A Survey of Neuromorphic Computing and Neural Networks in Hardware

I. INTRODUCTION

• The survey reviews more than 3,000 papers across 35 years to synthesize neuromorphic models, algorithms, learning, hardware, applications, materials, and supporting systems.It aims to provide a broad historical perspective and identify open issues for future research.Source: §I · p. 1
• Neuromorphic computing emerged as a complementary, biologically inspired alternative to von Neumann systems, motivated by differences in organization, power, and processing capability.The term was coined by Carver Mead in 1990 and now includes biologically inspired or artificial neural networks in non-von Neumann architectures.Source: §I · p. 1
• Highly connected, parallel architectures that collocate memory and processing promise faster, more power-efficient, and smaller-footprint computation while addressing the von Neumann bottleneck.Interest increased alongside the approaching end of Moore’s law and rising power demands associated with Dennard scaling.Source: §I · pp. 1–2
• Machine learning provides a second major motivation through potential improvements on some tasks and support for online, real-time learning resembling biological brains.The survey presents neuromorphic architectures as promising platforms for future machine-learning algorithms.Source: §I · p. 2
• The field spans materials science, neuroscience, electrical and computer engineering, and computer science, linking device fabrication, biological modeling, circuitry, algorithms, and supporting software.The paper organizes its review around models, algorithms and learning, hardware, supporting components, applications, and future challenges.Source: §I · p. 2

II. MOTIVATION

• Early neuromorphic hardware emphasized many simple, densely interconnected components to exploit biological-style parallelism, faster computation, and real-time performance in applications such as control and image reconstruction –.Fault tolerance and adaptation were also attractive because they could mitigate device variation and hardware errors,.Source: §II · pp. 2–3
• Low-power operation became the dominant contemporary motivation, complemented by small device footprints, continued interest in speed and real-time performance, and hardware-error tolerance.The human brain’s approximately 20-watt power consumption has been a key inspiration,.Source: §II · p. 3
• Recent motivations also include neuroscience simulation, scalability to large neuron and synapse populations, mitigation of Moore’s-law and memory-processing limits, and low-power online learning.Online learning targets unsupervised adaptation to changing tasks, while custom systems enable neuroscience simulations infeasible on traditional supercomputers in scale, speed, or power,.Source: §II · pp. 3–4

Model foundations

• Model selection determines the neurons, synapses, component behavior, and interactions that constitute a neuromorphic network.The section covers neuron, synapse, and network models spanning biologically inspired to computationally driven designs.Source: §III · p. 4

Application-driven selection

• Application requirements guide model choice: biologically realistic models support brain simulation, whereas convolutional networks may suit high-accuracy image recognition.The appropriate model depends partly on whether the objective is neuroscience simulation or computational performance on a recognition task.Source: §III · p. 4

Hardware-driven selection

• Device and material characteristics also constrain model choice; memristors support spike-timing-dependent-plasticity-like mechanisms suited to spiking neural networks.In many cases, the appropriate model or complexity level remains unclear.Source: §III · p. 4

Model foundations

• Neuron models in neuromorphic systems generally accumulate charge and fire, but their mechanisms and biological fidelity vary substantially.Biologically implausible models typically omit axons and dendrites, with some exceptions.Source: A. Neuron Models · p. 4

Hardware model taxonomy

• Hardware implementations are organized into five categories: biologically-plausible, biologically-inspired, Neuron+Other, integrate-and-fire, and McCulloch-Pitts models.Neuron+Other models may include axons, dendrites, or glial cells; integrate-and-fire models are simpler biologically-inspired spiking models; McCulloch-Pitts models derive from the original McCulloch-Pitts neuron.Source: A. Neuron Models · pp. 4–5

McCulloch-Pitts models

• McCulloch-Pitts derivatives compute each neuron’s output from an activation function applied to weighted inputs from N presynaptic neurons.The notation defines y_j as output, f as activation function, w_i,j as synaptic weight, and x_i as the preceding neuron’s output.Source: A. Neuron Models · p. 5

