Reward-based training of recurrent neural networks for cognitive and value-based tasks

• DOI: 10.7554/elife.21492
• Publication Date: 2017-01-13T13:00:20Z
• Authors: H Francis Song; Guangyu R Yang; Xiao-Jing Wang
• Keywords: Maximization; Representation; Value (mathematics)

Trained neural network models, which exhibit features of activity recorded from behaving animals, may provide insights into the circuit mechanisms cognitive functions through systematic analysis and connectivity. However, in contrast to graded error signals commonly used train networks supervised learning, animals learn reward feedback on definite actions reinforcement learning. Reward maximization is particularly relevant when optimal behavior depends an animal’s internal judgment confidence or subjective preferences. Here, we implement reward-based training recurrent a value guides learning by using decision predict future reward. We show that such models capture behavioral electrophysiological findings well-known experimental paradigms. Our work provides unified framework for investigating diverse value-based computations, predicts role representation essential but not executing, task.