Abstract This work presents a novel spiking neural network (SNN) decoding method, combining SNNs with hyperdimensional computing (HDC). This decoding method is designed to achieve high accuracy, high noise robustness, low inference latency and low energy consumption. Compared to analogous architectures decoded with existing approaches, the SNN-HDC model attains generally better classification accuracy, lower inference latency, lower spike count and lower estimated energy consumption on multiple test cases from the literature. The SNN-HDC achieved spike count reductions of 1.74 × to 3.36 × on the DvsGesture dataset and 1.36 × to 2.70 × on the SL-Animals-DVS dataset. The SNN-HDC achieved estimated energy consumption reductions of 1.24 × to 3.67 × on the DvsGesture dataset and 1.38 × to 2.27 × on the SL-Animals-DVS dataset. The proposed decoding method enables detection of classes unseen during training. On the DvsGesture dataset, the SNN-HDC model can detect 100% of samples from an unseen/untrained class. The findings suggest the proposed decoding method is a compelling alternative to both rate and latency decoding.

Abstract This paper introduces the ‘rebound winner-take-all (RWTA)’ motif as the basic element of a scalable neuromorphic control architecture. From the cellular level to the system level, the resulting architecture combines the reliability of discrete computation and the tunability of continuous regulation: it inherits the discrete computation capabilities of WTA state machines and the continuous tuning capabilities of excitable biophysical circuits. The proposed event-based framework addresses continuous rhythmic generation and discrete decision-making in a unified physical modeling language. We illustrate the versatility, robustness, and modularity of the architecture through the nervous system design of a snake robot.

Abstract This review examines the practical advantages of two-dimensional materials for energy-efficient in-memory computing by assembling a curated, experiment-only dataset covering 32 material systems across diverse device structures, mechanisms, and fabrication routes. Energy analysis was standardized using an averaged pulse-based metric, and key figures of merit—switching energy, on/off ratio, endurance, retention, and linearity—were compared against structural and mechanistic factors. Two low-energy-consumption design pathways emerge: ultrathin (<10 nm) two-terminal devices exploiting filament formation for sub- μ s updates and three-terminal heterojunction devices leveraging charge trapping to achieve nA-level programming currents over longer timescales. However, dynamic on/off ratios remain modest and are often overstated by DC sweep data. Endurance improves with shorter switching times, and the most intrinsically linear conductance evolution is observed in three-terminal gate-controlled devices employing charge trapping, Schottky barrier modulation, or ion intercalation. No universal optimum exists, as enhancing one performance metric typically compromises another. Based on the comparative analysis presented in this review, three near-term levers emerge as particularly relevant for translating selective material advantages into reproducible system-level gains: standardized pulsed benchmarking, scalable chemical vapor deposition growth with controlled defects and interfaces, and device–circuit co-design.

Abstract The introduction of large language models has significantly expanded global demand for computing; addressing this growing demand requires novel approaches that introduce new capabilities while addressing extant needs. Although inspiration from biological systems served as the foundation on which modern artificial intelligence (AI) was developed, many modern advances have been made without clear parallels to biological computing. As a result, the ability of techniques inspired by ``natural intelligence'' (NI) to inflect modern AI systems may be questioned. However, by analyzing remaining disparities between AI and NI, we argue that further biological inspiration can contribute towards expanding the capabilities of artificial systems, enabling them to succeed in real-world environments and adapt to niche applications. To elucidate which NI mechanisms can contribute toward this goal, we review and compare elements of biological and artificial computing systems, emphasizing areas of NI that have not yet been effectively captured by AI. We then suggest areas of opportunity for NI-inspired mechanisms that can inflect AI hardware and software.

Abstract The surrogate gradient descent algorithm enabled spiking neural networks to be trained to carry out challenging sensory processing tasks, an important step in understanding how spikes contribute to neural computations. However, it is unclear the extent to which these algorithms fully explore the space of possible spiking solutions to problems. We investigated whether spiking networks trained with surrogate gradient descent can learn to make use of information that is only encoded in the timing and not the rate of spikes. We constructed synthetic datasets with a range of types of spike timing information (interspike intervals, spatio-temporal spike patterns or polychrony, and coincidence codes). We find that surrogate gradient descent training can extract all of these types of information. In more realistic speech-based datasets, both timing and rate information is present. We therefore constructed variants of these datasets in which all rate information is removed, and find that surrogate gradient descent can still perform well. We tested all networks both with and without trainable axonal delays. We find that delays can give a significant increase in performance, particularly for more challenging tasks. To determine what types of spike timing information are being used by the networks trained on the speech-based tasks, we test these networks on time-reversed spikes which perturb spatio-temporal spike patterns but leave interspike intervals and coincidence information unchanged. We find that when axonal delays are not used, networks perform well under time reversal, whereas networks trained with delays perform poorly. This suggests that spiking neural networks with delays are better able to exploit temporal structure. To facilitate further studies of temporal coding, we have released our modified speech-based datasets.

