[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-93886":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":9,"language":10,"languages":9,"totalLinesOfCode":9,"stars":11,"forks":12,"watchers":11,"openIssues":13,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":13,"stars7d":13,"stars30d":13,"stars90d":13,"forks30d":13,"starsTrendScore":13,"compositeScore":14,"rankGlobal":9,"rankLanguage":9,"license":15,"archived":16,"fork":16,"defaultBranch":17,"hasWiki":18,"hasPages":16,"topics":19,"createdAt":9,"pushedAt":9,"updatedAt":34,"readmeContent":35,"aiSummary":36,"trendingCount":13,"starSnapshotCount":13,"syncStatus":37,"lastSyncTime":38,"discoverSource":39},93886,"Awesome-Spiking-Neural-Networks-Hub","haoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub","haoran-zha","The comprehensive bilingual (EN\u002F中文) hub for Spiking Neural Networks — 340+ papers, models, neuromorphic hardware, datasets, tools & research groups.",null,"Python",106,1,0,40.9,"MIT License",false,"main",true,[20,21,22,23,24,25,26,27,28,29,30,31,32,33],"awesome","awesome-list","brain-inspired-computing","deep-learning","event-camera","machine-learning","neural-networks","neuromorphic-computing","neuromorphic-hardware","paper-list","snn","spiking-neural-networks","spiking-transformer","surrogate-gradient","2026-09-21 04:01:26","\u003Cdiv align=\"center\">\n\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhaoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub\">\u003Cimg src=\"assets\u002Fbanner.webp\" alt=\"Awesome Spiking Neural Networks Hub\" width=\"100%\">\u003C\u002Fa>\n\n\u003Ch1>Awesome Spiking Neural Networks Hub\u003C\u002Fh1>\n\n\u003Cp>\u003Cem>A comprehensive, deeply-annotated guide to the world of Spiking Neural Networks (SNNs)\u003C\u002Fem>\u003Cbr>\n\u003Csub>papers · models · neuromorphic hardware · datasets · tools · research groups\u003C\u002Fsub>\u003C\u002Fp>\n\n\u003Cp>\n\u003Ca href=\"https:\u002F\u002Fawesome.re\">\u003Cimg src=\"https:\u002F\u002Fawesome.re\u002Fbadge-flat2.svg\" alt=\"Awesome\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhaoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub\u002Fstargazers\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fstars\u002Fhaoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub?style=flat-square&logo=github&color=e3b341\" alt=\"Stars\">\u003C\u002Fa>\n\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fhaoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub\u002Fnetwork\u002Fmembers\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Fforks\u002Fhaoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub?style=flat-square&logo=github&color=8b949e\" alt=\"Forks\">\u003C\u002Fa>\n\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fgithub\u002Flast-commit\u002Fhaoran-zha\u002FAwesome-Spiking-Neural-Networks-Hub?style=flat-square&color=blue\" alt=\"Last commit\">\n\u003Ca href=\"#contributing\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPRs-welcome-brightgreen?style=flat-square\" alt=\"PRs Welcome\">\u003C\u002Fa>\n\u003Ca href=\"LICENSE\">\u003Cimg src=\"https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-MIT-yellow?style=flat-square\" alt=\"MIT\">\u003C\u002Fa>\n\u003C\u002Fp>\n\n\u003Cp>\u003Cb>English\u003C\u002Fb>  ·  \u003Ca href=\"README.zh-CN.md\">中文\u003C\u002Fa>  |  \u003Ca href=\"TIMELINE.md\">Timeline edition\u003C\u002Fa>\u003C\u002Fp>\n\n\u003Cp>\u003Cb>340+\u003C\u002Fb> papers & resources  ·  \u003Cb>40+\u003C\u002Fb> seminal works  ·  \u003Cb>42\u003C\u002Fb> research groups  ·  \u003Cb>45+\u003C\u002Fb> open-source projects\u003C\u002Fp>\n\n\u003Cp>\u003Cb>Jump to\u003C\u002Fb>  \n\u003Ca href=\"#start-here\">Start Here\u003C\u002Fa> ·\n\u003Ca href=\"#1--foundations--neural-coding\">Fundamentals\u003C\u002Fa> ·\n\u003Ca href=\"#3--training-methods\">Learning & Models\u003C\u002Fa> ·\n\u003Ca href=\"#6--neuromorphic-hardware\">Hardware\u003C\u002Fa> ·\n\u003Ca href=\"#7--applications\">Applications\u003C\u002Fa> ·\n\u003Ca href=\"#8--energy-robustness--security\">Topics\u003C\u002Fa> ·\n\u003Ca href=\"#10--datasets--benchmarks\">Resources\u003C\u002Fa> ·\n\u003Ca href=\"#13--research-groups--labs\">Groups\u003C\u002Fa>\u003C\u002Fp>\n\n\u003Csub>If this guide helps your work, please star the repo and \u003Ca href=\"#citation\">cite it\u003C\u002Fa> — contributions are very welcome.