Conference on Cognitive Computational Neuroscience

Semantic cognition allows for flexible acquisition, storage, and deployment of conceptual representations. Previously, different computational models have attempted to explain a wealth of empirical phenomena in semantic learning (Rogers & McClelland, 2004; Saxe et al., 2019). Our work focuses on phenomena documented in the form of analytical solutions to learning dynamics of deep linear networks which display progressive differentiation and stage-like transitions during learning. These phenomena have been frequently observed during human development, but have not been tested in the formal setting of a controlled experiment. Here, we asked if they are general features of human learning that persist when training adults over short time-spans. We trained participants on a semantic learning experiment that invoked hierarchical constraints on learned properties. Given our theoretical predictions, we compared human data to several classes of differently initialised neural networks. Findings indicate that human learning respects the hierarchical constraints. Furthermore, we find that human learning is most closely mirrored by neural networks which learn from small rather than large random weights. Such networks in particular are known to display patterns of progressive differentiation and stage-like transitions. Our results validate that simple neural networks can usefully describe phenomena observed in human semantic cognition.