(A) The spatial selectivity of the nine most informative ACC cells for decoding position in one session chosen by the decoding network. Spike density on the track is coded by color, from blue to red. The units are rank ordered by importance from top left to bottom right. (B) The root mean squared error (RMSE) of the position decoded from the ACC activity as compared to the actual position of the rat for each session (line) and rat (shade), showing that the deep artificial neural network generates lower prediction error than does a Bayesian decoder for each of the seven sessions tested. These session-averaged errors are inflated by occasional large errors around the reward zones, as described below. Error bars show standard deviation of 20 randomly selected training and test sets for each session and method. (C) The distribution of changes in decoded position by noise. (D) Error vectors for two representative laps of the task. The arrows indicate the magnitude and direction of the decoding error every 50 ms. (E) Cumulative probabilities of the prediction error magnitude for the seven sessions. The dotted lines indicate the median, and the arrows indicate the median error for the left (blue) and right (black) laps session shown in panel C (green curve). (F) Decoded position for test data from one session, color coded by the actual position (inset) (G) The error computed every 50 ms in one representative session, represented as a box plot according to track position as shown in Figure 1. The box plot shows the median (horizontal lines in boxes), 95% confidence intervals (notches), first and third quartiles (boxes ends), and outliers (dots). There are a disproportionate number of outliers in the bin corresponding to the target feeder locations, but the median prediction accuracy is as good at these feeders as anywhere else on the track. (H) The mean (top) and maximal (bottom) prediction error for discretized positions on the track, showing that the very large errors occurred exclusively at the location of the target feeders. These show mean of means computed from all sessions.