Figure 4.

Learning speed increases with GCL. A: example of learned predictions after 1, 5, and 50 trials of learning, with predictions in red and target function in black. B: example learning trajectory of MSE fit with a double exponential. Black circles: MSE of network output on each trial. Red line: double exponential fit MSE during learning. Here, step size was 10⁻⁶ and z-scored GCL threshold was 0. We use the exponents k from the exponential fit to measure learning speed. C: learning speed as a function of GCL threshold (red dots). Blue line: learning speed in networks lacking GCL, i.e., mossy fibers directly innervate output Purkinje unit, gradient descent step size was 10⁻⁶. GCL, granule cell layer; MSE, mean squared error.