Model components

• Implemented spiking models may include membrane leakage, ion-channel dynamics, axonal delays, and dendritic effects on presynaptic and postsynaptic inputs.Izhikevich provides an overview of these spiking-neuron model types.Source: A. Neuron Models · p. 5

1) Biologically-Plausible:

• Hodgkin-Huxley is the most popular biologically plausible model, while Morris-Lecar provides a simpler two-dimensional alternative.Hodgkin-Huxley uses four-dimensional nonlinear differential equations to model ion transfer, whereas Morris-Lecar reduces the formulation to two dimensions.Source: 1) Biologically-Plausible · p. 5
• Simplified Hodgkin-Huxley variants, including Fitzhugh-Nagumo and Hindmarsh-Rose, reduce computational and parameter complexity but emphasize behavior over physical biological activity.Their simpler computation can enable smaller, more efficient hardware implementations, while fewer parameters can simplify setting or training.Source: 2) Biologically-Inspired · p. 5
• Izhikevich, Mihalas-Niebur, and quartic models reproduce biologically relevant bursting or spiking behavior with simpler or linear-differential-equation-based formulations.Izhikevich is especially popular because it combines computational simplicity with biologically accurate behavior; Mihalas-Niebur uses linear differential equations, while the quartic model uses two nonlinear equations.Source: 1) Biologically-Plausible · p. 5

3) Neuron + Other Biologically-Inspired Mechanism:

• Other biologically inspired mechanisms add membrane, ion-channel, axon, dendrite, or glial-cell detail, while some models are adapted specifically for low-power VLSI, FPGA, static CMOS, MOSFET, or single-electron hardware.These models may update or abstract established neuron equations to ease implementation or accommodate new hardware developments.Source: 3) Neuron + Other… · p. 5

4) Integrate-and-Fire Neurons:

• Integrate-and-fire models range from basic charge accumulation to more complex nonlinear and adaptive variants, offering useful spiking behavior with generally lower biological realism.Leaky integrate-and-fire adds temporal potential decay and is one of the most widely used neuromorphic neuron models; quadratic and adaptive exponential variants add further complexity.Source: 4) Integrate-and-Fire Neurons · pp. 5–6
• Digital spiking neurons replace differential-equation dynamics with cellular-automaton-based dynamics and include hybrid analog/digital, resonate-and-fire, rotate-and-fire, asynchronous, and pulse-coupled implementations.Hardware implementations also include neurons for random neural networks.Source: 4) Integrate-and-Fire Neurons · p. 6
• Spiking neural networks may use any surveyed neuron model, and programmable hardware such as SpiNNaker can realize different neuron models within one neuromorphic implementation.The review uses “spiking neural network” for full network models without specifying a particular neuron model.Source: 4) Integrate-and-Fire Neurons · p. 6

5) McCulloch-Pitts Neurons:

• McCulloch-Pitts hardware commonly uses perceptron thresholding, while sigmoid and hyperbolic-tangent activations dominate alternatives that trade hardware complexity against accuracy and usefulness.Other implementations include ramp-saturation, linear, piecewise-linear, step, multi-threshold, radial-basis, tangent-sigmoid, and periodic functions.Source: 5) McCulloch-Pitts Neurons · p. 6
• Activation-function circuits may implement both a function and its derivative for back-propagation, with some designs offering programmable activations or neuron-building blocks.Both sigmoid and hyperbolic-tangent derivatives have been implemented in hardware.Source: 5) McCulloch-Pitts Neurons · p. 6
• Hardware neurons also implement binary, fuzzy, and Hopfield models, reflecting a user tradeoff between computational complexity and biological inspiration.Figure 5 qualitatively compares neuron models using those two factors.Source: 5) McCulloch-Pitts Neurons · p. 6, Figure 5 · p. 6

