State-dependent plasticity routes learning signals in recurrent circuits

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Neuroscience

Background:

• * Synaptic plasticity is crucial for learning in biological and artificial systems.
• * Current computational models often rely on non-local signals like backpropagation (BP) for complex tasks.
• * Biological data suggests synapses may employ simpler, local plasticity rules.

Purpose of the Study:

• * To investigate computational models with synapses that switch plasticity rules based on their strength.
• * To develop a recurrent neural network (RNN) integrating Hebbian learning and BP.
• * To reconcile local learning rules in the brain with non-local rules in AI models.

Main Methods:

• * Designed a recurrent neural network (RNN) with plasticity-switching synapses (psRNN).
• * Synapses switch between Hebbian-like learning for weak connections and backpropagation (BP) for strong connections.
• * Evaluated psRNN performance on cognitive tasks like working memory.

Main Results:

• * The psRNN learned cognitive tasks in fewer trials compared to BP-only RNNs.
• * This efficiency stems from BP's improved gradient estimation, Hebbian initialization, and rule-switching.
• * The model developed a lower-rank, more feedforward recurrent structure.

Conclusions:

• * Combining simple and complex synaptic plasticity rules can enhance learning efficiency in computational models.
• * The psRNN framework offers a biologically plausible mechanism for credit assignment.
• * Findings provide testable connectomic predictions and new hypotheses for brain computation.