A neuron model defines how a single unit integrates its inputs and fires; an architecture defines how many such units are wired together into a network that computes something useful. The same design vocabulary developed for conventional artificial neural networks (feed-forward stacks, convolutions, recurrence, attention) has been carried over to the spiking domain, but each pattern must be adapted to two peculiarities of spiking neurons: information unfolds over a number of discrete timesteps rather than in a single forward pass, and the spike itself is a non-differentiable, all-or-nothing event. The architectures below span from the simplest layered networks to the spiking transformers now competing at the research frontier.
Feed-forward Networks
The simplest spiking architecture is a feed-forward stack: layers of spiking neurons (typically leaky integrate-and-fire units) connected in sequence with no recurrent loops, where activity propagates from one layer to the next over a fixed number of timesteps. Because the threshold (Heaviside) spike function has a derivative that is zero almost everywhere and undefined at threshold, these networks cannot be trained by naive backpropagation; they are instead trained either by converting a pre-trained conventional network into a spiking one, or by surrogate-gradient methods that substitute a smooth function for the spike derivative during the backward pass.
Feed-forward SNNs are the standard proving ground for static-image classification on datasets such as MNIST and Fashion-MNIST. They also expose the field's central trade-off: accuracy generally improves as more timesteps are integrated, but at a direct cost in latency and energy, which has motivated a large body of work on training accurate networks with as few timesteps as possible.
Convolutional Networks
Spiking convolutional networks replace the activations of a conventional convolutional layer with spiking neurons, producing spiking analogues of the VGG and ResNet backbones that dominate computer vision. These are the workhorses for the harder vision benchmarks: CIFAR-10/100, ImageNet and event-camera datasets such as CIFAR10-DVS and DVS-Gesture. Building them deep, however, is not simply a matter of stacking more layers.
A naively "spiking" ResNet suffers from spike degradation: once activations are reduced to binary spikes, the standard residual block can no longer implement the clean identity mapping that allows conventional ResNets to scale to hundreds of layers, so deeper spiking networks paradoxically train worse. Fang et al. (2021) addressed this with the Spike-Element-Wise (SEW) ResNet, which combines the spikes of the residual and shortcut paths with an element-wise function that provably restores identity mapping and overcomes vanishing and exploding gradients. SEW-ResNet was the first directly-trained SNN to exceed 100 layers, improving accuracy while using fewer timesteps on ImageNet, DVS-Gesture and CIFAR10-DVS.
Recurrent Networks
Adding recurrent, lateral or feedback connections lets a network hold state over time, making it suited to sequential and temporal-pattern tasks. Such networks are trained predominantly by surrogate-gradient backpropagation-through-time, in which the network is unrolled across its timesteps and gradients flow backward through the unrolled graph.
A key idea in recurrent SNNs is the adaptive leaky integrate-and-fire neuron, whose firing threshold rises slightly after each spike and decays slowly, endowing the neuron with a memory far longer than its membrane time constant. Bellec et al. (2018) combined regular and adapting spiking neurons into the Long Short-Term Memory Spiking Neural Network (LSNN), showing that this adaptation lets a recurrent SNN approach the performance of an LSTM on demanding temporal benchmarks, and that learning-to-learn protocols developed for LSTMs transfer to the spiking network.
Reservoir Computing and Liquid State Machines
Reservoir computing sidesteps the difficulty of training recurrent weights entirely. A liquid state machine, introduced by Maass, Natschläger and Markram (2002), is a fixed, randomly-connected recurrent reservoir of spiking neurons; only a simple (often linear) readout is trained, while the costly recurrent connections are left untouched. The reservoir projects a time-varying input into a high-dimensional transient state from which the readout extracts the answer.
The approach rests on firm theory: provided the reservoir has the separation property (distinct inputs drive distinguishable states) and the readout is a sufficiently good approximator, liquid state machines are universal for real-time computation on time-varying signals with fading memory, requiring no stable attractor states. In practice their performance is sensitive to reservoir hyperparameters, working best when the dynamics sit near the "edge of chaos". Because the expensive weights are never learned, reservoirs map naturally onto neuromorphic hardware.
Hybrid ANN–SNN Architectures
Hybrid designs pair a conventional artificial network with a spiking one (for example an ANN front-end for feature extraction feeding a spiking back-end), trading some of the energy advantage of a pure SNN for the accuracy and training stability of standard layers. The most widely used route to a spiking network is itself a hybrid procedure: ANN-to-SNN conversion, in which a conventional network is trained normally and its rectified-linear activations are then mapped onto the firing rates of spiking neurons, transferring the learned weights. Conversion reaches high accuracy but typically requires many timesteps, so mixed pipelines often convert first and then fine-tune with surrogate gradients to recover low-latency operation. Unlike the other patterns here, "hybrid" names a broad family of methods rather than a single canonical model.
Spiking Transformers
The most active current frontier is porting self-attention into the spiking domain while keeping the computation multiplication-free, so that the energy advantage of spikes is preserved. Zhou et al. (2022) introduced Spikformer and its Spiking Self-Attention, in which queries, keys and values are all spike trains and the softmax is dropped, turning attention into a lightweight, largely multiplication-free operation that outperformed earlier SNN frameworks on both static and neuromorphic image classification.
Yao et al. (2023) pushed this further with the Spike-Driven Transformer, whose attention uses only masking and addition (no multiplication at all) and which reached 77.1% top-1 accuracy on ImageNet-1K, a state of the art for spiking networks at the time. Follow-up designs such as Spikingformer have refined the residual pathways so that signals remain genuinely spike-based throughout the network.