Debunking Grad-ECLIP: A Comprehensive Study on Its Incorrectness and Fundamental Principles for Model Interpretation
For researchers in model interpretability, this paper corrects a flawed method and establishes guiding principles to prevent similar errors.
This paper proves that Grad-ECLIP, an ICML 2024 method for Transformer interpretation, is not novel and is actually equivalent to a simpler attention-based method (Attention-ECLIP). It also demonstrates that Grad-ECLIP produces interpretation results misaligned with the model's performance, and proposes two fundamental principles for correct model interpretation.
Grad-ECLIP is published at ICML 2024 and represents a new Transformer interpretation technical route (intermediate features-based). First, this paper demonstrates that the intermediate features-based technical route is not a novel one. Based on the existing attention-based route, we have developed Attention-ECLIP, which is completely equivalent to Grad-ECLIP but with simpler computation. Both through formal derivation and experimental validation, we prove that the intermediate feature-based route represented by Grad-ECLIP is actually an equivalent variant of the attention-based route. Next, this paper demonstrates that the Grad-ECLIP method is flawed. The model interpretation results obtained by Grad-ECLIP are not those of the original model, and the interpretation results are misaligned with the model's performance. We analyze the causes of Grad-ECLIP's flaws and propose, or rather, explicitly emphasize two fundamental principles that model interpretation should adhere to in order to avoid similar errors.