Understanding Inhibition Through Maximally Tense Images
This work addresses a specific interpretability issue in neural networks for researchers, but it is incremental as it builds on existing tools and focuses on a narrow aspect of model behavior.
The paper tackles the problem of understanding feature inhibition in vision models by proposing maximally tense images (MTIs) as a tool, and introduces two visualization techniques to study them, though it notes challenges from superposition effects.
We address the functional role of 'feature inhibition' in vision models; that is, what are the mechanisms by which a neural network ensures images do not express a given feature? We observe that standard interpretability tools in the literature are not immediately suited to the inhibitory case, given the asymmetry introduced by the ReLU activation function. Given this, we propose inhibition be understood through a study of 'maximally tense images' (MTIs), i.e. those images that excite and inhibit a given feature simultaneously. We show how MTIs can be studied with two novel visualization techniques; +/- attribution inversions, which split single images into excitatory and inhibitory components, and the attribution atlas, which provides a global visualization of the various ways images can excite/inhibit a feature. Finally, we explore the difficulties introduced by superposition, as such interfering features induce the same attribution motif as MTIs.