Taming the chaos gently: a predictive alignment learning rule in recurrent neural networks

Area of Science:

• Computational neuroscience
• Machine learning
• Neuroplasticity

Background:

• Recurrent neural circuits can exhibit chaotic spontaneous activity, complicating learning.
• Existing methods like FORCE learning require non-local plasticity and rapid adaptation.
• Biological plausibility on local timescales for synaptic adaptation remains a challenge.

Purpose of the Study:

• To propose a novel, biologically plausible framework for training chaotic recurrent neural networks.
• To investigate a learning rule that suppresses chaos by aligning predictions with network activity.
• To demonstrate the framework's ability to perform supervised learning on diverse tasks.

Main Methods:

• Developed a "predictive alignment" framework for recurrent neural networks.
• Implemented a biologically plausible, local plasticity rule.
• Trained networks to suppress chaotic dynamics by aligning recurrent predictions with spontaneous activity.

Main Results:

• Successfully trained networks to generate patterned activities from chaotic dynamics.
• Achieved supervised learning of complex low-dimensional attractors and delay matching tasks.
• Demonstrated capability in learning dynamic, high-dimensional data like movie clips.

Conclusions:

• Predictive alignment offers a biologically plausible approach to learning in chaotic recurrent circuits.
• Aligning predictions with chaotic activity is an effective strategy for network training.
• This framework advances understanding of how predictions support learning in neural systems.