Learning to select actions shapes recurrent dynamics in the corticostriatal system

Area of Science:

• Neuroscience
• Computational Neuroscience
• Cognitive Science

Background:

• Adaptive behavior relies on learning to select actions based on their values.
• The fronto-striatal system, particularly the dorsal-lateral prefrontal cortex (dlPFC) and dorsal striatum (dSTR), supports this learning.
• Neural recordings show these areas represent learning aspects, but computational mechanisms remain unclear.

Purpose of the Study:

• To investigate the computational mechanisms underlying action selection learning in the dlPFC-dSTR circuit.
• To model the dynamic interplay between value representation and action selection.
• To compare model activity with neurophysiological recordings from monkeys.

Main Methods:

• Developed a recurrent neural network (RNN) model of the dlPFC-dSTR circuit.
• Trained the RNN on an oculomotor sequence learning task.
• Compared model-generated activity with neural data from monkey dlPFC and dSTR.

Main Results:

• The trained RNN model autonomously represented and updated action values, closely matching corticostriatal recordings.
• Learning to select correct actions increased the distance between action-sequence representations in both the model and neural data.
• The model demonstrated that increased representational distance enhances the probability of selecting appropriate action sequences.

Conclusions:

• Neural circuit dynamics in the corticostriatal system support task learning by increasing representational distances.
• This mechanism enhances the likelihood of generating correct actions as learning progresses.
• The study advances understanding of neural computation and circuit dynamics in learning.