STDP installs in Winner-Take-All circuits an online approximation to hidden Markov model learning

Area of Science:

• Computational Neuroscience
• Neural Circuits
• Machine Learning

Background:

• Navigating dynamic environments requires rapid extraction and prediction of hidden causes from multi-modal sensory input.
• Cortical microcircuits are fundamental to information processing in the brain.

Purpose of the Study:

• To investigate how generic cortical microcircuit motifs enable predictive processing capabilities.
• To demonstrate the emergence of hidden Markov model (HMM) functionality through unsupervised learning mechanisms.

Main Methods:

• Modeling of pyramidal cells with lateral excitation and inhibition.
• Simulation of spike-timing-dependent plasticity (STDP) in Winner-Take-All circuits.
• Analysis of emergent functional properties analogous to hidden Markov models.

Main Results:

• Cortical microcircuit motifs with STDP exhibit noise-robust emergence of HMM-like properties.
• Online STDP application achieves significant HMM functionality without supervision or rewards.
• Reward-gated STDP enables near-optimal learning via rejection sampling.

Conclusions:

• Generic cortical microcircuits possess inherent capabilities for probabilistic inference and prediction.
• Unsupervised learning via STDP is a powerful mechanism for developing complex computational functions in neural systems.
• Reward-gated STDP offers a pathway to achieve optimal learning in artificial and biological systems.