Fig. 6. A two-pathway neural network trained to perform a reaching task exhibits enhanced retention of highly practiced behaviors.

a In this task, the agent, given one of four cues, is trained to perform a center-out reach over 10 timesteps to a particular target. b The neural network trained to perform the task receives a target-specific cue and the current position as input. The network contains one pathway learned with supervised learning (SL) and another with Hebbian learning (HL). The output is a two-dimensional vector determining the movement velocity in the next timestep. c The loss (mean squared error) on each of the four targets in sequence during training, with a longer training period for the first target, and with Hebbian learning either active (SL+HL) or inactive (SL only). d The loss (mean ± SEM) during testing on each of the four targets, after training is complete. The green bar shows a control network in which targets were randomly interleaved during training (*p = 0.01, two-sample two-sided t test, n = 21 randomly initialized networks; boxes and whiskers denote 25th and 75th percentiles and 5th and 95th percentiles, respectively).