Sequence Learning in a Single Trial

While recurrent neural networks can store pattern sequences through incremental learning, there could be a trade-off between network capacity and the speed of learning. The brain may solve this problem by using a two-stage system: a compact, low-capacity subsystem for rapid temporary storage of a few sequences, and a larger, high-capacity, slow learning subsystem for long-term storage of all sequences. In this study, we evaluate the ability of sparsely connected networks to learn pattern sequences in a single exposure using very high learning rates. The key factor is the amount of recurrent inhibition in the system. Our results indicate that post-synaptic gating in the learning rule enhances the rapid learning ability of networks. We also suggest how such rapid learning networks could transfer their memories to long-term storage.