Exploiting cognitive constraints to improve machine-learning memory models

Michael C. Mozer, University of Colorado Boulder

From the earliest days, theories of human vision have inspired the design of computer vision systems. In this talk, I'll argue that theories of human learning and memory can also inform the architecture of machine-learning systems that reason based on selective recall of their past experience. I'll describe the psychological literature on memory-trace storage and forgetting, focusing on how memory strength is influenced by the temporal distribution of experience and the passage of time. Phenomena as diverse as perceptual adaptation, sequential dependencies, and long-term declarative memory suggest that cortical memories have a temporal multiscale encoding. Incorporating this multiscale representation into dynamical neural models may help make task-relevant information accessible, and may also improve the performance of systems that attempt to predict nonstationary human preferences and desires.