B. Synapse Models

• Synapses are often the most abundant or area-intensive neuromorphic element, so implementations commonly prioritize compactness, while plasticity changes synaptic strength or weight over time.Biologically inspired systems may instead model chemical interactions, ion pumps, neurotransmitters, or conductance-based ion channels.Source: B. Synapse Models · pp. 6–7
• Spiking synapses commonly implement potentiation, depression, and STDP, while also modeling synaptic dynamics or homeostasis to stabilize network activity.These mechanisms can represent short- and long-term biological plasticity and regulate outgoing spikes or postsynaptic potentials.Source: B. Synapse Models · p. 7
• Non-spiking synapses support feed-forward, winner-take-all, and convolutional networks, using learning rules such as Hebbian, least-mean-square, and Gaussian mechanisms for back-propagation.These implementations cover both network-specific synaptic structures and learning-rule adaptations.Source: B. Synapse Models · p. 7

C. Network Models

• Network selection depends on biological inspiration, neuron and synapse complexity, topology, hardware connectivity, training feasibility, and application applicability.Figure 6 illustrates network topologies and emphasizes that matching required connectivity to feasible hardware is non-trivial.Source: C. Network Models · p. 7, Figure 6 · p. 8
• Spiking and feed-forward implementations are by far the most common, with spiking activity rising in the last decade and convolutional-neural-network activity increasing recently.General feed-forward activity had begun to taper off before convolutional neural networks increased activity alongside deep learning.Source: C. Network Models · p. 8

• Model choice depends heavily on the neuromorphic system’s intent: neuroscience projects favor biologically plausible or inspired models, whereas application systems favor proven artificial neural networks.Examples include image processing on remote sensors and autonomous robots.Source: D. Summary and Discussion · p. 9
• Hardware characteristics and efficiency also shape model selection, including STDP-compatible models for memristors and event-driven spiking neural networks.Source: D. Summary and Discussion · p. 9
• Most neural-network models have been implemented in hardware at some point in their history.The surveyed implementation landscape spans a wide variety of model types.Source: D. Summary and Discussion · pp. 8–9

IV. ALGORITHMS AND LEARNING

• Algorithm selection is constrained by neuron, synapse, network, and topology choices, while learning may be implemented on-chip, off-chip, or through chip-in-the-loop methods.The section focuses primarily on algorithms tailored directly to hardware, including on-chip and chip-in-the-loop approaches.Source: §IV · p. 9

A. Supervised Learning

• Back-propagation is the most commonly used supervised algorithm across feed-forward, recurrent, spiking, and convolutional neural networks.Implementations range from optimized off-line host training to on-chip and chip-in-the-loop training adapted to device characteristics such as component variation.Source: A. Supervised Learning · p. 9
• Back-propagation is restrictive for some neuron models, network models, and topologies, and can be difficult or costly to implement in hardware.Other on-chip supervised approaches include least-mean-squares, weight perturbation, and mechanisms specialized for particular model types.Source: A. Supervised Learning · pp. 9–10
• Nature- and evolution-inspired supervised methods support hardware-aware optimization because they need not depend on particular model characteristics and can evaluate solutions through chip-in-the-loop training.Reported implementations include differential evolution, evolutionary or genetic algorithms, and particle swarm optimization, both off-chip and in hardware or companion hardware.Source: A. Supervised Learning · p. 10

B. Unsupervised Learning

• On-chip online unsupervised training is presented as likely necessary to realize the full potential of neuromorphic implementations.Early implementations used self-organizing maps or related self-organizing learning rules.Source: B. Unsupervised Learning · p. 10
• Hebbian learning rules are popular online mechanisms, spanning variations that support both supervised and unsupervised learning.The cited work describes a broad family of Hebbian-type rules implemented in neuromorphic systems.Source: B. Unsupervised Learning · p. 10
• Spike-timing-dependent plasticity is identified as perhaps the most popular online unsupervised learning mechanism in neuromorphic systems.The supplied passage begins its extensive list of implementations and variants but does not provide further comparative results.Source: B. Unsupervised Learning · p. 10

Biologically-Inspired

• STDP increases or decreases synaptic weight according to the timing of pre- and post-synaptic spikes, but its broad applicability remains unestablished.Shorter intervals between spikes produce larger weight changes, and custom potentiation and depression circuits have been developed for biologically inspired implementations.Source: Biologically-Inspired · p. 11

