Bridging Function Approximation and Device Physics via Negative Differential Resistance Networks

Songyuan Li, Teng Wang, Jinrong Tang, Ruiqi Liu, Yuyao Lu, Feng Xu, Bin Gao, Xiangwei Zhu · 2025

Read Bridging Function Approximation and Device Physics via Negative Differential Resistance Networks by Songyuan Li, Teng Wang, Jinrong Tang, Ruiqi Liu, Yuyao Lu, Feng Xu, Bin Gao, Xiangwei Zhu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-23638.txt.

Introduction

Songyuan Li and nine co-authors build a neural network that computes entirely in analogue hardware, with no digital help. Analogue in-memory chips do matrix multiplication cheaply, but they lack learnable nonlinear steps, the curves that give networks their power. The team turns to Kolmogorov-Arnold Networks, which place trainable one-variable functions on each connection instead of fixed activations.

Their design, KANalogue, realizes those functions directly in devices with negative differential resistance, where current falls as voltage rises over a range. Using cold-metal tunnel diodes, they shape varied nonlinear curves and sum them through a crossbar array, keeping the whole signal path analogue. On MNIST, FashionMNIST, and CIFAR-10 the system reached strong accuracy with far fewer parameters than rival analogue designs, nearing digital KAN performance. The approach works for a broad class of such devices, pointing toward larger, low-energy analogue AI.

Catalog record

CREATORSongyuan Li, Teng Wang, Jinrong Tang, Ruiqi Liu, Yuyao Lu, Feng Xu, Bin Gao, Xiangwei Zhu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-23638

