Fig. 3: Evolution of oddball representation across the neuronal population in experimental data and an RNN model.

a, Tone responses as a function of oddball identity and index in an example unit. The red bar indicates tone presentation. b, The accuracy of neuronal population oddball decoding for patients p5 and p6 (n = 43 oddball-responsive units), for the first half of each block (left) or second half of each block (right), combined across blocks. The boxes represent the 25th–75th percentiles, notched at the median. The whiskers are the extrema within 150% of the IQR past the box, plus symbols represent outliers. Statistical analysis was performed using two-sided paired t-tests; **P = 0.004. c, Similar to b but for tone identity. d, Decoding accuracy for tone (purple) and oddball (green) identity as a function of trial position. Each point represents SVM accuracy within a set of 50 trials starting at the index location. The dashed lines are linear fits. e, The Euclidean distance (left) and cosine angle (right) between standard and oddball neuronal population response vectors, computed for each oddball trial. The lines are linear fits with the 95% confidence intervals. f, Schematic of the RNN model trained to differentiate between two different tone frequencies, indicated as tone A and tone B. g, The training paradigm for the RNN as compared with the human experiment. h, Network response across trials for a single example RNN model unit for the three training phases. a.u., arbitrary units. i, The decoding accuracy of RNN for tone identity (purple) and oddball identity (green) (n = 100 generated trials; Methods). The boxes represent the 25th–75th percentiles, notched at the median; the whiskers are the extrema. j, Evolution of Euclidean distance (brown) and cosine angle (pink) between oddball trials and the average standard trial across the RNN population. The shading represents the s.e.m. across ten runs. No adjustments were made for multiple comparisons.