Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Trading Generalization for Working Memory Capacity in Neural Network Representations

Jonathan König¹, Steven M Frankland², Taylor Whittington Webb³; ¹Universität Osnabrück, ²Dartmouth College, ³Université de Montréal

Presenter: Jonathan König

Why can we hold so few items in mind at the same time? Recent theoretical work suggests this owes to a tradeoff between generalization and working memory capacity in representation learning systems. This approach provides a potential explanation for observations of capacity limits in both humans and artificial neural networks. In the present study, we seek to validate this theory by modeling behavior in the change detection task that is commonly used to evaluate human working memory capacity. Modeling is based on artificial neural networks that we evaluate in terms of the change detection task, their ability to generalize, and their representational structure. Applying various learning objectives during training allows us to probe the predicted tradeoff. We show that generalization-optimized models develop structured representations that strongly limit working memory capacity. Models trained to avoid representational structure through orthogonal representations do not show these limitations but fail to generalize. Finally, models that are optimized for high working memory capacity form orthogonal representations and also fail to generalize. Together, these findings provide evidence that one set of representations cannot serve both generalization and working memory capacity, further strengthening evidence for shared limitations in artificial and human cognition.

Topic Area: Memory, Learning & Knowledge Structures