Reproduce “Mechanistic Interpretability of Deep Neural Networks: Methods, Frameworks, and Perspectives” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lorenzo Cirillo, Irene Amerini, Andrea Cavallaro
The authors present a causal workflow for mechanistic interpretability, combining visualization, disentanglement, activation interventions, and circuit tracking. In one ViT-B/16 sound-classification example, replacing an activation from a vacuum-cleaner run in a rain run changes the prediction to vacuum cleaner; the vacuum-cleaner logit rises from 0.002 to 7.753. This demonstrates an effect in these runs, not a general explanation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.