Holographic Reduced Representations for Working Memory Concept Encoding

Abstract

Artificial neural networks (ANNs) utilize the biological principles of neural computation to solve many engineering problems while also serving as formal, testable hypotheses of brain function and learning. However, since ANNs often employ distributed encoding (DE) methods they are underutilized in applications where symbolic encoding (SE) is preferred. The Working Memory Toolkit was developed to aid the integration of an ANN-based cognitive neuroscience model of working memory into symbolic systems by mitigating the details of ANN design and providing a simple DE interface. However, DE/SE conversion is still managed by the user and tuned specifically to each task. Here we utilize holographic reduced representation (HRR) to overcome this limitation since HRRs provide a framework for manipulating concepts using a hybrid DE/SE formalism that is compatible with ANNs. We validate the performance of the new toolkit and show how it automates the process of DE/SE conversion while providing additional cognitive capabilities.