C. Summary and Discussion

• Spiking neuromorphic systems offer event-driven power and energy efficiency with closer biological inspiration, but their full capabilities require new training and learning paradigms.Existing non-spiking models provide practical applications, yet the survey argues that hardware itself should be exploited to train larger-scale spiking networks more efficiently than von Neumann systems.Source: C. Summary and Discussion · p. 11
• Algorithm selection depends on the model, device, learning location and timing, convergence speed, applicability, and desired biological inspiration or plausibility.Back-propagation is broadly applicable and fast but device- and model-restricted; evolutionary methods are flexible but can converge slowly, while Hebbian learning and STDP are biologically attractive yet not widely demonstrated.Source: C. Summary and Discussion · p. 11
• Neuromorphic algorithm development remains incomplete because STDP is common in spiking systems, many systems omit training rules, and conventional algorithms target von Neumann architectures.Progress requires algorithm developers to account for device capabilities and collaborate with hardware developers and neuroscientists to integrate learning into future systems.Source: C. Summary and Discussion · p. 11

V. HARDWARE

• Neuromorphic hardware is surveyed as digital, analog, or mixed analog/digital, with the analysis emphasizing system-level implementations over unusual device components.This categorization provides the high-level structure for examining hardware platforms and their device-level components.Source: §V · p. 12

A. High-Level

• Analog systems compute through physical device properties and continuous values, whereas digital systems use discrete values and Boolean logic, although neuromorphic timing conventions often overlap.Analog hardware is generally noisier, but neural-network robustness to noise and faults has motivated arguments for analog implementations.Source: A. High-Level · p. 12

1) Digital:

• FPGAs support reconfigurable neuromorphic models, topologies, and algorithms and can accelerate software simulations, but they are generally unsuitable for the smallest, lowest-power systems.They offer shorter design and fabrication time, reuse across applications, problem-specific optimization, and straightforward host-computer interfacing; some implementations instead use them as stepping stones toward custom chips.Source: 1) Digital · p. 12
• 25 pJ per connection for TrueNorth versus 10 nJ for SpiNNaker illustrates the energy-efficiency cost of SpiNNaker’s greater model, topology, and learning flexibility.TrueNorth uses a fixed leaky integrate-and-fire model, limited programmable connectivity, and no on-chip learning, whereas SpiNNaker supports flexible neuron, synapse, topology, and learning configurations.Source: 1) Digital · p. 13

2) Analog:

• Analog systems include programmable and custom implementations, including FPAAs, field programmable neural arrays, and NeuroFPAAs with programmable neuron and synapse components.Custom neuromorphic arrays differ from general-purpose FPAAs by targeting neural and synaptic circuitry.Source: 2) Analog · p. 13
• Analog circuitry aligns with biological neural characteristics such as charge conservation, amplification, thresholding, and integration, while spiking networks can address asynchrony, noise, and unreliability.The paper identifies spiking neural networks as natural analog applications because they operate asynchronously and tolerate noisy, unreliable components.Source: 2) Analog · p. 13
• Subthreshold circuitry is a common analog neuromorphic approach, typically selected for power efficiency and associated with the original definition of neuromorphic circuits.The survey also records a broad range of other analog neuromorphic implementations.Source: 2) Analog · p. 13

3) Mixed Analog/Digital:

• Mixed analog/digital neuromorphic systems are widespread because analog circuitry naturally implements neuron and synapse processing, while digital components can overcome analog unreliability.The survey identifies mixed systems as a common implementation family across neuromorphic platforms.Source: 3) Mixed Analog/Digital · p. 13
• Mixed analog/digital systems combine analog neuron and synapse processing with digital components that improve reliability, memory, communication, programmability, or learning.Digital memory commonly stores synaptic weights, digital spikes support communication, and digital components can implement programmability or learning mechanisms.Source: 3) Mixed Analog/Digital · p. 13
• BrainScaleS uses superthreshold operation to achieve a reported 10,000x speed-up over subthreshold Neurogrid, while both systems use digital communication.Neurogrid is primarily analog and operates in subthreshold mode; BrainScaleS is wafer-scale, includes analog components, and operates in superthreshold mode.Source: 3) Mixed Analog/Digital · p. 14

