Learning spatiotemporal signals using a recurrent spiking network that discretizes time.

Area of Science:

• Computational Neuroscience
• Neural Networks
• Spiking Neuron Models

Background:

• The brain must learn to generate complex spatiotemporal sequences for various behaviors.
• Existing computational models often lack biological plausibility in their learning mechanisms.
• Understanding how neural networks encode and learn sequential information is a key challenge.

Purpose of the Study:

• To propose a biologically plausible spiking recurrent neural network model for learning spatiotemporal sequences.
• To investigate how such networks encode temporal information and map it to other dimensions like space or phase.
• To demonstrate robust sequence replay during spontaneous neural activity.

Main Methods:

• Developed a spiking recurrent network comprising excitatory and inhibitory neurons.
• Trained the network's dynamics to encode time.
• Utilized a read-out layer and Hebbian learning rules for encoding spatiotemporal patterns.

Main Results:

• The model successfully learned spatiotemporal dynamics on behaviorally relevant timescales.
• Synaptic weights to the read-out neurons encoded diverse spatiotemporal patterns.
• Learned sequences were robustly replayed during spontaneous activity regimes.

Conclusions:

• The proposed spiking neural network model offers a biologically plausible mechanism for learning and encoding spatiotemporal sequences.
• Hebbian learning rules are sufficient for encoding complex temporal and spatial information.
• The model's ability to replay sequences during spontaneous activity has implications for memory consolidation and retrieval.