Local online learning in recurrent networks with random feedback.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Recurrent neural networks (RNNs) are vital for processing time-dependent data like movement and memory.
• Traditional training algorithms lack biological plausibility due to violations of causality and locality.

Purpose of the Study:

• To develop a biologically constrained learning rule for RNNs.
• To enable efficient training of RNNs that mimics brain function.
• To address limitations in training RNNs over extended time scales.

Main Methods:

• Derived a novel, local learning rule approximating gradient-based methods for RNNs.
• Incorporated local synaptic activity and random feedback projections for weight updates.
• Proposed an augmented circuit architecture for handling long sequences.

Main Results:

• The new learning rule demonstrates effectiveness through mathematical analysis and simulations.
• Trained RNNs successfully performed various time-dependent tasks.
• The augmented architecture facilitated learning over extended timesteps.

Conclusions:

• The novel learning rule offers a biologically plausible alternative for training RNNs.
• This approach enhances RNN capabilities for complex, sequential data processing.
• Future research can explore further biological constraints in artificial neural networks.