\u003C\u002Fsub>\n\n\u003C\u002Fdiv>\n\n---\n\n## What Is a Spiking Neural Network? (Read Me First)\n\nConventional deep networks (ANNs) pass **continuous numbers** between neurons at every layer, synchronously, every forward pass. A **Spiking Neural Network** instead communicates with **discrete, binary events — \"spikes\" — in time**, exactly like biological neurons. A spiking neuron integrates incoming current onto a *membrane potential*; when that potential crosses a threshold, it emits a single spike and resets. Nothing happens when there is no spike.\n\nThree consequences make SNNs compelling:\n\n- **Event-driven & sparse** → computation (and energy) is spent only when a spike occurs. On neuromorphic hardware this can mean *orders-of-magnitude* lower power than a GPU.\n- **Temporal by construction** → information lives not just in *how many* spikes fire but in *when* they fire, giving a natural substrate for time-series, audio, and event-camera data.\n- **Brain-inspired** → SNNs are often called the **\"third generation\"** of neural networks (after perceptrons and rate-based deep nets), bridging neuroscience and machine learning.\n\nThe central difficulty is that a spike is a **step function — non-differentiable** — so ordinary backpropagation does not directly apply. The whole field, in a sense, is a set of answers to *\"how do we train these things?\"* — which is why the [Training Methods](#3--training-methods) section is the heart of this list.\n\n> This is an **awesome-style hub for the whole SNN field** — papers, models, hardware, datasets, tools, and research groups. Each entry carries a one-line \"why it matters\" note (English here; 中文 in the [Chinese edition](README.zh-CN.md)) so you can navigate without drowning.\n\n---\n\n## Start Here\n\n**12 must-read landmarks that trace the field's arc** — from the \"third generation\" idea to today's spiking Transformers and large models. *(The ★ markers throughout the guide flag many more.)*\n\n| Year | Landmark | Venue | Link |\n|:---:|---|:---:|:---:|\n| 1997 | **Networks of Spiking Neurons: The Third Generation** — *the founding idea* | Neural Networks | [paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002FS0893-6080(97)00011-7) |\n| 2014 | **TrueNorth** — *a million-neuron neuromorphic chip* | Science | [paper](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.1254642) |\n| 2015 | **Unsupervised Learning with STDP** — *bio-plausible learning* | Front. Comput. Neurosci. | [paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffncom.2015.00099) |\n| 2018 | **Loihi** — *on-chip-learning neuromorphic processor* | IEEE Micro | [paper](https:\u002F\u002Fieeexplore.ieee.org\u002Fdocument\u002F8259423) |\n| 2018 | **STBP** — *spatio-temporal backprop, the training workhorse* | Front. Neurosci. | [paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffnins.2018.00331) |\n| 2020 | **e-prop** — *biologically plausible online learning* | Nature Comm. | [paper](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41467-020-17236-y) |\n| 2021 | **SEW-ResNet** — *directly training 100+ layer SNNs* | NeurIPS | [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2102.04159) |\n| 2022 | **QCFS** — *near-lossless ANN→SNN conversion* | ICLR | [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.04347) |\n| 2023 | **Spikformer** — *the first spiking Transformer* | ICLR | [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.15425) |\n| 2023 | **Spike-driven Transformer** — *pure spike-driven attention* | NeurIPS | [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.01694) |\n| 2024 | **SpikeGPT** — *the first generative spiking LLM* | TMLR | [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.13939) |\n| 2025 | **SpikingBrain** — *7B\u002F76B spiking brain-inspired LLM* | arXiv | [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.05276) |\n\n---\n\n## Recent Updates\n\n\u003Cdetails open>\n\u003Csummary>\u003Cb>Changelog\u003C\u002Fb> — newest first (click to collapse)\u003C\u002Fsummary>\n\n- **2026-07** — **Audit & ecosystem update.