Modern computation based on von Neumann architecture is now a mature cutting-edge science. In the von Neumann architecture, processing and memory units are implemented as separate blocks interchanging data intensively and continuously. This data transfer is responsible for a large part of the power consumption. The next generation computer technology is expected to solve problems at the exascale with 1018 calculations each second. Even though these future computers will be incredibly powerful, if they are based on von Neumann type architectures, they will consume between 20 and 30 megawatts of power and will not have intrinsic physically built-in capabilities to learn or deal with complex data as our brain does. These needs can be addressed by neuromorphic computing systems which are inspired by the biological concepts of the human brain. This new generation of computers has the potential to be used for the storage and processing of large amounts of digital information with much lower power consumption than conventional processors. Among their potential future applications, an important niche is moving the control from data centers to edge devices. The aim of this roadmap is to present a snapshot of the present state of neuromorphic technology and provide an opinion on the challenges and opportunities that the future holds in the major areas of neuromorphic technology, namely materials, devices, neuromorphic circuits, neuromorphic algorithms, applications, and ethics. The roadmap is a collection of perspectives where leading researchers in the neuromorphic community provide their own view about the current state and the future challenges for each research area. We hope that this roadmap will be a useful resource by providing a concise yet comprehensive introduction to readers outside this field, for those who are just entering the field, as well as providing future perspectives for those who are well established in the neuromorphic computing community.

Purpose-led Publishing is a coalition of three not-for-profit publishers in the field of physical sciences: AIP Publishing, the American Physical Society and IOP Publishing.
Together, as publishers that will always put purpose above profit, we have defined a set of industry standards that underpin high-quality, ethical scholarly communications.
We are proudly declaring that science is our only shareholder.
ISSN: 2634-4386
Neuromorphic Computing and Engineering is a multidisciplinary, open access journal publishing cutting edge research on the design, development and application of artificial neural networks and systems from both a hardware and computational perspective. For detailed information about subject coverage see the About the journal section.
Stay informed about the latest journal news and announcements
Dennis V Christensen et al 2022 Neuromorph. Comput. Eng. 2 022501
Matteo Cucchi et al 2022 Neuromorph. Comput. Eng. 2 032002
This manuscript serves a specific purpose: to give readers from fields such as material science, chemistry, or electronics an overview of implementing a reservoir computing (RC) experiment with her/his material system. Introductory literature on the topic is rare and the vast majority of reviews puts forth the basics of RC taking for granted concepts that may be nontrivial to someone unfamiliar with the machine learning field (see for example reference Lukoševičius (2012 Neural Networks: Tricks of the Trade (Berlin: Springer) pp 659–686). This is unfortunate considering the large pool of material systems that show nonlinear behavior and short-term memory that may be harnessed to design novel computational paradigms. RC offers a framework for computing with material systems that circumvents typical problems that arise when implementing traditional, fully fledged feedforward neural networks on hardware, such as minimal device-to-device variability and control over each unit/neuron and connection. Instead, one can use a random, untrained reservoir where only the output layer is optimized, for example, with linear regression. In the following, we will highlight the potential of RC for hardware-based neural networks, the advantages over more traditional approaches, and the obstacles to overcome for their implementation. Preparing a high-dimensional nonlinear system as a well-performing reservoir for a specific task is not as easy as it seems at first sight. We hope this tutorial will lower the barrier for scientists attempting to exploit their nonlinear systems for computational tasks typically carried out in the fields of machine learning and artificial intelligence. A simulation tool to accompany this paper is available online7.