Abstract The characteristic of our hearing is essentially based on the mechanics in our inner ear. Around 3000 hair cells in the cochlea decode vibrations into electrical signals, covering frequencies from 0.020-20 kHz with relative resolutions normalized by their natural frequency of 0.1-0.4% and a high dynamic range of 0-120 dB. These dynamic properties can be described by critical oscillators as they provide high resolution and nonlinear response near their critical points. However, the wide frequency range cannot be achieved as high sensitivity requires high Q-factors and is therefore associated with narrow frequency range. To overcome this, frequency tunability could be used to increase the detectable frequency range while maintaining high sensitivity. One solution to achieve frequency tuning is the mutual coupling of oscillators. To this end, a bio-inspired sensing system based on coupled resonators tuned near their critical points is presented, whose frequency can be tuned by varying the feedback of the individual resonator. In the coupled system three Andronov-Hopf bifurcations are identified, where two of them enable frequency tunability. We show that this adaptability of the frequency enables the coverage of a wide frequency range with limited number of resonators and yet preserves a high resolution with low number of resonators, which make them suitable for hardware implementation.

Abstract Radar-based human activity recognition (HAR) offers privacy and robustness over camera-based methods, yet remains computationally demanding for edge deployment. We present the first application of spiking neural networks (SNNs) for radar-based HAR on aircraft marshaling signal classification. Our novel hybrid architecture combines a pre-trained convolutional backbone for spatial feature extraction and leaky integrate-and-fire neurons for temporal processing, inherently capturing gesture dynamics. The model reduces trainable parameters by 88% with under 1% accuracy loss compared to existing state of the art methods, and generalizes well to the Soli gesture dataset. Through systematic comparisons with other three artificial neural network architectures, we demonstrate the trade-offs of spiking computation in terms of accuracy, latency, memory, and energy, establishing SNNs as an efficient and competitive solution for radar-based HAR.

Abstract Nanofluidic memristive systems exhibit the nonlinear behavior and the short-time plasticity needed for reservoir computing (RC) networks. They use ions as information carriers and operate in an electrochemical environment, in resemblance to the biological synapses. Here we present simulation results of an RC model implementation using a parallel array of memristive nanopores as reservoir. Each nanopore of the array is simulated under distinct chemical conditions using an experimentally justified theoretical model. We demonstrate the potential of the proposed network by performing three different RC tasks: sine wave nonlinear transformation, waveform classification, and forecasting of the Mackey–Glass chaotic time series.

Abstract Reservoir computing (RC) is a computational framework where a fixed dynamical reservoir projects an input into a higher-dimensional state that is then analyzed by a readout , which is trained to map the reservoir state into the desired output. While the conventional RC paradigm is based on dynamics of in-silico implemented recurrent neural networks, this computing paradigm can be efficiently implemented in hardware by exploiting dynamics of a wide range of physical systems in a paradigm denoted as Physical RC (PRC), attracting interest from a broader research community spanning from computer scientists to physicists, and material scientists. Here, we present RCbench, an open-source RC benchmark toolkit that implements a standardized and comprehensive suite for benchmarking computational reservoir models and physical implementations of RC. RCbench integrates widely recognized metrics such as Memory capacity, Nonlinear autoregressive moving average of order N, Kernel rank , and generalization rank, along with nonlinear transformation tasks. It also allows testing and comparing different readout algorithms, the evaluation of computational capabilities with diverse accuracy metrics, and includes feature selection methods to unravel the effect of specific reservoir outputs on computational performances. In particular, the toolkit enables easy benchmarking of PRC systems, providing a comprehensive benchmark tool that can be easily integrated with experimental data acquisition processes. By standardizing performance assessments, RCbench aims to facilitate inter-study comparisons and to accelerate the exploration, characterization and optimization of RC systems.

Abstract Analog in-memory computing (AIMC) offers a promising solution to the von Neumann bottleneck. However, deploying transformer models on AIMC remains challenging due to their inherent need for flexibility and adaptability across diverse tasks. For the benefits of AIMC to be fully realized, weights of static vector-matrix multiplications must be mapped and programmed to analog devices in a weight-stationary manner. This poses two challenges for adapting a base network to hardware and downstream tasks: (i) conventional analog hardware-aware (AHWA) training requires retraining the entire model, and (ii) reprogramming analog devices is both time- and energy-intensive. To address these issues, we propose AHWA low-rank adaptation (AHWA-LoRA) training, a novel approach for efficiently adapting transformers to AIMC hardware. AHWA-LoRA training keeps the analog weights fixed as meta-weights and introduces lightweight external LoRA modules for both hardware and task adaptation. We validate AHWA-LoRA training on SQuAD v1.1 and the GLUE benchmark, demonstrate its scalability to larger models, and show its effectiveness in instruction tuning and reinforcement learning. We further evaluate a practical deployment scenario that balances AIMC tile latency with digital LoRA processing using optimized pipeline strategies, with RISC-V-based programmable multi-core accelerators. This hybrid architecture achieves efficient transformer inference with only a 4% per-layer overhead compared to a fully AIMC implementation.