Researchers Demonstrate Fully Analogue Neural Networks Using NDR Devices

Global: Researchers Demonstrate Fully Analogue Neural Networks Using NDR Devices

A team of researchers has introduced KANalogue, a fully analogue implementation of Kolmogorov‑Arnold Networks that leverages negative‑differential‑resistance (NDR) devices to perform both linear and nonlinear transformations without digital assistance. The work, posted on arXiv, details how the approach enables hardware‑level function approximation while maintaining an entirely analogue signal path, addressing a long‑standing bottleneck in analogue in‑memory computing.

Architecture Overview

KANalogue maps the intrinsic current‑voltage characteristics of NDR components to learnable, coordinate‑wise nonlinear functions, effectively embedding basis functions directly into device physics. These univariate functions are combined through crossbar‑based analogue summation, forming a hardware realization of Kolmogorov‑Arnold Networks that operates without intermediate digitisation.

Device Implementation

The prototype utilizes cold‑metal tunnel diodes as a representative NDR technology. By exploiting the devices’ negative differential resistance, the researchers constructed diverse nonlinear bases that can be programmed through voltage biasing, demonstrating that the framework is not confined to a single device class and can extend to other NDR platforms.

Experimental Results

Benchmarks on standard image‑classification tasks—MNIST, FashionMNIST, and CIFAR‑10—show that KANalogue attains competitive accuracy while employing substantially fewer parameters than analogue multilayer perceptrons. The crossbar node efficiency reported exceeds that of comparable analogue MLP designs, indicating a potential reduction in area and power consumption.

Performance Comparison

When evaluated against digital KAN implementations constrained by the same hardware limits, KANalogue approaches the digital performance levels, highlighting the viability of fully analogue networks for energy‑efficient inference. The results suggest that analogue KANs can close the gap with digital counterparts in scenarios where hardware resources are tightly bounded.

Future Prospects

The authors emphasize that the methodology can be generalized to a broad spectrum of NDR devices, paving the way for scalable, low‑power analogue neural networks. Continued exploration of device materials and integration strategies could further enhance the practicality of analogue AI accelerators for edge and IoT applications.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.