Neural hardware accelerators have demonstrated notable energy efficiency in tackling tasks, which can be adapted to artificial neural network (ANN) structures. Research is currently directed toward leveraging resistive random-access memories (RRAMs) among various memristive devices. In conjunction with complementary metal-oxide semiconductor (CMOS) technologies within integrated circuits (ICs), RRAM devices are used to build such neural accelerators. In this study, we present a neural accelerator hardware design and verification flow, which uses a lookup table (LUT)-based Verilog-A model of IHP’s one-transistor-one-RRAM (1T1R) cell. In particular, we address the challenges of interfacing between abstract ANN simulations and circuit analysis by including a tailored Python wrapper into the design process for resistive neural hardware accelerators. To demonstrate our concept, the efficacy of the proposed design flow, we evaluate an ANN for the MNIST handwritten digit recognition task, as well as for the CIFAR-10 image recognition task, with the last layer verified through circuit simulation. Additionally, we implement different versions of a 1T1R model, based on quasi-static measurement data, providing insights on the effect of conductance level spacing and device-to-device variability. The circuit simulations tackle both schematic and physical layout assessment. The resulting recognition accuracies exhibit significant differences between the purely application-level PyTorch simulation and our proposed design flow, highlighting the relevance of circuit-level validation for the design of neural hardware accelerators.

Nerve net simulators of C. elegans heavily support research of its nerve net functionality by offering the possibility to conduct digital experiments instead of real ones. However, current software tools are complex and difficult to use for non-programmers. With WDWorm, we offer a user-friendly toolbox with graphical user interface for simulating and experimenting with C. elegans’ nerve net. It does not require an installation and allows for several modifications of the nerve net, including parameter changes of each neuron and connection or the deactivation of individual neurons. Furthermore, a comparison with other software tools highlights that WDWorm is currently the most runtime-efficient approach for simulating and digitally experimenting with C. elegans . To invite other developers and researchers, we provide the source code in an open-access format under a CC-BY 4.0 Creative Commons license. The code is publicly available at https://github.com/dsacri/WDWorm .

Biological systems are a great source of inspiration for developing novel computing technologies, as they exhibit favourable properties such as self-organisation and energy efficiency. The sheer complexity of these systems, however, makes developing such technologies both difficult and time-consuming. Tools providing the ability of emulating the behavior of complex systems, therefore become very essential. In this work, we present a method for emulating the behavior of electrical networks containing a coupling network with two-port coupling elements. Here, we make use of the wave digital concept due to the robustness and parallelism of the associated algorithms, which are closely related to the underlying electrical circuit. This work should serve as a basis for future implementations of close-to-biology neural networks.

Hardware realizations of neuronal networks should also consider axon growth in addition to synaptic weight changes, because this accounts for an additional dynamic aspect due to the delayed signal transmission varying with the grown axon length. Our aim is to model axon growth by electrical circuits to enable a dynamic, self-organized topology formation, extending the state of the art hardware realizations. We start from a lossless transmission line whose underlying partial differential equations serve as a modeling of a locally distributed static axon. We then derive a memory-dependent axon model implementing a growth mechanism by utilizing the wave digital concept as a modeling tool based on mapping voltages and currents to wave quantities. The growth mechanism utilizes memristive Jaumann structures and can be seen as a fundamental building block enabling a spike-timing-dependent topology formation. Emulation results show that our approach offers an axon model with reconfigurable axon length capable of mimicking spike-timing-dependent axon growth. Moreover, we consider Pavlov's dog experiment to demonstrate the delay-based self-organized topology formation due to Hebbian learning in a small neuronal network.

Self-organizing circuits are subject to current research, especially self-organizing logic gates (SOLGs). The self-organizing aspect enables the latter to be operated `backwards', meaning that outputs can be used as inputs and vice versa. This makes SOLGs potential candidates to solve mathematically complex problems efficiently. Although SOLGs have been subject to up to date research, the investigated circuits center around purely mathematical models of memristors. In this work, we aim to solve the NP-complete subset sum problem with SOLGs based on a physical model of real memristors. The memristors of choice are RRAM-cells, as they provide a promising performance and a fast convergence due to their rapid switching behavior. For this purpose, we exploit the wave digital emulation technique which differs from many other emulation techniques by preserving passivity. It is shown that a simple subset sum problem is properly solved.

Optimal Wiener filtering is a popular method for the estimation of stationary processes which can be completely derived system-theoretically. Although there exist several optimal filtering concepts for non-stationary processes, there is still a lack of fundamental time-variant system theory that describes the problem statement for non-stationary process estimation. This paper provides an intuitively understandable theory for the fundamental optimal filtering concept. By interpreting cross- and autocorrelations as a time-variant impulse response of a linear system, the problem statement can be illustrated with a network of linear systems. This paper introduces a double periodic system model that approximates a time-variant transfer function in both time and frequency domain which leads to an analytic solution for the optimal time-variant filter. The presented results are generally applicable and degenerate for simplified process properties (e.g. stationarity) to the well known results. We also present how the problem statement can easily be extended due to the fundamental and uniform theoretical approach.

The degrees of freedom (DoF) of an X-network with M transmit and N receive nodes utilizing interference alignment with the support of J relays each equipped with Ljantennas operating in a half-duplex non-regenerative mode is investigated. Conditions on the feasibility of interference alignment are derived using a proper transmit strategy and a structured approach based on a Kronecker-product representation. The advantages of this approach are twofold: First, it extends existing results on the achievable DoF to generalized antenna configurations. Second, it unifies the analysis for time-varying and constant channels and provides valuable insights and interconnections between the two channel models. It turns out that a DoF ofNM/M+N-1 is feasible whenever the sum of the L2j≥ [N - 1][M - 1].

The degrees of freedom (DoF) of the MIMO Y-channel, a multi-way communication network consisting of 3 users and a relay, are characterized for arbitrary number of antennas. The converse is provided by cut-set bounds and novel genie-aided bounds. The achievability is shown by a scheme that uses beamforming to establish network coding on-the-fly at the relay in the uplink, and zero-forcing pre-coding in the downlink. It is shown that the network has min{2M2+2M3, M1+ M2+ M3,2N} DoF, where Mjand N represent the number of antennas at user j and the relay, respectively. Thus, in the extreme case where M1+M2+M3dominates the DoF expression and is smaller than N, the network has the same DoF as the MAC between the 3 users and the relay. In this case, a decode and forward strategy is optimal. In the other extreme where 2N dominates, the DoF of the network is twice that of the aforementioned MAC, and hence network coding is necessary. As a byproduct of this work, it is shown that channel output feedback from the relay to the users has no impact on the DoF of this channel.