Interpret multi-input neural networks
Opened this issue · 0 comments
felixpeters commented
Word embedding
- Analyze word embedding layers using t-SNE (use TensorBoard for analysis)
Feature visualization: What does the network detect?
- Perform feature visualization techniques (semantic dictionary, activation vectors)
Attribution methods: How are detectors assembled?
See Distill article and framework
-
Perform spatial attribution
-
Perform channel attribution