Integrated Gradients / Integral of gradients

anchor

Integrates gradients along a straight path from a baseline input to the real input. Requires choosing a baseline, and the choice changes the result.

Findings

• IC-018Object information is localized to specific visual tokens in LLaVA-1.5 [compared-to]
• IC-1005Social bias neurons in BERT-base-cased and RoBERTa-base are concentrated in the deepest transformer layers [builds-on]
• IC-1156The VPT advantage over FT for ViT-B/16 is not explained by overfitting resistance or additional optimization dimensions; the specific feature-preservation mechanism of VPT is the key factor [supporting]
• IC-1409FF blocks in BERT and GPT-2 modify token-to-token contextualization, with the effect concentrated in specific layers and targeting specific linguistic compositions rather than simple word co-occurrence [primary]
• IC-1410FF's contextualization effects in BERT and GPT-2 are largely canceled by the residual connection and layer normalization, with LN's γ weights specifically shrinking the outlier dimensions in FF output [primary]
• IC-153ResNet50 relies on flower petals and green background features as shortcuts when classifying bee images [supporting]
• IC-160Pythia-70m and Gemma-2-2b implement subject-verb agreement across a relative clause via a circuit of number detectors, PP/RC boundary detectors, and verb form promoters, with Gemma-2-2b additionally using NP number trackers [supporting]
• IC-528The knowledge localization assumption fails for a large fraction of facts in GPT-2, Llama2-7B, and Llama3-8B, with 77% of facts classified as inconsistent knowledge in Llama3-8B [primary]
• IC-636Syntactic phenomena (determiner-noun and subject-verb agreement) localize to the same topmost-layer MLP neurons as factual information in BERT, GPT-2, and Llama-2 [primary]
• TM-010Fixed corner patches dominate Integrated Gradients maps regardless of image content [primary]