Fig. 10 | Nature Communications

Fig. 10

From: Deviation from the matching law reflects an optimal strategy involving learning over multiple timescales

Fig. 10

The metaplastic model of synapses (the cascade model52) can capture key aspects of experimental data. a Network model. A decision is made according to the competition between the two action selection populations (G: choosing Green target; R: choosing Red target), mediated by the inhibitory population. The competition is determined by synaptic strength between the input population and the action selection populations. These synapses are plastic, and they encode the value of Green target and Red target8,12,20,56,61. b The cascade model of synapses. Each synapse makes Markov transitions between states with a different strength (depressed −, or potentiated +) and plasticity (upper states are more plastic) with given probabilities, where transition probabilities are designed to be ordered as \(\alpha _1 \gg \alpha _2 \gg \alpha _3,q_1 \gg q_2\). Note that transition probabilities are logarithmically distributed so that deeper states are harder to enter, and harder to leave. In general, the number of metaplastic states (vertical states) can be more than three8,52. Here we show three states for an illustrative purpose. c The model captures changes in matching behavior. The model was simulated in the same conditions experienced by Monkey F, according to the learning rules that allow reward-dependent transitions during experimental sessions, and transitions that incorporate forgetting during long breaks between experimental sessions8. d The model also captures the tradeoff between undermatching and the variance of choice

Back to article page