Computationally Efficient Forgetting via Base-Level Activation

As we apply cognitive models to complex, temporally extended tasks, removing declarative knowledge from memory, or forgetting, will become important both to model human behavior, as well as to scale computationally. The base-level activation (BLA) model predicts that the availability of specific memories is sensitive to frequency and recency of use. Memory decay based on this model has long been a core commitment of the ACT-R theory (Anderson et al., 2004), as it has been shown to account for a class of memory retrieval errors (Anderson, Reder, & Lebiere, 1996), and has been used in Soar (Laird, 2012) to investigate task-performance effects of forgetting short-term (Chong, 2003) and procedural (Chong, 2004) knowledge. Prior work has addressed many of the computational challenges associated with retrieving a single memory according to the BLA model (Petrov, 2006; Derbinsky, Laird, & Smith, 2010; Derbinsky & Laird, 2011). However, efficiently removing items from memory, while preserving BLA-model fidelity, is a different problem, which we address here. We formally describe the computational problem; present a novel approach to forget according to BLA in large memories; and evaluate using synthetic data.