Memory Models: Quantitative

Abstract

Quantitative models of human memory differ on a number of dimensions—how items are represented (images, vectors of features, discrete- or continuous-valued features), how associations are represented (partitioned vectors, convolutions of item vectors, or outer-product matrices), how serial-order information is represented (inter-item associations or item-to-position associations), how information is stored (localized storage or superposition), how retrieval occurs (search or direct access), how context is bound to items and associations (concatenation, autoassociations, or three-way outer products), how forgetting occurs (context change or trace degradation), and how repetition works (laying down new traces or strengthening old traces). Most current models are global-matching models—which assume many items enter into the comparison (for recognition) or retrieval (for recall) process. The matched-filter and linear-association model of Anderson (1970) were two early models. Current models also discussed here are the resonance–retrieval theory, the SAM (search of associative memory) model, MINERVA2, the dual-trace array-perturbation model, OSCAR (Oscillator-based Associative Recall), and TODAM2 (theory of distributed associative memory).