4 papers
A Good Initialization is All You Need for Faithful Visual Attribution
Zihan Gu, Jiayu Wang, Hua Zhang +1
Faithful visual attribution identifies which image regions support a model prediction. Search-based perturbation methods lead the insertion--deletion faithfulness frontier by maski…
Deconstructing Positional Information: From Attention Logits to Training Biases
Zihan Gu, Ruoyu Chen, Han Zhang +2
Positional encodings enable Transformers to incorporate sequential information, yet their theoretical understanding remains limited to two properties: distance attenuation and tran…
PhaseWin Search Framework Enable Efficient Object-Level Interpretation
Zihan Gu, Ruoyu Chen, Junchi Zhang +3
Attribution is essential for interpreting object-level foundation models. Recent methods based on submodular subset selection have achieved high faithfulness, but their efficiency…
Beyond Progress Measures: Theoretical Insights into the Mechanism of Grokking
Zihan Gu, Ruoyu Chen, Hua Zhang +2
Grokking, referring to the abrupt improvement in test accuracy after extended overfitting, offers valuable insights into the mechanisms of model generalization. Existing researches…