Fig. 3: Uncertainty analysis for ‘total’ in choice decoding.

a The extraction of 9 frames from a total of 120 frames for each neuron’s neural activity in each trial from 2P data. All neurons’ neural activities for each trial were processed through a signal encoder, followed by input into a transformer block, which accommodates varying numbers of neurons owing to its architecture. During the inference process, dropout was enabled for 1,000 iterations to calculate variance and derive uncertainty. b–d Uncertainties in response decoding over the course of learning (b), influenced by stimulus frequency (c) and modulated by response type (d). Uncertainty decreases with learning (b) and shows an increase in incorrect trials (d). e A significant negative correlation between d′ and neural uncertainty, indicating that improved task performance (higher d′) is associated with reduced uncertainty. The red line represents the regression line, and the shaded area around it indicates the 95% confidence interval. f A conceptual illustration of response uncertainty. There was a correlation of lower uncertainty with correct responses and higher uncertainty with incorrect responses during the go stimulus, and elevated uncertainty at toss-up frequency. In the illustration, the uncertainty related to the response (licking) is depicted for different stimuli: go (green), no-go (red) and probe (blue). The degree of uncertainty is conveyed through the number of tongue-shaped question marks. A check mark indicates a correct result, while an X mark indicates an incorrect result. The error bars represent the s.e.m.