Path Choice Matters for Clear Attribution in Path Methods
This work addresses the problem of unclear attributions in DNN interpretations for researchers and practitioners, though it appears incremental as it builds on existing path methods.
The paper tackles the ambiguity in interpreting deep neural networks (DNNs) by introducing the Concentration Principle and SAMP, a model-agnostic interpreter that efficiently finds near-optimal paths for attributions, resulting in significant outperformance over existing methods in quantitative experiments.
Rigorousness and clarity are both essential for interpretations of DNNs to engender human trust. Path methods are commonly employed to generate rigorous attributions that satisfy three axioms. However, the meaning of attributions remains ambiguous due to distinct path choices. To address the ambiguity, we introduce \textbf{Concentration Principle}, which centrally allocates high attributions to indispensable features, thereby endowing aesthetic and sparsity. We then present \textbf{SAMP}, a model-agnostic interpreter, which efficiently searches the near-optimal path from a pre-defined set of manipulation paths. Moreover, we propose the infinitesimal constraint (IC) and momentum strategy (MS) to improve the rigorousness and optimality. Visualizations show that SAMP can precisely reveal DNNs by pinpointing salient image pixels. We also perform quantitative experiments and observe that our method significantly outperforms the counterparts. Code: https://github.com/zbr17/SAMP.