Videos / AI

Episode 5 of 22 · AI, end to end

CNNs: how a machine learned to see

Chapters

1. One small picture, unravelled0:00
2. Flattening loses the grid0:16
3. The same cat, in a new corner0:32
4. Learn it again, in every place0:47
5. What an image really is1:04
6. Two demands, one fix1:23
7. Nine pixels, nine weights1:41
8. The same weights, reused1:58
9. The feature map2:14
10. Move the input, the map moves2:31
11. Stride and padding2:48
12. The size arithmetic3:06
13. Colour is three sheets3:25
14. Many kernels, many maps3:43
15. The one by one convolution4:01
16. Pooling throws detail away4:17
17. The window that never grows4:38
18. Space out, depth in4:55
19. Strokes, then a face5:13
20. Layer one is always edges5:31
21. Corners, then textures5:50
22. The middle: parts6:07
23. The top: whole objects6:22
24. Borrow the bottom6:39
25. Digits on a cheque6:57
26. Twenty twelve, and the G P Us7:17
27. V G G: just go deeper7:38
28. Deeper stopped helping7:58
29. The skip connection8:17
30. Variations on the same parts8:36
31. A camera over a conveyor8:59
32. A prior, borrowed from data9:17
33. Patches, and attention9:36
34. It only wins with the data9:55
35. The honest scoreboard10:16
36. The idea outlived pictures10:37
37. The whole thing10:57

CNNs: how a machine learned to see. Here is a photograph, two hundred pixels square, in colour. That is one hundred and twenty thousand numbers. Feed it to an ordinary fully connected layer of a thousand units and yo

Mark as watchedSaved in this browser only.