Ole Richter et al 2024 Neuromorph. Comput. Eng. 4 014003
With the remarkable progress that technology has made, the need for processing data near the sensors at the edge has increased dramatically. The electronic systems used in these applications must process data continuously, in real-time, and extract relevant information using the smallest possible energy budgets. A promising approach for implementing always-on processing of sensory signals that supports on-demand, sparse, and edge-computing is to take inspiration from biological nervous system. Following this approach, we present a brain-inspired platform for prototyping real-time event-based spiking neural networks. The system proposed supports the direct emulation of dynamic and realistic neural processing phenomena such as short-term plasticity, NMDA gating, AMPA diffusion, homeostasis, spike frequency adaptation, conductance-based dendritic compartments and spike transmission delays. The analog circuits that implement such primitives are paired with a low latency asynchronous digital circuits for routing and mapping events. This asynchronous infrastructure enables the definition of different network architectures, and provides direct event-based interfaces to convert and encode data from event-based and continuous-signal sensors. Here we describe the overall system architecture, we characterize the mixed signal analog-digital circuits that emulate neural dynamics, demonstrate their features with experimental measurements, and present a low- and high-level software ecosystem that can be used for configuring the system. The flexibility to emulate different biologically plausible neural networks, and the chip’s ability to monitor both population and single neuron signals in real-time, allow to develop and validate complex models of neural processing for both basic research and edge-computing applications.
Tianyi Liu et al 2025 Neuromorph. Comput. Eng. 5 032001
The field of neuromorphic tactile sensing aims to emulate the biological mechanisms of touch to enable artificial systems with efficiency, adaptability, and precision akin to natural tactile perception. Inspired by the spike-based data encoding of biological mechanoreceptors and neural processing, neuromorphic tactile sensors (NTSs) leverage event-driven architectures to handle sensory information through sparse, low-power, and efficient formats. This review explores the state of neuromorphic tactile sensing, emphasizing its biological foundations, sensor technologies and encoding techniques within the field of robotics. By bridging biological touch mechanisms with neuromorphic engineering, NTSs have the potential to enhance robotic manipulation, prosthetics, and human–machine interfaces. Challenges and future directions include developing novel materials for sensors, improving the performance of spiking neural networks and lowering the barrier to entry into neuromorphic touch research through open-sourcing code and datasets. This review underscores the potential of neuromorphic tactile sensing in creating highly efficient and versatile tactile systems for robotics and beyond.
Kevin Max et al 2025 Neuromorph. Comput. Eng. 5 034010
In this study, we explore how the combination of synthetic biology, neuroscience modeling, and neuromorphic electronic systems offers a new approach to creating an artificial system that mimics the natural sense of smell. We argue that a co-design approach offers significant advantages in replicating the complex dynamics of odor sensing and processing. We propose a hybrid system of synthetic sensory neurons that provides three key features: (a) receptor-gated ion channels, (b) interface between synthetic biology and semiconductors and (c) event-based encoding and computing based on spiking networks. Our approach is validated using simulation-based modeling of the complete sensing and processing pipeline. This research seeks to develop a platform for ultra-sensitive, specific, and energy-efficient odor detection, with potential implications for environmental monitoring, medical diagnostics, and security.
Charles P Rizzo et al 2025 Neuromorph. Comput. Eng. 5 024012
Neuromorphic computing is a novel style of computing that features low-power spiking neural networks (SNNs) as the main compute components. It is an event-driven computational paradigm that naturally pairs with event-based cameras and their asynchronous event output. In this work, we present NeuroPong, a novel closed-loop neuromorphic hardware system composed of an event-based camera, a neuromorphic system, and an Atari 2600 console. The system facilitates the implementation of SNN Atari agents capable of playing Atari games in real time using event camera capture as input. We perform a small parameter optimization experiment to examine how software agents translate to hardware, discuss some of the challenges intrinsic to the hardware system, and propose some future improvements of the system and its components.
