Biological neurons act as generalization filters in reservoir computing

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Reservoir computing, a machine learning paradigm, processes time-series data using nonlinear system dynamics.
• The role of nonrandom network architecture, like modularity, in biological neuronal networks (BNNs) for information processing remains unclear.
• Understanding how cortical modularity integrates with neuronal biophysics is crucial for characterizing BNN function.

Purpose of the Study:

• To investigate the computational capabilities of cultured BNNs using the reservoir computing framework.
• To explore how embedded modular architecture influences information processing in BNNs.
• To assess BNNs' potential for tasks like short-term memory and spoken digit classification.

Main Methods:

• Utilized optogenetics and calcium imaging to record multicellular responses in cultured BNNs.
• Employed micropatterned substrates to engineer modular architecture within BNNs.
• Applied reservoir computing framework and linear decoders to analyze BNN dynamics and computational abilities.

Main Results:

• Modular BNNs exhibited dynamics classifiable by linear decoders, with modularity positively correlating with classification accuracy.
• BNNs demonstrated short-term memory (several 100 ms) and successfully performed spoken digit classification.
• BNN-based reservoirs enabled categorical learning and generalization, outperforming direct linear decoding.

Conclusions:

• Cultured BNNs with engineered modularity exhibit functional reservoir computing properties.
• BNNs possess inherent short-term memory and generalization capabilities crucial for complex computations.
• Findings advance mechanistic understanding of information representation in BNNs and support physical reservoir computing systems.