Reproduce “Emergence of Episodic Memory in Transformers: Characterizing Changes in Temporal Structure of Attention Scores During 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
Deven Mahesh Mistry, Anooshka Bajaj, Yash Aggarwal, Sahaj Singh Maini, Zoran Tiganj
Researchers at Indiana University Bloomington characterized the emergence of temporal biases, akin to human episodic memory effects, in Transformer models during training. They found that in-context temporal contiguity is highly localized, spanning 2-4 tokens, and is causally linked to the presence and function of induction heads.
Key insights
- Transformer temporal contiguity is localized to a narrow window of 2-4 lags, contrasting with the broader temporal scales observed in human episodic memory.
- Induction heads are causally responsible for the in-context serial recall (contiguity effect) observed in transformer outputs.
- Positional encoding magnitude positively influences the emergence of induction heads and the strength of recency effects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.