Model-based inference of synaptic plasticity rules.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Machine Learning in Neuroscience

Background:

• Inferring synaptic plasticity rules is crucial for understanding brain learning.
• Existing methods struggle with complex, nonlinear dependencies in neural and behavioral data.

Purpose of the Study:

• To develop a novel computational method for inferring synaptic plasticity rules from experimental data.
• To apply this method to both neural and behavioral data, uncovering complex learning dynamics.

Main Methods:

• Parameterized function approximation of plasticity rules (truncated Taylor series or multilayer perceptrons).
• Gradient descent optimization of plasticity parameters using full data trajectories.
• Validation through simulations and application to Drosophila reward-learning data.

Main Results:

• Successfully recovered known plasticity rules (e.g., Oja's rule) and complex reward-modulated rules.
• Identified an active forgetting component in Drosophila reward learning, enhancing model accuracy.
• Demonstrated robustness to noise in experimental data.

Conclusions:

• The proposed computational framework effectively infers complex synaptic plasticity rules.
• This method provides new insights into the computational principles of learning and memory.
• Reveals active forgetting as a key mechanism in invertebrate reward learning.