Основные понятия
A novel dense associative memory model, Correlated Dense Associative Memory (CDAM), integrates both auto- and hetero-association in a unified framework for continuous-valued memory patterns using an arbitrary graph structure to semantically link memory patterns.
Аннотация
The author introduces a new dense associative memory model called Correlated Dense Associative Memory (CDAM) that integrates both auto- and hetero-association for continuous-valued memory patterns. CDAM uses an arbitrary graph structure to semantically link the memory patterns.
The key highlights and insights are:
CDAM's dynamics exhibit four distinct modes: auto-association, narrow hetero-association, wide hetero-association, and neutral quiescence, which can be controlled by modulating the balance between auto- and hetero-association.
Anti-Hebbian learning rules can be used to: (i) widen the range of hetero-association across memories; (ii) extract multi-scale representations of community structures in the memory graph; (iii) stabilize recall of temporal sequences; and (iv) enhance performance in non-traditional auto-association tasks.
CDAM can handle real-world data, replicate a classical neuroscience experiment on hetero-association, perform image retrieval, and simulate arbitrary finite automata.
The model provides insights into the potential mixture of auto- and hetero-associative dynamics in attention mechanisms of Transformer models, opening up new interpretability approaches.