AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

On the Convergence of Encoder-only Shallow Transformers

Yongtao Wu, Fanghui Liu, Grigorios Chrysos, Volkan Cevher

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Attention head and block redesign · 1 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
Use τ₀ = dₘ⁻¹/² rather than τ₀ = dₘ⁻¹Shallow Transformer · MNIST, especially at large widthconvergence rate: d_m = 1024 and 16384Improved
Use τ₀ = dₘ⁻¹/² rather than τ₀ = dₘ⁻¹Shallow Transformer · Small-width convergenceconvergence rateNo clear effect
Use LeCun or He initialization instead of NTK initializationShallow Transformer · Synthetic regressionconvergence rateImproved
Represent vector inputs along the embedding dimension (dₛ = 1, d = 100) (ablation)Shallow Transformer · Synthetic vector-input regressiontraining convergenceImproved
Represent vector inputs along the sequence dimension (dₛ = 100, d = 1) (ablation)Shallow Transformer · Synthetic vector-input regressiontraining convergenceHurt
Increase width from the sub-linear regime to the tested linear-or-larger widths (ablation)Shallow Transformer with LeCun initialization and τ₀ = dₘ⁻¹/² · Synthetic regressiontraining loss convergence: d_m= \{10, 100,1000, 4000\}Improved
Use LeCun or He initialization instead of NTK initializationShallow Transformer · Global convergence analysisconvergence rateImproved

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