Computational mechanisms of distributed value representations and mixed learning strategies

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Real-world reward learning involves complex options with multiple attributes, posing a significant challenge.
• The neural mechanisms behind value representation and learning strategy adoption in such environments are not well understood.

Purpose of the Study:

• To investigate the computational and neural underpinnings of complex learning strategies in humans.
• To model human learning behavior using recurrent neural networks (RNNs) in a multi-dimensional probabilistic learning task.

Main Methods:

• Human participants performed a multi-dimensional probabilistic learning task.
• Recurrent neural networks (RNNs) were trained to replicate human learning and choice data.
• Analysis included examining RNN representations, connectivity, and simulated neural lesions.

Main Results:

• Human participants estimated stimulus-outcome associations by combining feature and conjunction reward probabilities.
• The RNN models revealed that this mixed strategy depends on a distributed neural code.
• Opponency between excitatory and inhibitory neurons, via value-dependent disinhibition, was identified as a key mechanism.

Conclusions:

• The study elucidates the computational mechanisms for complex learning strategies in naturalistic settings.
• Findings suggest a neural basis involving distributed coding and specific neuronal interactions for adaptive reward learning.