B. Device-Level Components

• Device-level components include memory technologies and optical elements, with Figure 11 summarizing their relative popularity in the neuromorphic literature.The figure’s box sizes represent the number of referenced works that include each component.Source: B. Device-Level Components · p. 14, Figure 11 · p. 14

1) Memristors:

• Memristors store history-dependent resistance and can reproduce synapse-like STDP, motivating their widespread use in neuromorphic systems.They have also attracted attention for energy-efficient circuitry, while neural-network fault tolerance can mitigate device variation.Source: 1) Memristors · p. 14
• Memristors most commonly implement synapses or crossbar arrays for weight storage, Hebbian learning, STDP, and other plasticity mechanisms.Multi-memristor designs can represent positive, negative, and zero weights, and specialized training algorithms address non-ideal device characteristics.Source: 1) Memristors · p. 14
• Memristors also implement complex or stochastic neurons and Hodgkin-Huxley axons, but nonideal behavior, sneak paths, geometry variation, and idealized theory limit practical reliability.Reported theory addresses stabilization, synchronization, and passivity, although these analyses commonly assume ideal memristor models.Source: 1) Memristors · pp. 14–15

2) CBRAM and Atomic Switches:

• CBRAM provides fast, nanoscale, low-power nonvolatile memory for neuromorphic synapses and neurons, while atomic switches form ion-mediated bridges for synaptic plasticity.CBRAM differs from ReRAM through its electrochemical connection formation and dissolution.Source: 2) CBRAM and Atomic Switches · p. 15

3) Phase Change Memory:

• Phase-change memory is mainly used for high-density synapses, STDP-capable elements, and synaptic weight storage, with some implementations combining neurons and synapses.Source: 3) Phase Change Memory · p. 15

4) Spin Devices:

• Spintronic devices offer tunable functionality, CMOS compatibility, nanoscale density, and implementations spanning neurons, online-learning synapses, and network modules.The surveyed devices include spin-transfer torque, spin-wave, and magnetic-domain-wall technologies.Source: 4) Spin Devices · p. 15

5) Floating Gate Transistors:

• Floating-gate transistors support analog, digital, and mixed-signal neuromorphic circuits, most often as analog memory for synaptic weights or parameters and as learning synapses.Learning mechanisms commonly include STDP.Source: 5) Floating Gate Transistors · p. 15
• Floating-gate transistors have also been used for linear-threshold and full-neuron implementations, dendrite models, and silicon-neuron firing-rate estimation.Source: 5) Floating Gate Transistors · p. 15

6) Optical:

• Optical and photonic platforms have reemerged because they offer ultrafast operation, moderate complexity, and programmability, supporting synapse and neuron implementations,.Earlier optical systems were limited by difficult storage despite inherent parallelism.Source: 6) Optical · p. 15

