Learning Long Temporal Sequences in Spiking Networks by Multiplexing Neural Oscillations

Area of Science:

• Computational Neuroscience
• Neuroscience
• Artificial Intelligence

Background:

• Cognitive and behavioral tasks require precisely-timed neural activation sequences.
• External stimuli alone cannot fully explain these complex temporal patterns.
• Spiking recurrent neural networks are a key model for understanding brain function.

Purpose of the Study:

• To demonstrate how chaotic and noisy spiking recurrent neural networks can generate repeatable and reliable spatiotemporal activity patterns.
• To propose a general solution for autonomous generation of rich neural activity patterns.
• To investigate the role of multi-periodic oscillatory input in neural computation.

Main Methods:

• Utilized spiking recurrent neural networks with multi-periodic oscillatory input signals.
• Trained the model on tasks including speech generation, motor control, and spatial navigation.
• Analyzed the generated spatiotemporal activity for patterns and features relevant to neural processing.

Main Results:

• The model successfully learned and generated complex activity patterns for various tasks.
• Demonstrated temporal rescaling of spoken words and exhibited sequential neural activity.
• In spatial navigation, the model learned compressed sequences, replayed place cells, and showed ripples and theta phase precession.

Conclusions:

• Multi-periodic oscillatory neuronal inputs are a key mechanism for generating precisely timed activity in recurrent neural circuits.
• This approach offers a framework for understanding autonomous neural pattern generation.
• The findings have implications for both neuroscience and artificial intelligence applications.