Training Physical Neural Networks for Analog In-Memory Computing

In-memory computing (IMC) architectures mitigate the von Neumann bottleneck encountered in traditional deep learning accelerators. Its energy efficiency can realize deep learning-based edge applications. However, because IMC is implemented using analog circuits, inherent non-idealities in the hardware pose significant challenges. This paper presents physical neural networks (PNNs) for constructing physical models of IMC. PNNs can address the synaptic current's dependence on membrane potential, a challenge in charge-domain IMC systems. The proposed model is mathematically equivalent to spiking neural networks with reversal potentials. With a novel technique called differentiable spike-time discretization, the PNNs are efficiently trained. We show that hardware non-idealities traditionally viewed as detrimental can enhance the model's learning performance. This bottom-up methodology was validated by designing an IMC circuit with non-ideal characteristics using the sky130 process. When employing this bottom-up approach, the modeling error reduced by an order of magnitude compared to conventional top-down methods in post-layout simulations.

Learning Reservoir Dynamics with Temporal Self-Modulation

Reservoir computing (RC) can efficiently process time-series data by transferring the input signal to randomly connected recurrent neural networks (RNNs), which are referred to as a reservoir. The high-dimensional representation of time-series data in the reservoir significantly simplifies subsequent learning tasks. Although this simple architecture allows fast learning and facile physical implementation, the learning performance is inferior to that of other state-of-the-art RNN models. In this paper, to improve the learning ability of RC, we propose self-modulated RC (SM-RC), which extends RC by adding a self-modulation mechanism. The self-modulation mechanism is realized with two gating variables: an input gate and a reservoir gate. The input gate modulates the input signal, and the reservoir gate modulates the dynamical properties of the reservoir. We demonstrated that SM-RC can perform attention tasks where input information is retained or discarded depending on the input signal. We also found that a chaotic state emerged as a result of learning in SM-RC. This indicates that self-modulation mechanisms provide RC with qualitatively different information-processing capabilities. Furthermore, SM-RC outperformed RC in NARMA and Lorentz model tasks. In particular, SM-RC achieved a higher prediction accuracy than RC with a reservoir 10 times larger in the Lorentz model tasks. Because the SM-RC architecture only requires two additional gates, it is physically implementable as RC, providing a new direction for realizing edge AI.