Fig. 1: Behavior and modeling overview.

a, Sensory decision-making paradigm. Mice indicated whether a contrast grating was on the left or right of a screen using a wheel. b, Representative behavior of mouse KS014 (also used later). This shows improvements in behavior and the concomitant extension of the contrast set. c, Visual representation of the main components of the model. Each state, represented by a circle, has an associated observation distribution, shown inside its circle. This is implemented via logistic regression, which considers the contrast on the current trial and a weighted history of previous choices (the latter is not shown here). The weights underlying these regressions can change from session to session, resulting in shifts of the PMFs they represent; we depict this evolution here with varying shades of color. States are connected to other existing states via transition probabilities P. In addition to that, states also have the option to transition into a new state, to describe a type of behavior that is not well captured by any of the existing states. Finally, staying in the same state for more than one trial is not modeled via a self-transition probability; instead, each state has its own duration distribution. Panel a reproduced from ref. ¹¹ under a Creative Commons licence CC BY 4.0.