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

In building spiking neural networks for edge devices, low power consumption and time scale matching with the input signal are essential characteristics for their analog implementation. In each node of the neural network, an activation function should be implemented to achieve nonlinearity between input spike frequency and output spike frequency. However, the conventional analog implementation often achieves nonlinearity in the voltage domain rather than in the spike frequency domain and consumes considerable power. In this study, a nonlinear frequency-conversion circuit based on a current-starved ring oscillator is proposed. In order to design nonlinearity in the frequency domain, the supply current for the ring oscillator is controlled as a function of input spike frequency. As a result, a hyperbolic-tangent nonlinearity is achieved in the simulation with the TSMC 180 nm process. Furthermore, the supply current is controlled in an extremely low range to achieve low power consumption of 0.2 nW and several hundred millisecond time constants, which are suitable for processing data with similar time scales such as biomedical data, environmental vibration, and so on.