A new interpretability method computes 'reasons vectors' from neuron activations and measures how strongly each neuron supports propositions about the input, with experiments on MNIST, Adult, and SST2.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Explaining Neural Networks with Reasons
A new interpretability method computes 'reasons vectors' from neuron activations and measures how strongly each neuron supports propositions about the input, with experiments on MNIST, Adult, and SST2.