Synaptic and nonsynaptic plasticity approximating probabilistic inference.

Area of Science:

• Computational Neuroscience
• Systems Neuroscience
• Theoretical Neuroscience

Background:

• Neural circuits use synaptic and nonsynaptic plasticity for learning and memory.
• The interplay of Hebbian plasticity, homeostatic plasticity, neuromodulation, and intrinsic excitability in unified learning remains unclear.
• Existing models struggle to reconcile diverse plasticity mechanisms for coordinated learning.

Purpose of the Study:

• To propose a unified Hebbian learning rule for spiking neurons inspired by Bayesian statistics.
• To investigate how synaptic and nonsynaptic mechanisms can jointly orchestrate learning.
• To provide a biophysical model for Bayesian computation in neural systems.

Main Methods:

• Developed a Hebbian learning rule for spiking neurons incorporating Bayesian statistics.
• Modeled on-line adaptation of synaptic weights and intrinsic currents upon single spike arrival.
• Simulated learning and inference tasks using integrate-and-fire (IAF) neurons in a biologically plausible regime.

Main Results:

• The model demonstrates spike-timing-dependent plasticity and stable set-point return over long timescales.
• Temporally interacting memory traces locally estimate neuronal activation probabilities.
• Spike-based reinforcement learning is enabled by linking traces to external signals.
• Persistent firing and regulation of postsynaptic input are generated by activity-dependent ion channels.

Conclusions:

• Neurons can represent information as probability distributions, supporting probabilistic inference as a byproduct of neural computation.
• Coupled synaptic and nonsynaptic mechanisms operating over multiple timescales can achieve Bayesian computation.
• The proposed model offers a biophysical realization of Bayesian inference, reconciling various neural phenomena.