AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Stable and low-precision training for large-scale vision-language models

Mitchell Wortsman, Tim Dettmers, Luke Zettlemoyer, Ari S. Morcos, Ali Farhadi, Ludwig Schmidt

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 11 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 SwitchBack int8, keeping weight-gradient computation in higher precisionCLIP ViT-Huge · ImageNet zero-shotaccuracy vs bfloat16: 0.1 percentage pointsNo clear effect
Use LLM.int8() for trainingCLIP ViT-Huge · ImageNet zero-shotaccuracy vs bfloat16: 5.9 percentage pointsHurt
Replace transformer linear layers with SwitchBack int8CLIP ViT-Huge · CLIP trainingspeedup: 13%; 25%Improved
Use SwitchBack simulated float8CLIP ViT-Huge · ImageNet zero-shotaccuracy vs bfloat16: 0.1 percentage pointsNo clear effect
Use tensor-wise float8 quantization for inputs, weights, and gradients without other modificationsCLIP ViT-Large · CLIP trainingtraining stabilityHurt
Add zero-initialized layer-scale to tensor-wise float8 training (ablation)CLIP ViT-Large · CLIP trainingtraining stabilityImproved
Add gradient clipping to tensor-wise float8 training (ablation)CLIP ViT-Large · CLIP trainingtraining stabilityNo clear effect
Add KQ layernorm to tensor-wise float8 training (ablation)CLIP ViT-Large · CLIP trainingtraining stabilityNo clear effect
Replace tensor-wise weight quantization with row-wise and column-wise weight quantization (ablation)SwitchBack · CLIP training at tested scalesaccuracyNo clear effect
Save quantized rather than full-precision tensors for the backward pass (ablation)SwitchBackM · CLIP ViT-HugeaccuracyNo clear effect
Add AdaFactor-style update clipping to AdamWCLIP ViT-Huge · CLIP trainingloss spikesImproved
Use update clipping instead of gradient clippingAdamW on CLIP ViT-Huge · ImageNet zero-shotaccuracyImproved

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