Principled neuromorphic reservoir computing.

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Machine Learning

Background:

• Reservoir computing utilizes nonlinear recurrent neural circuits for signal encoding.
• Monolithic reservoir networks face challenges in simultaneously buffering signals and expanding them into nonlinear features.
• Separate configuration of memory buffering and higher-order feature expansion outperforms traditional reservoir computing for time-series prediction.

Purpose of the Study:

• Propose a configurable neuromorphic representation scheme for enhanced reservoir computing.
• Achieve competitive prediction performance with improved scaling properties.
• Implement the proposed scheme on neuromorphic hardware.

Main Methods:

• Combined randomized representations from reservoir computing with principles for approximating polynomial kernels.
• Utilized Sigma-Pi neurons for computing higher-order features, enabling summation and multiplication of inputs.
• Implemented the memory buffer and Sigma-Pi networks on the Loihi 2 neuromorphic platform.

Main Results:

• The proposed configurable scheme demonstrates competitive performance on prediction tasks.
• The approach exhibits significantly better scaling properties compared to prior methods that directly materialize higher-order features.
• Successful implementation on Loihi 2 hardware validates the practical feasibility of the proposed neuromorphic representation.

Conclusions:

• The configurable neuromorphic representation scheme offers an efficient alternative to traditional reservoir computing for complex prediction tasks.
• The use of Sigma-Pi neurons and separate configuration of components enhances scalability and performance.
• Neuromorphic hardware platforms like Loihi 2 are suitable for implementing advanced computational models such as this.