> 00:01:07,350 created by tree if it was a color. 22 00:01:07,350 --> 00:01:09,300 RGV image will always remember that. 23 00:01:09,930 --> 00:01:12,060 And this produces these feature maps here. 24 00:01:12,330 --> 00:01:13,710 So we use these filters here. 25 00:01:14,190 --> 00:01:17,410 So what are these filters look like after? 26 00:01:17,460 --> 00:01:23,700 Because remember, remember, we initialize these tree budget by one filters with random values. 27 00:01:24,090 --> 00:01:30,210 But after the training process is finished, can we visualize and look at these filters and see what 28 00:01:30,210 --> 00:01:30,540 they learn? 29 00:01:30,660 --> 00:01:32,130 And actually, yes, we can. 30 00:01:33,060 --> 00:01:36,510 This is what a tributary filters what your budget filters look like. 31 00:01:36,510 --> 00:01:39,300 You know when a CNN upgrade before. 32 00:01:40,080 --> 00:01:43,230 So you can see these are basically what this one looks like a plus sign. 33 00:01:43,230 --> 00:01:44,670 This one looks like a diagonal. 34 00:01:45,030 --> 00:01:48,360 They all look like some sort of diagonal straight line like these here. 35 00:01:48,930 --> 00:01:55,440 So you can immediately see some patterns and we can use a ModCloth lib to visualize these filters and 36 00:01:55,440 --> 00:02:01,200 what the darker areas correspond to here that correspond to low values of zero and the lighter areas 37 00:02:01,200 --> 00:02:03,750 correspond to the high values of of white. 38 00:02:04,050 --> 00:02:05,130 So that's 255. 39 00:02:06,180 --> 00:02:12,210 So this means here that a white or lighter areas in the filter correspond to the areas that have a higher 40 00:02:12,210 --> 00:02:12,660 widths. 41 00:02:13,020 --> 00:02:19,590 So these are the areas here that you can see them right here that have had got some larger widths, 42 00:02:19,590 --> 00:02:25,170 which means that they will be activated when this filter aligns with something on the image and we'll 43 00:02:25,170 --> 00:02:27,300 talk about that in the next section. 44 00:02:27,510 --> 00:02:29,580 Let's take a look at how these filters work. 45 00:02:30,270 --> 00:02:31,980 So remember what this filter does? 46 00:02:31,980 --> 00:02:39,030 Remember this for this slightly slid across the image at a stride of one usually, and it produces this 47 00:02:39,120 --> 00:02:40,380 this feature map right here. 48 00:02:41,340 --> 00:02:44,130 Well, let's take a look at the convolution operation. 49 00:02:44,160 --> 00:02:48,450 No, the convolution operation is the dot product of the input vector. 50 00:02:48,750 --> 00:02:51,960 That's what we're multiplying here by the width vector. 51 00:02:51,990 --> 00:02:53,370 That's what the filter is. 52 00:02:54,270 --> 00:03:00,450 Now, the dot product between two vectors is proportional to the angle between two vectors between those 53 00:03:00,450 --> 00:03:00,930 vectors. 54 00:03:01,260 --> 00:03:02,280 So what does that mean? 55 00:03:02,340 --> 00:03:08,160 Well, that means that the output that's the feature map is high when the angle between these vectors 56 00:03:08,160 --> 00:03:08,640 are zero. 57 00:03:09,180 --> 00:03:11,220 That means that vectors on the same direction. 58 00:03:11,580 --> 00:03:18,060 This also means you can consider it as they're basically along the lines of the same portion of the 59 00:03:18,060 --> 00:03:18,340 image. 60 00:03:18,360 --> 00:03:19,350 That means they line up. 61 00:03:19,770 --> 00:03:25,980 That means if we're looking for an edge to edge with an edge detect filter, when the edge covers the 62 00:03:25,980 --> 00:03:28,890 edge in the image here it activates that filter. 63 00:03:28,950 --> 00:03:30,250 That's exactly what it means. 64 00:03:30,270 --> 00:03:31,620 That means when the output is high. 65 00:03:32,370 --> 00:03:36,720 So these are some examples of filters that look for features in images. 66 00:03:37,290 --> 00:03:39,180 This is this was a trained CNN. 67 00:03:39,630 --> 00:03:41,280 And these are the filters that learn here. 68 00:03:41,490 --> 00:03:46,770 So you can see it learns quite distinct patterns, and you can tell immediately that these patterns 69 00:03:46,770 --> 00:03:53,820 would correspond to maybe some sort of texture or maybe some sort of like exact pattern in the image. 70 00:03:54,240 --> 00:04:00,120 So you can immediately see that these filters have learned to extract features from that image. 71 00:04:00,360 --> 00:04:02,220 Here's a look at some of the filters here. 72 00:04:02,820 --> 00:04:07,650 These are also you can tell these look more like diagonals and stripes and stuff like that. 73 00:04:07,680 --> 00:04:08,130 So. 74 00:04:08,640 --> 00:04:10,040 And here are some more filters. 75 00:04:10,050 --> 00:04:11,430 These look like little blobs. 76 00:04:12,390 --> 00:04:14,220 These are low level filters, by the way. 77 00:04:14,220 --> 00:04:19,230 You can tell because the dimensions are usually quite small, like in this case, it looks quite small. 78 00:04:19,920 --> 00:04:24,240 So how do we actually visualize visualize these photos now? 79 00:04:24,780 --> 00:04:25,830 It's actually quite simple. 80 00:04:26,250 --> 00:04:29,010 We can obtain the width and bias of a filter. 81 00:04:29,400 --> 00:04:30,270 You can increase. 82 00:04:30,270 --> 00:04:31,650 It's quite easy to do. 83 00:04:32,040 --> 00:04:35,290 So we'll start off with examples with cameras in this case. 84 00:04:35,610 --> 00:04:40,350 So you can use get with some cameras to extract the width and bias of a trained filter. 85 00:04:40,950 --> 00:04:43,410 Then we just normalized weights between zero and one. 86 00:04:43,980 --> 00:04:49,680 And we use Matloff Lib to plot the weight values in 2-D, which we'll do in the next lesson. 87 00:04:50,460 --> 00:04:54,990 However, actually, before we move on to the code, I'll just talk about this present. 88 00:04:55,320 --> 00:04:59,220 The slides on filter activations just because the code deals with. 89 00:04:59,840 --> 00:05:04,880 Them together, so instead of jumping around, it will do the two terrorist lights first. 90 00:05:05,030 --> 00:05:07,610 That's the section and the section here. 91 00:05:07,970 --> 00:05:11,930 Then we'll dive into the coda where we actually see the implementations of these. 92 00:05:12,560 --> 00:05:14,190 So I'll see you in the next section. 93 00:05:14,300 --> 00:05:14,750 Thank you.