Can the inference logic of a DNN be faithfully explained as symbolic concepts? Our #ICLR2024 paper (arxiv.org/abs/2305.01939) makes an initial theoretical attempt to address this question. We prove that under three sufficient conditions, a DNN only encodes a small number of

Reasoning or memorization? Our recent work (arxiv.org/abs/2405.11880) proposes a new axiomatic framework to formulate and quantify the exact reasoning and memorization effects in an LLM. See our project page for a brief introduction: sjtu-xai-lab.github.io/ReasoningMemor…

What is the mathematical mechanism shared by different attribution methods? Our TPAMI paper (ieeexplore.ieee.org/document/10414…) unifies 14 post-hoc attribution methods from the perspective of interactions. We prove that all 14 attribution methods can be represented as a weighted

How to explain a DNN’s generalization ability and learning dynamics through the lens of interaction concepts? Our recent works (arxiv.org/abs/2405.10262, and arxiv.org/abs/2407.19198 in #NeurIPS2024) discover and theoretically prove a two-phase dynamics of interaction concepts