Reward prediction error neurons implement an efficient code for reward.

Area of Science:

• Computational Neuroscience
• Neuroscience
• Systems Neuroscience

Background:

• Efficient coding principles optimize information transmission in neural systems.
• Dopaminergic reward prediction error (RPE) neurons signal discrepancies between expected and received rewards.

Purpose of the Study:

• To determine the optimal neural population for encoding reward distributions using efficient coding principles.
• To investigate the relationship between efficient coding and the responses of RPE neurons.

Main Methods:

• Applied efficient coding theory to derive the ideal neural population for representing reward distributions.
• Analyzed the properties of RPE neuron responses in mice and macaques.
• Derived novel learning rules based on efficient coding principles.

Main Results:

• RPE neuron responses exhibit characteristics of an efficient code, including broad midpoint distributions and specific gain/slope relationships.
• Neuron properties varied with the reward distribution's width, with narrower distributions leading to higher slopes.
• Derived learning rules that converge to the efficient code, with one rule resembling distributional reinforcement learning.

Conclusions:

• RPE neuron responses may be optimized for broadcasting efficient reward signals.
• This study establishes a connection between efficient coding and reinforcement learning frameworks in computational neuroscience.