Willian S Girão et al 2025 Neuromorph. Comput. Eng. 5 034004
Working memory serves as a crucial building block in cognitive systems due to its fundamental role in information processing and decision-making. Acting as a temporary storage and manipulation mechanism, working memory enables individuals to actively hold and manipulate relevant information essential for ongoing tasks. This cognitive function is pivotal for various complex processes, including problem-solving, language comprehension, and decision-making. This paper introduces a novel working memory model designed as a spiking recurrent neural network of excitatory and inhibitory neurons. The proposed model incorporates biological mechanisms that allows attractors to be active phasically, thereby reducing the energy budget associated with maintaining attractor states. The core innovation of the model lies in its ability to leverage phasic attractors for state-dependent computation in a probabilistic manner. By modulating the activity of attractors via synaptic delays, the model demonstrates context-sensitive information processing. The results showcase the efficiency gains achieved by the proposed phasic attractor mechanism and highlight the model’s capacity for flexible and adaptive information processing.
Sandeep Soni et al 2025 Neuromorph. Comput. Eng. 5 034001
This work proposes a Spintronics-based Hopfield oscillatory neural network (HONN) that leverages dynamic frequency-encoded electrical synchronization between two spin-torque vortex nano-oscillators (SVNOs) as oscillatory neurons, with a non-volatile memristor as a coupling element (synaptic connection). The frequency synchronization mechanism, inspired by the brain’s oscillatory dynamics, enables the synchronization of SVNOs, facilitating efficient information processing of the dynamic oscillatory signals within the network. This coupling mechanism has been investigated to design SVNOs-based neural circuit design topology for enhanced frequency-encoded computing using SVNOs neurons and memristive coupling synapses. The proposed transmission gate-based SVNO oscillatory neural circuit has been implemented, offering efficient frequency synchronization, non-linearity, and a less complex neural circuit design. Further, a hybrid Spintronic/complementary metal oxide semiconductor 16-SVNOs HONN is designed, and circuit-based simulations are performed, which offer a promising solution for building robust and scalable HONNs. We achieve fast computation (∼4 ns) and offer significantly lower energy consumption (∼24 fJ/neuron) as compared to VO2-based ONN architectures (8× faster and 4× reduced power/neuron). Finally, we demonstrate an image denoising application on the proposed SVNO-based HONN hardware-compatible accelerator using an image-splitting approach with parallel processing. The 32 × 32 street view house number image dataset is efficiently split into blocks and processed through the 16-SVNOs HONN design, dividing the image into 4 × 4 blocks. Lastly, we examined the peak signal-to-noise ratio and structural similarity index measure for denoising the images with an efficient splitting approach for scalability. The network effectively denoises images while maintaining image quality, demonstrating the potential of the HONN hardware-compatible architecture for large-scale and real-time applications.
Abhronil Sengupta and Asif I Khan 2025 Neuromorph. Comput. Eng. 5 030201
Giulia D’Angelo et al 2025 Neuromorph. Comput. Eng. 5 024019
Active vision enables dynamic and robust visual perception, offering an alternative to the static, passive nature of feedforward architectures commonly used in computer vision, which depend on large datasets and high computational resources. Biological selective attention mechanisms allow agents to focus on salient regions of interest (ROIs), reducing computational demand while maintaining real-time responsiveness. Event-based cameras, inspired by the mammalian retina, further enhance this capability by capturing asynchronous scene changes, enabling efficient, low-latency processing. To distinguish moving objects while the event-based camera is also in motion, the agent requires an object motion segmentation mechanism to accurately detect targets and position them at the centre of the visual field (fovea). Integrating event-based sensors with neuromorphic algorithms represents a paradigm shift, using spiking neural networks (SNNs) to parallelise computation and adapt to dynamic environments. This work presents a spiking convolutional neural network bioinspired attention system for selective attention through object motion sensitivity. The system generates events via fixational eye movements using a dynamic vision sensor integrated into the Speck neuromorphic hardware, mounted on a Pan–Tilt unit, to identify the ROI and saccade toward it. The system, characterised using ideal gratings and benchmarked against the event camera motion segmentation dataset, reaches a mean IoU of 82.2% and a mean structural similarity index of 96% in multi-object motion segmentation. Additionally, the detection of salient objects reaches an accuracy of 88.8% in office scenarios and 89.8% in challenging indoor and outdoor low-light conditions, as evaluated on the event-assisted low-light video object segmentation dataset. A real-time demonstrator showcases the system’s capabilities of detecting the salient object through object motion sensitivity in 0.124 s in dynamic scenes. Its learning-free design ensures robustness across diverse perceptual scenes, making it a reliable foundation for real-time robotic applications and serving as a basis for more complex architectures.
