Spike-Based Bayesian-Hebbian Learning of Temporal Sequences.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience

Background:

• Temporal processing is vital for cognitive and motor functions.
• Neocortical microcircuits' ability to encode and replay stimulus sequences remains an open question.

Purpose of the Study:

• To propose and validate a computational model for reliable sequential information processing in neural networks.
• To investigate the role of synaptic plasticity in learning and replaying sequences.

Main Methods:

• A modular attractor memory network utilizing the spike-based Bayesian Confidence Propagation Neural Network (BCPNN) learning rule.
• Simulations employing adaptive exponential integrate-and-fire model neurons (AdEx).

Main Results:

• The model successfully learned and replayed sequences through meta-stable attractor transitions driven by synaptic plasticity.
• Sequence replay speed and learning efficiency were influenced by biophysical parameters like stimulus duration and noise levels.
• Demonstrated flexible combinatorial coding where sequence elements could participate multiple times.

Conclusions:

• The proposed spiking attractor network model provides a principled framework for understanding sequence learning and replay.
• Multiple interacting plasticity mechanisms can coordinate hetero-associative learning for temporal information processing.