Reproduce “Sigmoid Loss for Language Image Pre-Training” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, Lucas Beyer

Google DeepMind researchers introduce Sigmoid Loss for Language-Image Pre-training (SigLIP), an alternative to softmax-based contrastive learning that enables training highly competitive models with significantly fewer computational resources, achieving up to 84.5% zero-shot ImageNet accuracy in two days on four TPUv4 chips. The work also demonstrates that performance for this task saturates around a batch size of 32k, challenging the assumption that ever-larger batch sizes are always beneficial.

Key insights

  • Performance in language-image pre-training saturates around a global batch size of 32k, indicating that indefinitely scaling batch size does not yield proportional gains.
  • The pairwise sigmoid loss demonstrates superior performance over softmax at smaller batch sizes and exhibits increased robustness to noisy web-scale data.
  • Crucial insights for stable training include the introduction of a learned bias term and adjustments to optimizer parameters (e.g., "β2") when dealing with large batch sizes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.