Media: The accompanying video can be found online7.
Kevin Max et al 2025 Neuromorph. Comput. Eng. 5 034010
In this study, we explore how the combination of synthetic biology, neuroscience modeling, and neuromorphic electronic systems offers a new approach to creating an artificial system that mimics the natural sense of smell. We argue that a co-design approach offers significant advantages in replicating the complex dynamics of odor sensing and processing. We propose a hybrid system of synthetic sensory neurons that provides three key features: (a) receptor-gated ion channels, (b) interface between synthetic biology and semiconductors and (c) event-based encoding and computing based on spiking networks. Our approach is validated using simulation-based modeling of the complete sensing and processing pipeline. This research seeks to develop a platform for ultra-sensitive, specific, and energy-efficient odor detection, with potential implications for environmental monitoring, medical diagnostics, and security.
Sidi Yaya Arnaud Yarga and Sean U N Wood 2025 Neuromorph. Comput. Eng. 5 034009
Enhancing speech in noisy environments is essential for applications like automatic speech recognition, hearing aids, and real-time voice interfaces, but remains challenging on low-power, always-on edge devices. Conventional systems rely on pulse code modulation (PCM) signals and artificial neural networks, both of which introduce significant preprocessing and computational overhead. In this work, we present PDMDNS, a novel end-to-end neuromorphic framework for real-time speech denoising that directly processes binary pulse density modulation (PDM) microphone output using a spiking neural network, entirely bypassing the conventional PDM-to-PCM conversion and preprocessing stages. PDMDNS simultaneously performs speech enhancement and signal format conversion, leveraging stateless spiking neurons to reduce computational cost while maintaining temporal modeling capabilities. Moreover, when evaluated on a dataset containing noisy signals with SNRs ranging from 20 dB to −5 dB, our system achieves an average improvement of +7 dB in SI-SNR and a +3% gain in STOI. Although this performance is slightly below the current state-of-the-art by less than 1 dB, PDMDNS requires only 33 M-Ops/s, which is nearly 3× fewer operations than the best-performing spiking models. While PDM signals require a trade-off between maximizing precision through high sampling rates and minimizing energy consumption with lower rates, PDMDNS demonstrates robust generalization across varying input sampling rates (−12.5% to +37.5%) without the need for retraining. This flexibility makes it a compelling solution for energy-efficient, low-latency speech processing in embedded and neuromorphic systems.
Tianyi Liu et al 2025 Neuromorph. Comput. Eng. 5 032001
The field of neuromorphic tactile sensing aims to emulate the biological mechanisms of touch to enable artificial systems with efficiency, adaptability, and precision akin to natural tactile perception. Inspired by the spike-based data encoding of biological mechanoreceptors and neural processing, neuromorphic tactile sensors (NTSs) leverage event-driven architectures to handle sensory information through sparse, low-power, and efficient formats. This review explores the state of neuromorphic tactile sensing, emphasizing its biological foundations, sensor technologies and encoding techniques within the field of robotics. By bridging biological touch mechanisms with neuromorphic engineering, NTSs have the potential to enhance robotic manipulation, prosthetics, and human–machine interfaces. Challenges and future directions include developing novel materials for sensors, improving the performance of spiking neural networks and lowering the barrier to entry into neuromorphic touch research through open-sourcing code and datasets. This review underscores the potential of neuromorphic tactile sensing in creating highly efficient and versatile tactile systems for robotics and beyond.