C. Materials for Neuromorphic Systems

• Atomic-switch and CBRAM behavior, stability, and reliability depend on material choice, making material selection application-dependent.Reported atomic-switch materials include Ag2S, Cu2S, Ta2O5, and WO3–x, while CBRAM uses several material combinations –,,,,,, –, –.Source: C. Materials for Neuromorphic Systems · p. 15
• Transition-metal-oxide memristors provide varied resistance states for stored weights, but their endurance, stability, and reliability differ across materials –.Examples include HfOx, TiOx, WOx, SiOx, TaOx, NiOx, FeOx, AlOx, HfOx/ZnOx, and PCMO.Source: C. Materials for Neuromorphic Systems · pp. 15–16
• Alternative memristor materials trade distinct properties: chalcogenides enable nanosecond-scale STDP, polymers offer low cost and tunable performance, and MgO magnetic devices have limited resistance ranges [2211], –.These materials have been proposed or used for neurons and synapses, with applicability depending on the required function.Source: C. Materials for Neuromorphic Systems · p. 16
• Ferroelectric three-terminal synaptic devices can realize learning processes such as STDP within the device itself, avoiding additional STDP circuitry,.Ferroelectric materials have been investigated for analog synaptic-weight memory and synaptic devices –.Source: C. Materials for Neuromorphic Systems · p. 16
• Graphene and carbon nanotubes support compact or biologically scaled neuromorphic components, while carbon nanotubes also enable interaction with living tissue for potential prosthetic applications –,.Carbon nanotubes have been proposed for dendrites, synapses, and spiking neurons; graphene has been used in transistors, resistors, and full synapses.Source: C. Materials for Neuromorphic Systems · p. 16
• Novel organic and inorganic transistors are pursued for low-cost, flexible neuromorphic components and for brain-machine interfaces, biosensors, and chemical or biological sensing –.The surveyed devices include silicon-based, oxide-based, organic electrochemical, organic nanoparticle, and polymer-membrane transistors.Source: C. Materials for Neuromorphic Systems · p. 16

D. Summary and Discussion

• Neuromorphic implementations span programmable FPGAs and FPAAs, custom digital, analog, and mixed-signal chips, diverse components, and materials that shape future system operation.Memristors are especially common, alongside CBRAM, phase-change memory, spin-based and optical components, and floating-gate transistors.Source: D. Summary and Discussion · p. 16

VI. SUPPORTING SYSTEMS

• Supporting systems are necessary for neuromorphic architectures to become feasible, usable, and accepted, with communication frameworks and software identified as key components.The survey notes that supporting software has received relatively limited attention despite its importance.Source: §VI · p. 16, B. Supporting Software · p. 17

A. Communication

• AER provides common asynchronous inter-chip communication, while SpiNNaker’s toroidal mesh enables tens of thousands of chips to simulate one network; BrainScaleS instead uses synchronous communication, –,,, –,,.Hierarchical AER and FPGA-based interfaces provide additional communication implementations –,,.Source: A. Communication · pp. 16–17
• On-chip communication uses AER, buses, bus-free designs, mesh, shared-bus, tree, point-to-point, multicast, ring, and feed-forward approaches, with network-on-chip multicast reported as highest-performing,,,,,,,,,, –.These alternatives reflect recognized limits of AER for intra-chip communication.Source: A. Communication · p. 17
• Three-dimensional integration supports neuromorphic pattern recognition, memory-processing stacks, and chip stacking through TSVs, microbumps, memristors, phase-change memory, and CMOL systems,,,, –,, –.TSVs can mitigate some neuromorphic integration issues such as parasitic capacitance.Source: A. Communication · p. 17

B. Supporting Software

• Hardware synthesis tools convert high-level descriptions into low-level, application-specific neural circuitry for particular neuromorphic systems and applications,, –.These tools operate within the constraints of a target neuromorphic system.Source: B. Supporting Software · p. 17
• Mapping tools adapt offline-trained neural networks to hardware restrictions, whereas programming tools configure architectures directly through parameters, topologies, custom training, Corelets, high-level languages, or instruction sets [414],,, –.Mapping accounts for connectivity restrictions and parameter bounds; Corelets provide reusable TrueNorth building blocks.Source: B. Supporting Software · pp. 17–18
• Software simulators support hardware verification, hardware-change testing, training-algorithm development, and user-base formation, while visualization tools expose system behavior from network-level to low-level hardware views,,,,,, –.Figure 12 illustrates complementary high-level spiking-network and low-level implementation visualizations,.Source: Figure 12 · p. 17, B. Supporting Software · p. 18

C. Summary

• Supporting software and communication systems remain underdeveloped relative to neuromorphic hardware, making co-development essential for practical utilization.The field particularly needs software tools developed alongside hardware to improve ease of use.Source: C. Summary · p. 18

