Fig. 3.

SRM demonstration. (Left) A graph that outlines the true polynomial function, the data drawn from the polynomial function (with added noise), and a neural network’s (NN) prediction. (Middle) The correlation between the raw residuals and the true residuals versus the correlation between the smoothed residuals and the true residuals for a simple linear model fit to the data. (Right) The average squared residual between the data and the true function versus the average residual between the neural network and the true function. As predicted, smoothed residuals correlate better with the true residuals when the error of the neural network falls below the noise in the data. Ten simulations were run for each dataset size, and error shading in Middle and Right reflect $\pm 1$ SEM.