Using neuron deletion to analyse neural networks, we find:
— DeepMind (@DeepMindAI) March 21, 2018
1) interpretable neurons ("cat neurons”) are no more important than confusing neurons for image classification;
2) networks that generalise better are harder to break
Blog: https://t.co/TCDJ3VdrHAhttps://t.co/xBkqGV6Chj
from Twitter https://twitter.com/DeepMindAI
March 21, 2018 at 09:20AM
via IFTTT
No comments:
Post a Comment