Neuronal Sequence Models for Bayesian Online Inference.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Neuronal dynamics across species and brain regions exhibit sequence-like structures during cognitive tasks and for stationary concepts.
• These sequences display robust and reproducible spatiotemporal activation patterns, suggesting a fundamental role in brain function.
• The brain is increasingly understood as an active predictor of sensory input, supported by evidence from ethology, physiology, and neuroscience.

Purpose of the Study:

• To illustrate how neuronal sequences are critical for probabilistic predictive information processing.
• To identify dynamical principles that generate neuronal sequences.
• To review how sequence-generating models can be embedded in a functional hierarchy of time scales for sensory recognition and prediction.

Main Methods:

• Review of imaging and electrophysiological studies on neuronal sequences.
• Exploration of dynamical principles underlying sequence generation.
• Integration of sequence-generating models within hierarchical brain organization frameworks.
• Introduction of the Bayesian brain hypothesis for online recognition and prediction.

Main Results:

• Neuronal sequences are fundamental to probabilistic predictive information processing.
• Dynamical principles can generate these sequences, which can be organized hierarchically.
• This hierarchical organization forms a generative model for sensory input recognition and prediction.

Conclusions:

• Investigating sequential brain activity reveals principles of information processing and prediction.
• These neuroscientific principles can inform the design of more efficient artificial neural networks for machine learning.
• Spatiotemporally structured methods and hierarchical networks in machine learning share principles with neuronal sequences.