VII. APPLICATIONS

• Neuromorphic systems have been applied broadly, but their “killer” applications remain undetermined, especially for spiking systems requiring further hardware and algorithm development.Figure 13 summarizes application types and their prevalence, while neuroscience simulation is explicitly excluded from this application overview.Source: §VII · p. 18, Figure 13 · p. 19
• Vision dominates sensory applications, alongside auditory, olfactory, and somatosensory systems, while medical and wearable uses exploit very small, low-power devices and spike-based biological communication.Applications include brain-machine interfaces, pacemakers, retina implants, fall detectors, and prosthetics.Source: §VII · pp. 18–19
• Robotics commonly uses neuromorphic systems for autonomous navigation, behavior learning, locomotion, social learning, target following, and central-pattern-generator-based motion.Small, power-efficient systems are especially useful for autonomous robots and oscillatory motions such as swimming or walking gaits.Source: §VII · p. 19
• Control applications benefit from real-time, compact, low-power, and temporal-processing requirements, with cart-pole or inverted-pendulum tasks serving as the most common test case.Recurrent connections and synaptic delays support these temporal control demands; neuromorphic systems have also been applied to Pong, PACMAN, and Flappy Bird.Source: §VII · p. 19
• Image, sound, and video applications span detection, classification, recognition, filtering, segmentation, speech, music, motion, tracking, and activity analysis, often leveraging temporal processing.Neuromorphic systems have evaluated MNIST, CIFAR10, SVHN, CalTech-101, and ImageNet, while recurrent or spiking models suit speech and music tasks with temporal or real-time requirements.Source: §VII · pp. 19–20, Figure 14 · p. 20
• Neuromorphic systems also support NLP, smart sensing, anomaly detection, general and medical data classification, and widely used benchmark datasets for comparison across systems and models.Applications include sentence processing, environmental and mobile sensing, cybersecurity, fault detection, diagnosis, physiological-signal analysis, and UCI or Proben1 benchmarks.Source: §VII · pp. 20–21
• Beyond neural-network applications, neuromorphic architectures have been used for graph algorithms, flock-motion simulation, and optimization, with simple temporal and logical tasks remaining common development benchmarks.Examples include maximum cut, graph coloring, traveling salesman, shortest path, two spirals, parity, function approximation, temporal patterns, and XOR.Source: §VII · p. 21

Cross-field coordination

• Neuromorphic computing could become a major part of future computing if research expands from materials through models and algorithms while coordinating innovation across these levels.Source: §VIII · pp. 21–22

Machine learning

• Learning research should develop neuromorphic training mechanisms beyond traditional back-propagation, using accelerated hardware exploration to improve spiking neural network methods and broaden applications.Source: §VIII · p. 21

Devices and materials

• New nanoscale technologies and composite materials could enable extremely small, ultra-fast devices with behaviors tailored specifically for neuromorphic computation.Source: §VIII · pp. 21–22

Software engineering

• Neuromorphic systems need dedicated supporting software and programming languages that expose characteristics such as massive parallelism and collocated memory and computation to non-experts.Source: §VIII · p. 22

End users and applications

• The field must communicate concrete use cases, including heterogeneous co-processors, IoT smart sensors, autonomous-vehicle controllers, and in situ satellite analysis platforms.Source: §VIII · p. 22

• Neuromorphic computing needs low-power, massively parallel, real-time, non-von Neumann architectures capable of potentially online training or learning.The field is expected to retain diverse models, spanning feed-forward networks to detailed biological neural-network emulators, because each model has distinct strengths and weaknesses.Source: §IX · p. 22
• Future work should develop training and learning algorithms specifically for neuromorphic architectures rather than adapting algorithms designed for different systems.The authors identify this as a particularly high-impact area for innovation, alongside continued development of hardware, devices, and materials.Source: §IX · p. 22
• Neuromorphic computing research also requires stronger supporting software and continued exploration of applications where these systems have already been successfully applied.The survey covers supporting systems, including software, while noting that relatively little software exists for the field.Source: §IX · p. 22
• The paper provides a comprehensive cross-disciplinary view intended to help researchers identify gaps, develop innovative systems, and consider neuromorphic computers for their own applications.The authors included all references in this version to support that broad overview and encourage further research.Source: §IX · pp. 22–23