AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Transformers learn through gradual rank increase

Enric Boix-Adserà, Etai Littwin, Emmanuel Abbe, Samy Bengio, Joshua M. Susskind

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 5 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
Train a practical ViT with Adam and track attention-weight perturbationsViT · CIFAR-10stable rank of attention-weight perturbationsImproved
Train a practical ViT with Adam and track attention-weight perturbationsViT · CIFAR-100stable rank of attention-weight perturbationsImproved
Train a practical ViT with Adam and track attention-weight perturbationsViT-Base/16 · ImageNetstable rank of attention-weight perturbationsImproved
Train a language transformer and track attention-weight perturbationsGPT-2 · Wikitext-103stable rank of attention-weight perturbationsImproved
Train with SGD rather than AdamViT · CIFAR-10 and CIFAR-100stable rank of attention-weight perturbationsImproved
Use small initialization and diagonal attention weights, training only attention-head weightsDiagonal attention-head transformer · Diagonal attention-head toy taskrank increase per training stageImproved
Reduce initialization scale (ablation)ViT · CIFAR-10stable rank after trainingImproved
Use default rather than smaller initialization scale (ablation)ViT · CIFAR-10visibility of incremental rank increaseHurt
Apply the practical-model analysis without diagonal-weight or small-initialization assumptionsVision and language transformers · CIFAR-10, CIFAR-100, ImageNet, and Wikitext-103gradual increase in stable rankImproved
Compare the learned perturbation spectrum with the initialized weight spectrumViT · CIFAR-10rank of learned attention-weight perturbationImproved

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