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bridging Function Approximation and Device Physics via Negative Differential Resistance Networks Songyuan Li, Teng Wang, Jinrong Tang, Ruiqi Liu, Yuyao Lu, Feng Xu, Bin Gao, Xiangwei Zhu Abstract Achieving fully analog neural computation requires hardware that can natively implement both linear and nonlinear operations with high efficiency. While analogue matrix-vector multiplication has advanced via compute-in-memory architectures, nonlinear activation functions remain a bottleneck, often requiring digital or hybrid solutions. Inspired by the Kolmogorov-Arnold framework, we propose KANalogue, a fully analogue implementation of Kolmogorov-Arnold Networks (KANs) using negative differential resistance devices as physical realizations of learnable univariate basis functions. By leveraging the intrinsic negative differential resistance characteristics of tunnel diodes fabricated from NbSi 2 N 4 /HfSi 2 N 4 heterostructures, we construct coordinate-wise nonlinearities with distinct curvature and support profiles. We extract I-V data from fabricated armchair and zigzag devices, fit high-order polynomials to emulate diode behavior in software, and train KANs on vision benchmarks using these learned basis functions. Our results demonstrate that KANalogue can approximate complex functions with minimal parameters while maintaining classification accuracy competitive with digital baselines. This work bridges device-level physics and function approximation theory, charting a path toward scalable, energy-efficient analogue machine learning systems. keywords Analogue In-Memory Computing, Negative Differential Resistance, Kolmogorov-Arnold Network † † equal-contributors: These authors contributed equally to this work. † † equal-contributors: These authors contributed equally to this work. 1 Introduction The growing demand for efficient machine learning on edge devices has intensified the search for alternatives to conventional digital architectures, which often struggle with power and latency constraints ambrogio2023analog ; wen2024fusion . Neuromorphic computing offer a compelling pathway, exploiting device physics to achieve orders-of-magnitude gains in energy efficiency and computational density mead1990neuromorphic ; kudithipudi2025neuromorphic . Recent advances in analogue in-memory computing (AIMC) based on resistive RAM (RRAM) and phase-change memory (PCM) have demonstrated near–software-level accuracy for large-scale matrix–vector multiplication buchel2024kernel ; huang2024memristor ; zhang2021computing ; LeGallo2023 ; Rasch2024 ; aguirre2024hardware ; Wu2024 ; lanza2025growing ; wang2023echo . By leveraging Ohm’s law and Kirchhoff’s current law, AIMC arrays physically co-locate memory and computation to perform these linear operations in situ , directly alleviating the von Neumann bottleneck that arises from shuttling data between separate memory and processing units sebastian2020memory ; xi2020memory ; seok2024beyond . Yet neural networks are not purely linear systems: they rely on alternating linear and nonlinear transformations. A critical obstacle remains because most analogue accelerators still delegate nonlinear activation functions to digital circuits, breaking the vision of fully analogue neural pipelines and limiting overall efficiency buchel2024kernel ; Yang2025 . Kolmogorov–Arnold Networks (KANs) provide a unique opportunity to address this gap. Inspired by the Kolmogorov–Arnold superposition theorem kolmogorov1957representations , KANs replace edge weights with learnable coordinate-wise basis functions. This structural distinction shifts expressive capacity from dense matrix multiplication to univariate nonlinear transformations, enabling compact models that differ fundamentally from traditional multilayer perceptrons liu2025kan . Crucially, it also makes KANs particularly suitable for analogue realization: each basis function can, in principle, be instantiated directly by a physical nonlinear device. Combined with their reduced parameter counts compared to conventional networks, KANs offer a smaller hardware footprint and lower energy demands. Recent variants have been applied to symbolic regression sidharth2024 , modeling of physical systems aghaei2025fkan ; bozorgzadeh2024wavelet ; toscano2024kkan , remote-sensing image classification cheon2024remote , and molecular property prediction li2025kolmogorov . Despite these advantages, KANs have so far been realized only in digital or hybrid platforms huang2025hardware ; peng2024photonic , leaving their analogue potential largely unexplored. Nonlinear devices with negative differential resistance (NDR) offer a promising route to close this gap. Esaki diodes esaki1958new , resonant tunneling diodes tsu1973tunneling ; chang1974resonant , and recently designed cold-metal devices theoretical ; Bodewei2024 naturally exhibit nonlinear current–voltage characteristics that resemble activation functions. Prior work has demonstrated the use of NDR devices to implement spiking neurons Donati2024 , bistable logic gates McNaughton2025 , or to emulate software-defined nonlinearities in memristor crossbars Yang2025 . However, integrating device-level nonlinearities directly into function-based architectures has yet to be realized. Here we introduce KANalogue, a framework that implements Kolmogorov–Arnold Networks using NDR devices as physical instantiations of univariate basis functions. NDR devices exhibit inherently nonlinear, non-monotonic current–voltage characteristics that provide rich expressive capability. Moreover, multiple NDR elements can be composed to realize more complex nonlinear functions, further extending the representational capacity of analogue nodes. This property allows KAN nodes to approximate complex mappings with fewer units, improving node efficiency and enabling more compact network realizations. By mapping these device characteristics to the functional components of KANs, KANalogue establishes a direct bridge between quantum electronic behavior and neural computation. Through simulation and surrogate modeling based on cold-metal tunnel diodes, we demonstrate that NDR-enabled KANs achieve competitive accuracy on standard benchmarks while reducing parameter counts and eliminating digital nonlinearity computation. When scaled to larger systems, such analogue implementations naturally exhibit statistical robustness: variations and noise across individual devices are averaged out through distributed computation, imparting a degree of error tolerance to the overall network. This work integrates theoretical function approximation with device-level physics, pointing toward scalable, fully analogue neural networks grounded in emerging, manufacturable device technologies. 2 Method 2.1 Tunnel Diodes Simulations For the device modeling, we constructed atomically thin heterostructures in which \text{NbSi}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{N}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}} monolayers serve as the electrodes and a \text{HfSi}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{N}{\vphantom{\text{X}}}_{\smash[t]{\text{4}}} monolayer acts as the tunnel barrier. Both armchair and zigzag orientations were considered, and barrier thicknesses of 15 Å and 21 Å were tested to examine orientation- and geometry-dependent transport. A vacuum spacing of 20 Å was added in the out-of-plane direction to eliminate spurious interactions between periodic replicas. These geometries were subsequently used as input for first-principles calculations. The cold mental tunnel diodes are designed To calculate the I–V characteristics, we performed first-principles calculations using DFT within the linear combination of atomic orbitals (LCAO) approach. A double-zeta polarized basis set and norm-conserving FHI pseudopotentials were employed. The exchange-correlation energy was treated using the Perdew–Burke–Ernzerhof (PBE) form of the generalized gradient approximation (GGA). The self-consistent DFT calculations are performed using a 24\times 24\times 1 k-point grid with a density mesh cut-off of 45 Hartree and the total energy converged to at least 10 -4 Hartree. To eliminate spurious interactions between periodic images, a vacuum spacing of 20\text{\,}\mathrm{\SIUnitSymbolAngstrom} was added in the out-of-plane direction. Transport properties were computed using DFT in combination with the nonequilibrium Green’s function (NEGF) formalism. The DFT-NEGF calculations utilized a 24\times 1\times 184 k-point grid for the armchair orientation and a 14\times 1\times 318 grid for the zigzag orientation of the lateral tunnel diodes. The current-voltage (I–V) characteristics were evaluated using the Landauer formula: I(V)=\frac{2e}{h}\int T(E,V)[f_{L}(E,V)-f_{R}(E,V)]dE (1) where T(E,V) denotes the transmission coefficient, and f_{L} , f_{R} are the Fermi-Dirac distributions of the left and right electrodes, respectively. A temperature of 300 K was maintained throughout all calculations. Besides, to improve the band gap of the tunnel barrier HfSi 2 N 4 , we employed a PBE+U method with a U value of 8 eV for the d orbitals of Hf in the transport calculations. 2.2 Kolmogorov–Arnold Network architecture with tunnel-diode basis functions To construct the neural architecture, we implemented a variant of Kolmogorov–Arnold Networks (KANs) that replaces the conventional spline- or polynomial-based basis functions with nonlinear responses derived from tunnel-diode current–voltage (I–V) characteristics. Each KAN layer

Read the complete text