You should apply some sort of brightness filter over the first layer which is dependent on the trained weights of the network for each corresponding node. This way it might be possible to understand how the network can differentiate between triangles and squares because we will "be seeing like the network"
great video, hope you make more. note regarding triangle and square looking similar: they enclose an area whereas the cross don't, that is probably what it corresponds to.
I love the animations and the deep dive into theories. Try this again but with maybe 64x64 or 128x128 layers. I feel like patterns would be a lot more noticeable
yes thats possible problem that for humans at least the pattern is complete white noise so the bigger you make the image the more confusing it just gets :)
You should apply some sort of brightness filter over the first layer which is dependent on the trained weights of the network for each corresponding node. This way it might be possible to understand how the network can differentiate between triangles and squares because we will "be seeing like the network"
great video, hope you make more. note regarding triangle and square looking similar: they enclose an area whereas the cross don't, that is probably what it corresponds to.
Very cool video!
Great explanation
I love the animations and the deep dive into theories. Try this again but with maybe 64x64 or 128x128 layers. I feel like patterns would be a lot more noticeable
yes thats possible
problem that for humans at least the pattern is complete white noise
so the bigger you make the image the more confusing it just gets :)
this is literally, like reverse engineering a brain.