Human-like learning Framework for frequency-skewed multi-level classification
- Amarjot Singh, Stanford University, Stanford, California, United States
- Jay McClelland, Stanford University, Stanford, California, United States
AbstractContemporary deep neural network based classification systems are typically designed to learn information at a single level of granularity from datasets in which all items occur with equal frequency. Humans, on the other hand, acquire information at several different levels of granularity from experiences that contain some items more frequently than others. This allows us to learn and differentiate frequent items better from other items. We investigate the consequence of learning from a natural frequency/multi-level dataset in a deep neural network designed to model the human neocortex, complemented in some simulations with a replay buffer, playing the role of the human hippocampus. The NC network, when trained on its own, is able to learn more frequent items relatively quickly and differentiate them better from other items, as human learners do. However, the network's performance on infrequent and unseen examples pays a price in generalization performance compared to a standard training regime. The replay buffer serves to ameliorate these deficiencies, and we introduce a computationally and psychologically motivated replay weighting scheme that performs better than two alternatives.