7 comments

[ 3.4 ms ] story [ 31.8 ms ] thread
This is an excellent introduction to neural networks, deep neural networks, and the recent use of GPUs to speed up training them.

Good explanation of Type 1/Type 2 errors at the end, too: 'it correctly identified birds 95% of the time, but only spotted 90% of the birds in our test data'.

Google also has something else than lots of people's images. It has a reCaptcha system where users have to identify things in images. I assume Google is crowd sourcing training its AI systems.
Zeiler's ECCV14 paper is good follow on reading from this. He was one of the first people to actually visualise what these networks were doing.

https://www.cs.nyu.edu/~fergus/papers/zeilerECCV2014.pdf

Have a look at Figure 2.9. What you see is that the upper layers tend to be very general, basic filters that act as edge, corner and circle detectors. When you move to lower levels, the filters get more complicated and you start seeing things like animal parts, faces and so on.

I really liked the visual representation of the "8" bitmap! It's kind of neat how `0` has a fairly pixel density compared to `255`, so even as numbers the `8` is still recognizable.
I'm the author. I just saw this post.

I actually cheated to make that illustration better. In a real grayscale image, 0 is black and 255 is white. I inverted the image to make the numbers easier to see as an "8". I swapped it so that black was 0 and white was 255 :)

Hey Aymeric, thanks for posting this. I'm a big fan of your work on tflearn!