A new technique for analyzing how neural nets make decisions. Not necessarily incomprehensible black boxes after all!
OpenAI
 @OpenAI
6 Mar 2019
In collaboration with Google, we're releasing Activation Atlases: a new technique for visualizing what interactions between neurons can represent. 💻Blog: blog.openai.com/introducing-… 📝Paper: distill.pub/2019/activation-… 🔤Code: github.com/tensorflow/lucid/… 🗺️Demo: distill.pub/2019/activation-…

Mar 6, 2019 · 5:08 PM UTC

4
83
4
317
Cool writeup in @WIRED about the technique: "Shark or Baseball? Inside the ‘Black Box’ of a Neural Network" wired.com/story/inside-black…

Inside the ‘Black Box’ of a Neural Network

New research from Google and OpenAI offers insight into how neural networks "learn" to identify images.

wired.com
1
9
1
41
Replying to @gdb
good start, but being limited to image recognition is indeed limiting. i was surprised that they didn't mention this, as there are so many other applications of NNs.
1
2
One step at a time :)!
3
Replying to @gdb
@OpenAI Such a nice technique to oversee the parameter space and connections driving them. It would be great to use this personally
Replying to @gdb
The tool is great, but clearly shows that DNN have a problem