Interleaving facilitates the rapid formation of distributed representations

AbstractDistributed representations, in which information is encoded in overlapping populations of neuronal units, are essential to the remarkable success of artificial neural networks (ANNs) in many domains, and have been posited to be employed throughout the brain, especially in neocortex. A fundamental signature of ANNs employing distributed representations is that learning requires exposure to information in an interleaved order; exposure to new information in a blocked order tends to overwrite prior knowledge (i.e., 'catastrophic interference’). Because it is difficult to match human learning to the learning conditions of these networks, it is not known whether human learning exhibits these properties, which, if true, would suggest use of similar representations. To test this, we leveraged a recent proposal that parts of the hippocampus host distributed representations of the kind typically ascribed to neocortex, and adopted a hippocampally dependent task that contrasts the effects of interleaved versus blocked learning on a short timescale. Experiments 1a and 1b demonstrate that interleaved exposure facilitates the rapid perception of shared structure across items. Experiment 2 shows that only interleaved exposure permits useful inference when item associations need to be inferred based on statistical regularities. Together, these results demonstrate the power of interleaved learning and implicate the use of distributed representations in human rapid learning of structured information.

Return to previous page