Designing nonlinearity in a current-starved ring oscillator for reservoir computing hardware.

Area of Science:

• Analog circuit design
• Spiking neural networks
• Low-power electronics

Background:

• Spiking neural networks (SNNs) require efficient analog implementations for edge devices.
• Conventional analog SNNs often lack frequency-domain nonlinearity and consume significant power.
• Matching signal time scales is crucial for processing data like biomedical signals.

Purpose of the Study:

• To propose a novel nonlinear frequency-conversion circuit for SNNs.
• To achieve nonlinearity in the spike frequency domain, not just the voltage domain.
• To enable low power consumption and appropriate time constants for edge computing.

Main Methods:

• Designed a nonlinear frequency-conversion circuit using a current-starved ring oscillator.
• Controlled the ring oscillator's supply current based on input spike frequency.
• Simulated the circuit using the TSMC 180 nm process.

Main Results:

• Achieved hyperbolic-tangent nonlinearity in the frequency domain.
• Demonstrated ultra-low power consumption of 0.2 nW.
• Obtained time constants in the hundreds of milliseconds range.

Conclusions:

• The proposed circuit effectively implements frequency-domain nonlinearity for SNNs.
• Ultra-low power consumption and suitable time constants make it ideal for edge devices.
• This approach is well-suited for processing time-scale-matched data, including biomedical and environmental signals.