Théophile Rageau and Julie Grollier 2025 Neuromorph. Comput. Eng. 5 034008
Oscillator networks represent a promising technology for unconventional computing and artificial intelligence. Thus far, these systems have primarily been demonstrated in small-scale implementations, such as Ising machines for solving combinatorial problems and associative memories for image recognition, typically trained without state-of-the-art gradient-based algorithms. Scaling up oscillator-based systems requires advanced gradient-based training methods that also ensure robustness against frequency dispersion between individual oscillators. Here, we demonstrate through simulations that the Equilibrium Propagation algorithm enables effective gradient-based training of oscillator networks, facilitating synchronization even when initial oscillator frequencies are significantly dispersed. We specifically investigate two oscillator models: purely phase-coupled oscillators and oscillators coupled via both amplitude and phase interactions. Our results show that these oscillator networks can scale successfully to standard image recognition benchmarks, such as achieving nearly 98% test accuracy on the Modified National Institute of Standards and Technology (MNIST) dataset, despite noise introduced by imperfect synchronization. This work thus paves the way for practical hardware implementations of large-scale oscillator networks, such as those based on spintronic devices.
Hermann Folke Johann Rolf et al 2025 Neuromorph. Comput. Eng. 5 034007
In changing environments, the mammalian hearing perception so far outperforms technical speech processing. This is enabled by the nonlinear dynamics of the cochlea. Inside of it, the processing is based on a frequency decomposition and a sophisticated feedback loop, so that the amplification of spectrum of an external signal can vary from linear to compressive. This behavior can be mimicked by implementing a controllable Andronov–Hopf bifurcation into neuromorphic oscillators, which enables a compressive and frequency-selective response. However, the frequency decomposition with these neuromorphic oscillators has not been investigated yet. Here, we show that any oscillator, which exhibits an Andronov–Hopf bifurcation, has a unique response to external stimuli, if its bifurcation parameter is in a neighborhood of the critical point. In addition, we propose three different algorithms to enable the frequency decomposition by implementing the Fourier transform in an acoustic sensor. We found that this Fourier transform can be done by applying amplitude demodulation to the output of any oscillator exhibiting at least one Andronov–Hopf bifurcation and investigated the convergence time of the different algorithms. Our results demonstrate that the Fourier transform can be utilized for either a single oscillator, which is simple to implement, or an array of oscillators, which has a fast convergence time. Thereby, it is shown that the neuromorphic acoustic sensor consisting of these oscillators can both mimic the processing of the cochlea and be implemented into the technical unit. Finally, a potential experimental implementation using micro-electromechanical system sensors is proposed.
Tianyi Liu et al 2025 Neuromorph. Comput. Eng. 5 032001
The field of neuromorphic tactile sensing aims to emulate the biological mechanisms of touch to enable artificial systems with efficiency, adaptability, and precision akin to natural tactile perception. Inspired by the spike-based data encoding of biological mechanoreceptors and neural processing, neuromorphic tactile sensors (NTSs) leverage event-driven architectures to handle sensory information through sparse, low-power, and efficient formats. This review explores the state of neuromorphic tactile sensing, emphasizing its biological foundations, sensor technologies and encoding techniques within the field of robotics. By bridging biological touch mechanisms with neuromorphic engineering, NTSs have the potential to enhance robotic manipulation, prosthetics, and human–machine interfaces. Challenges and future directions include developing novel materials for sensors, improving the performance of spiking neural networks and lowering the barrier to entry into neuromorphic touch research through open-sourcing code and datasets. This review underscores the potential of neuromorphic tactile sensing in creating highly efficient and versatile tactile systems for robotics and beyond.
