Paper detail

Training recurrent neural networks with sparse, delayed rewards for flexible decision tasks

Recurrent neural networks in the chaotic regime exhibit complex dynamics reminiscent of high-level cortical activity during behavioral tasks. However, existing training methods for such networks are either biologically implausible, or require a real-time continuous error signal to guide the learning process. This is in contrast with most behavioral tasks, which only provide time-sparse, delayed rewards. Here we show that a biologically plausible reward-modulated Hebbian learning algorithm, previously used in feedforward models of birdsong learning, can train recurrent networks based solely on delayed, phasic reward signals at the end of each trial. The method requires no dedicated feedback or readout networks: the whole network connectivity is subject to learning, and the network output is read from one arbitrarily chosen network cell. We use this method to successfully train a network on a delayed nonmatch to sample task (which requires memory, flexible associations, and non-linear mixed selectivities). Using decoding techniques, we show that the resulting networks exhibit dynamic coding of task-relevant information, with neural encodings of various task features fluctuating widely over the course of a trial. Furthermore, network activity moves from a stimulus-specific representation to a response-specific representation during response time, in accordance with neural recordings in behaving animals for similar tasks. We conclude that recurrent neural networks, trained with reward-modulated Hebbian learning, offer a plausible model of cortical dynamics during learning and performance of flexible association.