Noise-aware training of neuromorphic dynamic device networks

Area of Science:

• Physics
• Computer Science
• Materials Science

Background:

• In materio computing leverages intrinsic system dynamics for embodied intelligence.
• Networks of devices enable complex tasks but are hard to design dynamically.
• Device noise and lack of physical models hinder network training.

Purpose of the Study:

• Introduce a framework for training networks of dynamical devices.
• Enable robust programming of embodied intelligence in complex systems.
• Address challenges in designing dynamic networks with inherent device noise.

Main Methods:

• Developed the Noise-Aware Dynamic Optimization (NADO) framework.
• Utilized Neural Stochastic Differential Equations (Neural-SDEs) as differentiable digital twins.
• Combined backpropagation through time with cascade learning for temporal property exploitation.

Main Results:

• Successfully trained networks of spintronic devices for temporal classification and regression.
• Demonstrated effective exploitation of physical device temporal properties.
• Showcased robust, gradient-based programming without analytical device descriptions.

Conclusions:

• The NADO framework enables effective training of dynamical device networks.
• Decoupling model training and network optimization reduces data needs.
• Facilitates gradient-based programming for embodied intelligence in complex systems.