Memory in humans and deep language models: Linking hypotheses for model augmentation
This work addresses computational complexity issues in deep learning models for researchers, but it appears incremental as it builds on existing memory-augmentation approaches.
The paper tackles the limitation of Transformer models in generalizing over long temporal durations by proposing memory-augmentation based on insights from human memory literature, using cross-domain linking hypotheses and evaluating surprisal as an example.
The computational complexity of the self-attention mechanism in Transformer models significantly limits their ability to generalize over long temporal durations. Memory-augmentation, or the explicit storing of past information in external memory for subsequent predictions, has become a constructive avenue for mitigating this limitation. We argue that memory-augmented Transformers can benefit substantially from considering insights from the memory literature in humans. We detail an approach for integrating evidence from the human memory system through the specification of cross-domain linking hypotheses. We then provide an empirical demonstration to evaluate the use of surprisal as a linking hypothesis, and further identify the limitations of this approach to inform future research.