AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Deep Equilibrium Models

Shaojie Bai, J. Zico Kolter, Vladlen Koltun

NeurIPS 2019 · Architectures and efficiency › Attention and transformers · approach: State space and recurrent sequence models · 264 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
Replace recurrent sequence modeling with a DEQ self-attention transformerCopy Memory T=400 · Copy Memory T=400loss: 3.5e-6 vs 0.0501; 3.5e-6 vs 0.0491Improved
Use a DEQ TrellisNet instead of a 60-layer TrellisNetTrellisNet · Penn Treebank word-level testperplexity: 57.0 -> 57.1No clear effect
Use a DEQ TrellisNet instead of a 70-layer TrellisNetTrellisNet · WikiText-103 word-level testperplexity: 29.2 -> 29.0Improved
Use a DEQ transformer instead of a layer-stacked Transformer-XLTransformer-XL medium, 16 layers · WikiText-103 word-level testperplexity: 24.3 -> 24.2Improved
Use a DEQ transformer instead of a layer-stacked adaptive-embedding Transformer-XLTransformer-XL medium, adaptive embedding · WikiText-103 word-level testperplexity: 23.6 -> 23.2Improved
Use a DEQ transformer instead of a weight-tied Transformer-XLTransformer-XL small · WikiText-103 word-level testperplexity: 34.9 -> 32.4Improved
Use the DEQ formulation instead of conventional layer stackingDEQ models versus deep networks · Language modeling memory benchmarksmemory: up to 88%Improved
Increase equilibrium tolerance or lower the Broyden iteration limit (ablation)DEQ-Transformer · WikiText-103performanceHurt
Set the equilibrium tolerance too high (ablation)DEQ · Language modelingaccuracyHurt
Add dropout (ablation)DEQ · Root-finding convergenceBroyden convergence speedHurt
Use weight normalization (ablation)DEQ · Root-finding convergenceBroyden iteration growthImproved
Pretrain a shallow weight-tied stack before DEQ training (ablation)DEQ · DEQ trainingtraining efficiencyImproved

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