Fig. 2: Decoding results of the neuron transformer.

a The structure of the neuron transformer for UQ. b Model accuracy according to stimulus over the ‘total’ 8 days, with ‘naive’ indicating the first 3 days and ‘expert’ the last 3 days. (total 0.850, naive 0.813, expert 0.880) c Model accuracy according to response type, lick versus no lick (total 0.765, naive 0.678, expert 0.834) For decoding, the output classes involved are identifying 200-Hz versus 600-Hz stimuli. The chance performance, given two output classes, is 1/2. d Decoding results of the model for each response type pair (hit/miss 0.907, hit/FA 0.886, CR/miss 0.856, CR/FA 0.798) The chance performance is 1/2. e The decoding performance of stimulus and response across days of training. The chance performance is 1/2. f Model accuracy across days for no-go/toss-up (240–320 Hz), go-like (440–560 Hz) and go stimuli. There was a sustained low decoding performance by the model for responses at toss-up frequencies over time. g Stimulus decoding according to stimulus frequency. h Response decoding according to stimulus frequency. There was a decreased efficacy in decoding responses to vibration stimuli around the psychometric threshold. i Excluding one mouse for the test set and using data from all others for training and validation, and running the deep learning model ten times with day 7 of learning as the test set and other days for training and validation. j Trials lacking premovement demonstrated satisfactory model performance, contrasting with the insufficient outcomes in trials featuring premovement. The error bars represent the s.e.m.