Stable learning in stochastic network states.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Synaptic Plasticity

Background:

• Mammalian cerebral cortex exhibits irregular spontaneous activity in vivo.
• Existing associative plasticity rules, often studied in silent networks, struggle with this spontaneous activity due to sensitivity to spurious correlations.
• The impact of ongoing neural dynamics on signal processing and learning remains largely unknown.

Purpose of the Study:

• To develop novel spike-timing-dependent plasticity (STDP) learning rules that can operate effectively in realistic, noisy neural network states.
• To address the limitations of current plasticity rules in handling spontaneous brain activity.
• To unify various learning rules into a coherent framework.

Main Methods:

• Introduction of a new class of STDP learning rules incorporating local floating plasticity thresholds.
• Modeling the slow dynamics of these thresholds to account for metaplasticity.
• Analyzing the algorithm's performance in predicting homeostasis and stable learning in noisy conditions.

Main Results:

• The novel algorithm successfully predicts homeostasis in synaptic weights.
• It resolves the issue of asymptotic stable learning in noisy network states.
• The proposed rules naturally integrate and unify several known learning rules into a single framework.
• The mixed pre- and postsynaptic dependency of the floating threshold is mechanistically justified by known molecular pathways.

Conclusions:

• The developed learning rules offer a robust solution for synaptic plasticity in the presence of spontaneous neural activity.
• This framework provides a unified perspective on various learning rules and offers experimentally testable predictions.
• The findings advance our understanding of learning and signal processing in the mammalian cerebral cortex.