Matching recall and storage in sequence learning with spiking neural networks

Area of Science:

• Computational neuroscience
• Neural networks
• Spiking neural networks

Background:

• The brain must store and recall spiking sequences for cognitive functions.
• Existing biologically plausible learning rules struggle to robustly learn diverse spatiotemporal patterns.

Purpose of the Study:

• To derive a generic, biologically plausible learning rule for recurrent networks of stochastic spiking neurons.
• To ensure the learning rule is robust and applicable to a wide class of spatiotemporal activity patterns.

Main Methods:

• Developed a recurrent network model with visible and hidden stochastic spiking neurons.
• Derived a learning rule by minimizing an upper bound on Kullback-Leibler divergence.
• Analyzed the rule's consistency with spike-timing dependent plasticity and voltage-triplet rules.

Main Results:

• The derived learning rule supports spike-timing dependent plasticity (STDP): presynaptic spikes before postsynaptic spikes cause potentiation, otherwise depression.
• Synaptic learning rules for visible neurons match the voltage-triplet rule.
• Synaptic learning rules for hidden neurons are modulated by a global factor with astrocyte-like properties.

Conclusions:

• The novel learning rule provides a robust mechanism for spiking neural networks to learn spatiotemporal patterns.
• The rule's consistency with STDP and voltage-triplet rules enhances biological plausibility.
• The astrocyte-modulated rule for hidden neurons offers novel insights and testable predictions for neural computation.