Extended Data Fig. 7: Electrophysiological data and controls.

a. Single-neuron task coding by subject. For each subject, the left panel shows the rat’s choice behavior for all the electrophysiology recording sessions. The dots with error bars show the probability of choosing lottery against ΔEV of the two options. The lines are the psychometric curves estimated by a logistic fit to the data, the thin gray lines are fit to each session, the thick gray line fit to all the sessions combined. The right panel shows the distribution of the t-statistic for lottery value (y-axis) and upcoming choice (x-axis) for all the neurons recorded in each animal. Gray dots indicate the non-task relevant neurons, light blue dots indicate the pure choice neurons, orange dots indicate the pure lottery selectivity neurons, and green dots indicate tuning for both upcoming choice and lottery values (n = 893 trials, 9 sessions for subject 2224; n = 1,754 trials, 15 sessions for subject 2238; n = 1,421 trials, 12 sessions for subject 2244; n = 1,228 trials, 9 sessions for subject 2263; n = 430 trials, 4 sessions for subject 2261; n = 1,040 trials, 11 sessions for subject 2264, the circles with error bars are the mean and 95% binomial confidence intervals.). b. We recorded neurons from the FOF of 4 rats to test whether the relationship between firing rate and lottery could be due to FOF encoding the percept of the different lottery sounds. Out of the 105 neurons recorded, only 6 had p < 0.05 encoding of lottery cues, which was not significantly different than expected by chance (χ²(1, 105) = 0.051, p = 0.82, one-sided). c. Left: Decoding accuracy (Pearson’s r) for pseudopopulation decoding with shuffled training labels. With only 6 lotteries, the correlation can be very high by chance, but the distributions of accuracy are clearly distinct from the real decoding. Right: Comparing decoding using mean squared error (MSE) instead of r. Using MSE avoids the problems of computing correlation with small n. The decoding with the real data is significantly better than the shuffled data for all population sizes (n = 50 pseudosessions, all p < 0.00001, The box whisker plots show the median, lower/upper quartile, minimum/maximum and the outliers of the data, the notch showed the \(median\pm (1.57\times interquartilerange)/\sqrt{n}\), not adjusting for multiple comparisons).