A Wilson-Cowan reservoir computer for interpretable spatiotemporal vision.

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Machine Learning

Background:

• Reservoir computing leverages complex dynamics for computation.
• Existing methods often lack interpretability and energy efficiency.
• Neural fields offer a biologically plausible framework for complex dynamics.

Purpose of the Study:

• To develop a Wilson-Cowan reservoir computer (WC-RC) for efficient and interpretable feature extraction.
• To evaluate the WC-RC's performance on benchmark image datasets (MNIST, Fashion-MNIST).
• To explore the functional role of neural dynamics in reservoir computation.

Main Methods:

• Utilized a retinotopic excitatory-inhibitory neural field as a structured reservoir.
• Employed a two-stage sampler to extract spatiotemporal features.
• Trained recurrent neural networks (Att-LSTM, GRU, LSTM) and MLPs/ridge classifiers on extracted features.
• Performed ablation studies on lateral couplings and neighbour interactions.

Main Results:

• WC-RC achieved strong performance with recurrent readouts (Att-LSTM, GRU, LSTM) within a fixed feature budget.
• The WC-RC representation is partially linearly decodable but benefits from temporal modeling.
• Wave dynamics are crucial for recurrent model performance, while excessive smoothing degrades accuracy.
• WC-RC outperforms Echo State Networks (ESNs) and offers comparable or better accuracy than CNNs with greater interpretability and lower resource requirements.

Conclusions:

• WC-RC provides an interpretable, energy-aware reservoir for neuromorphic vision.
• Spatiotemporal dynamics, specifically travelling waves, are functionally important for computation.
• The WC-RC framework is suitable for efficient deployment on hardware accelerators like FPGAs/ASICs.