Jennifer Hasler and Arindam Basu 2025 Neuromorph. Comput. Eng. 5 012001
The effort addresses the research activity around the usage of non-volatile memories (NVM) for storage of ‘weights’ in neural networks and the resulting computation through these memory crossbars. In particular, we focus on the CMOS implementations of, and comparisons between, memristor/resistive random access memory (RRAM) devices, and floating-gate (FG) devices. A historical perspective for illustrating FG and memristor/RRAM devices enables comparison of nonvolatile storage (addressing issues related to resolution, lifetime, endurance etc), feedforward computation (different variants of vector matrix multiplication, tradeoffs between power dissipation and signal to noise ratio etc), programming (addressing issues of selectivity, peripheral circuits, charge trapping etc), and learning algorithms (continuous time LMS or batch update), in these systems. We believe this historical perspective is necessary and timely given the increasing interest in using computation in memory with NVM for a wide variety of memory bound applications.
Ali Safa 2024 Neuromorph. Comput. Eng. 4 042001
Recently, the use of bio-inspired learning techniques such as Hebbian learning and its closely-related spike-timing-dependent plasticity (STDP) variant have drawn significant attention for the design of compute-efficient AI systems that can continuously learn on-line at the edge. A key differentiating factor regarding this emerging class of neuromorphic continual learning system lies in the fact that learning must be carried using a data stream received in its natural order, as opposed to conventional gradient-based offline training, where a static training dataset is assumed available a priori and randomly shuffled to make the training set independent and identically distributed (i.i.d). In contrast, the emerging class of neuromorphic CL systems covered in this survey must learn to integrate new information on the fly in a non-i.i.d manner, which makes these systems subject to catastrophic forgetting. In order to build the next generation of neuromorphic AI systems that can continuously learn at the edge, a growing number of research groups are studying the use of sparse and predictive Coding (PC)-based Hebbian neural network architectures and the related spiking neural networks (SNNs) equipped with STDP learning. However, since this research field is still emerging, there is a need for providing a holistic view of the different approaches proposed in the literature so far. To this end, this survey covers a number of recent works in the field of neuromorphic CL based on state-of-the-art sparse and PC technology; provides background theory to help interested researchers quickly learn the key concepts; and discusses important future research questions in light of the different works covered in this paper. It is hoped that this survey will contribute towards future research in the field of neuromorphic CL.
Hyeon Ji Lee et al 2024 Neuromorph. Comput. Eng. 4 032003
The growing demand for artificial intelligence has faced challenges for traditional computing architectures. As a result, neuromorphic computing systems have emerged as possible candidates for next-generation computing systems. Two-dimensional (2D) materials-based neuromorphic devices that emulate biological synapses and neurons play a key role in neuromorphic computing hardware due to their unique properties such as high strength, thermal conductivity, and flexibility. Although several studies have shown the simulations of individual devices, experimental implementation of large-scale crossbar arrays is still unclear. In this review, we explore the working principles and mechanisms of memristive devices. Then, we overview the development of neuromorphic devices based on 2D materials including transition metal dichalcogenides, graphene, hexagonal boron nitride, and layered halide perovskites. We also highlight the requirement and recent progress for building crossbar arrays by utilizing the advantageous properties of 2D materials. Lastly, we address the challenges that hardware implementation of neuromorphic computing systems currently face and propose a path towards system-level applications of neuromorphic computing.
Wei Wang et al 2024 Neuromorph. Comput. Eng. 4 032002
Crossbar arrays of memristors are promising to accelerate the deep learning algorithm as a non-von-Neumann architecture, where the computation happens at the location of the memory. The computations are parallelly conducted employing the basic physical laws. However, current research works mainly focus on the offline training of deep neural networks, i.e. only the information forwarding is accelerated by the crossbar array. Two other essential operations, i.e. error backpropagation and weight update, are mostly simulated and coordinated by a conventional computer in von Neumann architecture, respectively. Several different in situ learning schemes incorporating error backpropagation and/or weight updates have been proposed and investigated through neuromorphic simulation. Nevertheless, they met the issues of non-ideal synaptic behaviors of the memristors and the complexities of the neural circuits surrounding crossbar arrays. Here we review the difficulties and approaches in implementing the error backpropagation and weight update operations for online training or in-memory learning that are adapted to noisy and non-ideal memristors. We hope this work will be beneficial for the development of open neuromorphic simulation tools for learning-in-memory systems, and eventually for the hardware implementation of such as system.
