New method traces distinguishability in transformers using stochastic LayerNorm

Researchers have developed a new method to analyze the internal workings of transformer models by introducing stochastic Layer Normalization. This modification allows for the statistical distinguishability of representations, grounding interpretability in functional outcomes rather than just proximity. The approach uses a shared global rate budget to distribute finite precision across the model's layers, enabling the tracing of how distinctions propagate through MLP blocks and attention heads. Experiments with ViT-S and GPT-2 small demonstrate this technique's ability to reveal continuous perturbations and head-specific sensitivities, offering a new perspective on transformer computation. AI

IMPACT Provides a new interpretability technique for understanding transformer computations, potentially aiding in model debugging and development.

RANK_REASON The cluster contains a research paper detailing a new method for analyzing transformer models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• Bhattacharyya distance
• CatalyzeX
• DagsHub
• Gotit.pub
• GPT-2 small
• Hugging Face
• IArxiv
• LayerNorm
• ScienceCast
• Vít Sopko

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →