Extended Data Figure 10. A model of expert dynamics and a model with synaptic depression.

a. Schema of the expert model (see Methods for details). The network structure is identical to that for learning (Fig. 4e). Following literature⁵², we provided a non-contextual tonic input to vary lick times across trials (no plasticity is imposed as this is a model of expert).

b. Dynamics of ALM neurons in the model (top) and corresponding energy landscape (bottom). Different color indicates activity in trials with different lick times. The amplitude of tonic input changed the slope of the landscape, which changed the speed of dynamics. This reproduced the ‘temporal scaling’ of ramping dynamics consistent with experimental data (Fig.3) and previous report⁵².

c. The same format as in Fig.4f, but for a network model with synaptic depression of the synapse between cue and PT neurons. The network architecture is identical to that in Fig.4e, but with different synaptic weights (Extended Data Table. 4) and reward-dependent synaptic depression instead of potentiation (Methods).

d. The same format as in Fig.4g, but for the network model with synaptic depression of the synapse between cue and PT neurons.

e. The same format as in b, but for the network model with synaptic depression of the synapse between cue and PT neurons. Altogether, similar to the potentiation model, the depression model can reproduce the experimental observations.