Reproduce “Attention that does not Explain Away” 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
Nan Ding, Xinjie Fan, Zhenzhong Lan, Dale Schuurmans, Radu Soricut
Research from Google and UT Austin proposes a Doubly-Normalized Attention Scheme (DNAS) to prevent information loss from input neurons in Transformer models, countering the "explaining away" effect of standard attention. This approach, alongside a hybrid variant (HNAS), yielded improved performance on Visual Question Answering, Natural Language Understanding tasks (e.g., SQuAD 1.1 F1 from 92.2 to 92.4), and abstractive summarization benchmarks, while also alleviating mode collapse.
Key insights
- The "explaining away" effect in standard Transformer attention (UNAS) was formally identified through a probabilistic GMM interpretation, revealing a potential for information loss.
- Doubly-Normalized Attention (DNAS) provides theoretical guarantees for information preservation, ensuring all input neurons contribute, and is mathematically linked to optimal transport via the Sinkhorn algorithm.
- The Hybrid Normalized Attention (HNAS) offers insights into task-specific information processing, demonstrating a strong preference for DNAS in VQA and NLU, with a more nuanced mix for summarization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.