Neural Circuits for Dynamics-Based Segmentation of Time Series.

Area of Science:

• Computational Neuroscience
• Time Series Analysis
• Machine Learning

Background:

• The brain processes continuous sensory streams to infer underlying environmental dynamics.
• Identifying changes in these dynamics (segmentation) is crucial for extracting behaviorally relevant information.
• Existing methods may not be biologically plausible due to computational demands or reliance on global information.

Purpose of the Study:

• To develop and evaluate biologically plausible algorithms for time series segmentation based on dynamic changes.
• To address the challenge of segmenting signals in a streaming setting with local learning rules.
• To offer solutions applicable to brain regions with and without feedback connections.

Main Methods:

• A model-based algorithm derived from optimizing a mixture of autoregressive processes, utilizing prediction error feedback.
• A model-free algorithm employing a running estimate of signal autocorrelation for segmentation, suitable for systems lacking feedback.
• Testing algorithms on synthetic autoregressive data with piecewise-constant parameters and real-world voice recordings.

Main Results:

• Both algorithms demonstrated high accuracy in segmenting time series signals with changing dynamics.
• Performance was comparable to oracle methods that assume knowledge of ground-truth model parameters.
• The model-free approach proved effective in scenarios lacking feedback mechanisms.

Conclusions:

• Biologically plausible algorithms can effectively perform time series segmentation based on dynamic shifts.
• These methods offer a framework for understanding neural signal processing and can be applied in various biological contexts.
• The developed algorithms provide practical tools for analyzing dynamic time series data.