Learning Resets of Neural Working Memory

Abstract. Working memory is a key component of intelligence that the brain implements as persistent neural activations. How do persistent neurons learn to store information, and how can they be made to forget this information once it is no longer relevant? When animals learn episodic tasks, neurons in prefrontal cortex learn to represent task ends. We show that a biologically plausible neural network model equipped with persistent memory and a ‘reset ’ action can learn to store and forget information at task ends by reinforcement learning. The new model has competitive performance compared to a variety of (biologically implausible) models.