Flamingo: a Visual Language Model for Few-Shot…

Flamingo connects a frozen vision encoder and a frozen language model with a fixed visual bottleneck and gated cross-attention layers. Read it to see how interleaved images and text share context, why the gates begin closed, and which celebrated few-shot claims remain hard to reproduce.

Reading focus: How a Perceiver Resampler turns variable image and video features into sixty-four visual tokens. Why zero-initialized gated cross-attention preserves the language model before the visual bridge learns to open. How the single-image attention mask, broad task results, proprietary data, and unreleased model bound the evidence.

NeurIPS 2022. Jean-Baptiste Alayrac et al.. 42 min read, medium difficulty.