AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Transformers are uninterpretable with myopic methods: a case study with bounded Dyck grammars

Kaiyue Wen, Yuchen Li, Bingbin Liu, Andrej Risteski

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Transformer expressivity and learnability · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Fix attention weights to 0, producing uniform attentionTwo-layer Transformer · Dyckaccuracy: > 99% accuracyImproved
Remove position embeddingTwo-layer Transformer · Dyck₂,₄ validationaccuracy: ≥ 97% accuracyNo clear effect
Add linear position embeddingTwo-layer Transformer · Prefix [[[[]]]](((())))average attention variation: 2.27 -> 2.20No clear effect
Use a minimal first layer with the specified embeddingTwo-layer Transformer · Prefix [[[[]]]](((())))average attention variation: 2.20 -> 0.24Hurt
Add contrastive regularization to reduce balance violationTransformer trained on Dyck · Dyck prefixes of length 400 to 500accuracyImproved
Train with standard training rather than contrastive regularizationTransformer with a minimal first layer · Dyck prefixes of length 400 to 500accuracyHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy