Emergence of Episodic Memory in Transformers: Characterizing Changes in Temporal Structure of Attention Scores During Training
This provides insights into how transformers organize temporal information during in-context learning, comparing them to human memory mechanisms.
The study investigated temporal biases in transformer attention heads, finding that GPT-2 models exhibit patterns like human episodic memory such as temporal contiguity, primacy, and recency, with these effects eliminated after ablation of induction heads.
We investigate in-context temporal biases in attention heads and transformer outputs. Using cognitive science methodologies, we analyze attention scores and outputs of the GPT-2 models of varying sizes. Across attention heads, we observe effects characteristic of human episodic memory, including temporal contiguity, primacy and recency. Transformer outputs demonstrate a tendency toward in-context serial recall. Importantly, this effect is eliminated after the ablation of the induction heads, which are the driving force behind the contiguity effect. Our findings offer insights into how transformers organize information temporally during in-context learning, shedding light on their similarities and differences with human memory and learning.