PulseAugur
LIVE 03:43:39
research · [2 sources] ·
0
research

From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models

By PulseAugur Editorial · Summary by None from 2 sources

Researchers have introduced a new framework for interpreting vision models, unifying local, global, and mechanistic analysis around instance-specific Effective Receptive Fields (iERFs). This approach uses pointwise feature vectors and their iERFs to generate activation-faithful explanations and semantic labels for latent vectors. The framework, which includes methods like Sharing Ratio Decomposition (SRD) and Concept-Anchored Feature Explanation (CAFE), has demonstrated superior fidelity and robustness across various model architectures compared to existing baselines. AI

Summary written by None from 2 sources. How we write summaries →

IMPACT Provides a unified method for understanding vision model decision-making, potentially improving debugging and trust.

RANK_REASON This is a research paper detailing a new framework for interpreting vision models.

Read on arXiv cs.CV →

COVERAGE [2]

  1. arXiv cs.CV TIER_1 · Yearim Kim, Sangyu Han, Nojun Kwak ·

    From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models

    arXiv:2605.00474v1 Announce Type: new Abstract: Modern vision models achieve remarkable accuracy, but explaining where evidence arises, what the model encodes, and how internal computations assemble that evidence remains fragmented. We introduce an iERF-centric framework that uni…

  2. arXiv cs.CV TIER_1 · Nojun Kwak ·

    From Local to Global to Mechanistic: An iERF-Centered Unified Framework for Interpreting Vision Models

    Modern vision models achieve remarkable accuracy, but explaining where evidence arises, what the model encodes, and how internal computations assemble that evidence remains fragmented. We introduce an iERF-centric framework that unifies local, global, and mechanistic interpretabi…

RELATED ENTITIES

RELATED TOPICS