Kaur et al
We present a novel skyrmion-based synaptic device featuring a multilayer structure of Ferromagnetic (FM)/Heavy metal (HM)/Ferroelectric (FE)/HM/FM, specifically (Co/Pt)n/(011)PMN-PT/(Co/Pt)n. The FE layer sandwiched between the two FM layers hosting the skyrmions enables electric field-induced strain-mediated modulation of perpendicular magnetic anisotropy (PMA) in the FM layers. This mechanism facilitates tunable skyrmion sizes, achieving continuous non-volatile conductance states due to remnant strain in the FE layer. The proposed device exhibits both synaptic potentiation and depression with the aid of differential MTJ readout unlike the prior skyrmion-based synaptic implementations and exhibits superior energy-efficiency compared to the other emerging non-volatile memory-based synaptic devices. Furthermore, a VGG-8 convolutional neural network utilizing the proposed synaptic element as weights achieves an accuracy of 90.39% after training on the CIFAR-10 dataset.
Şeker et al
The fields of machine learning and artificial intelligence drive researchers to explore energy-efficient, brain-inspired new hardware. Reservoir computing encompasses recurrent neural networks for sequential data processing and matches the performance of other recurrent networks with less training and lower costs. However, traditional software-based neural networks suffer from high energy consumption due to computational demands and massive data transfer needs. Photonic reservoir computing overcomes this challenge with energy-efficient neuromorphic photonic integrated circuits or NeuroPICs. Here, we introduce a reservoir NeuroPIC used for modulation format identification in C-band telecommunication network monitoring. It is built on a silicon-on-insulator platform with a 4-port reservoir architecture consisting of a set of physical nodes connected via delay lines. We comprehensively describe the NeuroPIC design and fabrication, experimentally demonstrate its performance, and compare it with simulations. The NeuroPIC incorporates non-linearity through a simple digital readout and achieves close to 100% accuracy in identifying several configurations of quadrature amplitude modulation formats transmitted over 20 km of optical fiber at a 32 GBaud symbol rate. The NeuroPIC performance is robust against fabrication imperfections like waveguide propagation loss, phase randomization, and delay line length variations. Furthermore, the experimental results exceeded numerical simulations, which we attribute to enhanced signal interference in the experimental NeuroPIC output. Our energy-efficient photonic approach has the potential for high-speed temporal data processing in various applications.
Narduzzi et al
Deploying energy-efficient deep neural networks on energy-constrained edge devices is an important research topic in both machine learning and circuit design communities. Both ANNs and SNNs have been proposed as candidates for these tasks and in particular SNNs have been proposed as energy-efficient because they leverage temporal sparsity in the outputs. However, existing computational frameworks fail to accurately estimate the cost of running sparse networks on modern time-stepped hardware, which exploits sparsity by skipping zero-valued operations. Meanwhile, weight sparsity-aware training remains underexplored for SNN and lacks systematic benchmarking against optimized ANN, making fair comparisons between the two paradigms difficult. To bridge this gap, we introduce EFLOP, a metric that accounts for the sparse operations during pre-activation updates of both ANN and SNN. Applying weight sparsity-aware training to both SNN and ANN, we achieve up to 8.9x reduction in EFLOP for GRU models and 3.6x for LIF models by sparsifying weights by 80%, without sacrificing accuracy on the SHD and SSC datasets. Our findings highlight the critical role of network sparsity in designing energy-efficient neural networks and establish EFLOP as a robust framework for cross-paradigm comparisons.