Anthony Joseph Kenyon

Physical neural networks promise low-power machine learning by computing directly with analogue device physics, yet many architectures require nonlinear devices to act as scalar weights or as fixed, identical activation functions, discarding potential computational resources. Inspired by Kolmogorov--Arnold networks, we place a trainable nonlinear function on every connection, making each physical edge a learnable computational element. We realise these edges as analogue band-pass filters on field-programmable analogue arrays and transfer trained networks directly to hardware, with quantified fidelity across approximately 35,000 experimentally realised connections. On smooth, continuously valued tasks learned from imperfect signals---six-axis robotic kinematics from actuator-corrupted observations, continuous-action reinforcement learning, and photovoltaic maximum-power-point tracking---the networks match or exceed multilayer perceptrons with substantially fewer trainable parameters and far fewer physical connections. A dedicated CMOS implementation of the photovoltaic controller is projected to consume approximately 35 muW. Replacing the filters with simulated memristor circuits reproduces the advantage, showing that it follows from the architecture instead of one particular device. Making device nonlinearity trainable, instead of suppressing it, offers a route to compact analogue networks for the types of smooth, continuous inference and control tasks studied here.

Previous research on transistor gate oxides reveals a clear link between hydrogen content and oxide breakdown. This has implications for redox‐based resistive random access memory (ReRAM) devices, which exploit soft, reversible, dielectric breakdown, as hydrogen is often not considered in modeling or measured experimentally. Here quantitative measurements, corroborated across multiple techniques are reported, that reveal ReRAM devices, whether manufactured in a university setting or research foundry, contain concentrations of hydrogen at levels likely to impact resistance switching behavior. To the knowledge this is the first empirical measurement depth profiling hydrogen concentration through a ReRAM device. Applying a recently‐developed Secondary Ion Mass Spectrometry analysis technique enables to measure hydrogen diffusion across the interfaces of SiO_x ReRAM devices as a result of operation. These techniques can be applied to a broad range of devices to further understand ReRAM operation. Careful control of temperatures, precursors, and exposure to ambient during fabrication should limit hydrogen concentration. Additionally, using thin oxynitride or TiO₂ capping layers should prevent diffusion of hydrogen and other contaminants into devices during operation. Applying these principles to ReRAM devices will enable considerable, informed, improvements in performance.

Artificial Intelligence (AI) is currently experiencing a bloom driven by deep learning (DL) techniques, which rely on networks of connected simple computing units operating in parallel. The low communication bandwidth between memory and processing units in conventional von Neumann machines does not support the requirements of emerging applications that rely extensively on large sets of data. More recent computing paradigms, such as high parallelization and near-memory computing, help alleviate the data communication bottleneck to some extent, but paradigm- shifting concepts are required. Memristors, a novel beyond-complementary metal-oxide-semiconductor (CMOS) technology, are a promising choice for memory devices due to their unique intrinsic device-level properties, enabling both storing and computing with a small, massively-parallel footprint at low power. Theoretically, this directly translates to a major boost in energy efficiency and computational throughput, but various practical challenges remain. In this work we review the latest efforts for achieving hardware-based memristive artificial neural networks (ANNs), describing with detail the working principia of each block and the different design alternatives with their own advantages and disadvantages, as well as the tools required for accurate estimation of performance metrics. Ultimately, we aim to provide a comprehensive protocol of the materials and methods involved in memristive neural networks to those aiming to start working in this field and the experts looking for a holistic approach.

Recent years have seen a rapid rise of artificial neural networks being employed in a number of cognitive tasks. The ever‐increasing computing requirements of these structures have contributed to a desire for novel technologies and paradigms, including memristor‐based hardware accelerators. Solutions based on memristive crossbars and analog data processing promise to improve the overall energy efficiency. However, memristor nonidealities can lead to the degradation of neural network accuracy, while the attempts to mitigate these negative effects often introduce design trade‐offs, such as those between power and reliability. In this work, authors design nonideality‐aware training of memristor‐based neural networks capable of dealing with the most common device nonidealities. The feasibility of using high‐resistance devices that exhibit high I‐V nonlinearity is demonstrated—by analyzing experimental data and employing nonideality‐aware training, it is estimated that the energy efficiency of memristive vector‐matrix multipliers is improved by almost three orders of magnitude (0.715 TOPs⁻¹W⁻¹ to 381 TOPs⁻¹W⁻¹) while maintaining similar accuracy. It is shown that associating the parameters of neural networks with individual memristors allows to bias these devices toward less conductive states through regularization of the corresponding optimization problem, while modifying the validation procedure leads to more reliable estimates of performance. The authors demonstrate the universality and robustness of this approach when dealing with a wide range of nonidealities.

New computing technologies inspired by the brain promise fundamentally different ways to process information with extreme energy efficiency and the ability to handle the avalanche of unstructured and noisy data that we are generating at an ever-increasing rate. To realize this promise requires a brave and coordinated plan to bring together disparate research communities and to provide them with the funding, focus and support needed. We have done this in the past with digital technologies; we are in the process of doing it with quantum technologies; can we now do it for brain-inspired computing?

Columnar microstructures are critical for obtaining good resistance switching properties in SiO_x resistive random access memory (ReRAM) devices. In this work, the formation and structure of columnar boundaries are studied in sputtered SiO_x layers. Using TEM measurements, we analyze SiO_x layers in Me–SiO_x–Mo heterostructures, where Me = Ti or Au/Ti. We show that the SiO_x layers are templated by the Mo surface roughness, leading to the formation of columnar boundaries protruding from troughs at the SiO_x/Mo interface. Electron energy-loss spectroscopy measurements show that these boundaries are best characterized as voids, which in turn facilitate Ti, Mo, and Au incorporation from the electrodes into SiO_x. Density functional theory calculations of a simple model of the SiO₂ grain boundary and column boundary show that O interstitials preferentially reside at the boundaries rather than in the SiO₂ bulk. The results elucidate the nature of the SiO_x microstructure and the complex interactions between the metal electrodes and the switching oxide, each of which is critically important for further materials engineering and the optimization of ReRAM devices.

The limited sensitivity of existing analysis techniques at the nanometer scale makes it challenging to systematically examine the complex interactions in redox-based resistive random access memory (ReRAM) devices. To test models of oxygen movement in ReRAM devices beyond what has previously been possible, we present a new nanoscale analysis method. Harnessing the power of secondary ion mass spectrometry, the most sensitive surface analysis technique, for the first time, we observe the movement of ¹⁶O across electrically biased SiO_x ReRAM stacks. We can therefore measure bulk concentration changes in a continuous profile with unprecedented sensitivity. This reveals the nanoscale details of the reversible field-driven exchange of oxygen across the ReRAM stack. Both the reservoir-like behavior of a Mo electrode and the injection of oxygen into the surface of SiO_x from the ambient are observed within one profile. The injection of oxygen is controllable through changing the porosity of the SiO_x layer. Modeling of the electric fields in the ReRAM stacks is carried out which, for the first time, uses real measurements of both the interface roughness and electrode porosity. This supports our findings helping to explain how and where oxygen from ambient moisture enters devices during operation.