Inferring the function performed by a recurrent neural network.

Area of Science:

• Systems Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Understanding neural circuit function is a central goal in systems neuroscience.
• Traditional top-down models require pre-defined functions, limiting their applicability to unknown neural systems.
• Inferring function directly from neural data remains a significant challenge.

Purpose of the Study:

• To propose an inverse reinforcement learning (RL) framework for inferring the function of neural networks from observed data.
• To develop a method that does not require prior assumptions about the specific function being performed.
• To enable predictions of neural network adaptation to changing environments or structures.

Main Methods:

• Utilizing an inverse reinforcement learning (RL) framework.
• Assuming neuron responses are optimized to reach 'rewarded' states.
• Inferring the reward function from observed neural network responses.

Main Results:

• Successfully demonstrated the use of inverse RL to infer reward functions from neural network activity.
• The inferred reward function can predict neural network dynamics adaptation.
• Provides a method to understand neural computations without prior functional hypotheses.

Conclusions:

• The proposed inverse RL framework offers a data-driven approach to uncover neural circuit functions.
• This method facilitates predictions of how neural networks adapt to dynamic conditions, such as cell death or altered sensory inputs.
• Advances the understanding of neural computation and adaptation in complex systems.