Conference on Cognitive Computational Neuroscience

We employed goal-driven deep reinforcement learning models to understand neural representations of human visuomotor control. The model learned a similar action policy to humans as a response to visual inputs. Moreover, low-level pixel inputs and high-level features related to a task goal are respectively represented in the early and late layers of the trained model. The later layers of the model better predict neural representations for visuomotor control although we did not find evidence of functional hierarchy along the dorsal stream.