** Fixed dead\u002Foutdated links and a moved lab affiliation; added six 2026 works, four major labs, and per-lab **latest-work (2024–2026)** tags across §13.\n- **2026-07** — **Launched the Hub.** 340+ entries across **6 Parts**; new **Research Groups & Labs** (§13), **Model Zoo & Community** (§12), and a dedicated **Spiking Large Models & LLMs** section (§5).\n- **2026-07** — **Visual overhaul.** New banner, \"Start Here\" table, and datasets\u002Fframeworks\u002Fmodel-zoo turned into tables.\n- **2026-07** — **Coverage push.** Added SpikingBrain, Sorbet, SpikeCLIP, Spike2Former, SDiT; Darwin3, Intel Hala Point, IBM NorthPole, Lynxi \u002F BrainChip \u002F Innatera \u002F Xylo \u002F GrAI; plus foundational works (Mead 1990, Missing Memristor 2008, Tempotron 2006) and frameworks (SPAIC, SNNAX, BrainPy, CARLsim…).\n\n> New paper \u002F model \u002F chip \u002F dataset \u002F tool? [Open a PR](#contributing) and add a line here.\n\n\u003C\u002Fdetails>\n\n---\n\n## Table of Contents\n\n**Part I · Fundamentals**\n- [1 · Foundations & Neural Coding](#1--foundations--neural-coding)\n- [2 · Neuron Models](#2--neuron-models)\n\n**Part II · Learning & Models**\n- [3 · Training Methods](#3--training-methods)\n- [4 · Architectures](#4--architectures)\n- [5 · Spiking Large Models & LLMs](#5--spiking-large-models--llms)\n\n**Part III · Hardware & Systems**\n- [6 · Neuromorphic Hardware](#6--neuromorphic-hardware)\n\n**Part IV · Applications**\n- [7 · Applications](#7--applications)\n\n**Part V · Cross-Cutting Topics**\n- [8 · Energy, Robustness & Security](#8--energy-robustness--security)\n- [9 · Theory & Neuroscience](#9--theory--neuroscience)\n\n**Part VI · Resources & Ecosystem**\n- [10 · Datasets & Benchmarks](#10--datasets--benchmarks)\n- [11 · Software & Frameworks](#11--software--frameworks)\n- [12 · Model Zoo & Community](#12--model-zoo--community)\n- [13 · Research Groups & Labs](#13--research-groups--labs)\n\n**Meta:** [Contributing](#contributing) · [Citation](#citation) · [Star History](#star-history) · [License](#license--acknowledgements)\n\n**Legend:** ★ = seminal \u002F must-read · [paper] = paper · [code] = official code\n\n---\n\n## Part I · Fundamentals\n\n### 1 · Foundations & Neural Coding\n\n> **In one breath:** the roots of the field — from the first threshold neuron and the Nobel-winning Hodgkin–Huxley model to Maass's \"third generation\" framing — plus *neural coding*: **how a real-valued signal becomes spikes** (and back). Rate coding counts spikes over a window (simple, robust, slow); temporal\u002Flatency coding puts information in spike *timing* (fast, efficient, harder to train); rank-order and population codes sit in between. Your choice of code sets the ceiling on both accuracy and latency.\n\n#### Historical Foundations\n\n- A Logical Calculus of the Ideas Immanent in Nervous Activity (**Bull. Math. Biophysics 1943**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1007\u002FBF02478259)\\]\n  > The McCulloch–Pitts threshold neuron — the conceptual seed of all artificial and spiking neurons.\n- A Quantitative Description of Membrane Current and Its Application to Conduction and Excitation in Nerve (**J. Physiology 1952**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1113\u002Fjphysiol.1952.sp004764)\\]\n  > The Hodgkin–Huxley model — the Nobel-winning biophysical basis for all spiking-neuron dynamics.\n- Networks of Spiking Neurons: The Third Generation of Neural Network Models (**Neural Networks 1997**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002FS0893-6080(97)00011-7)\\]\n  > Maass's landmark defining SNNs as the computationally more powerful \"third generation.\"\n- Spiking Neuron Models: Single Neurons, Populations, Plasticity (**Cambridge Univ. Press 2002**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1017\u002FCBO9780511815706)\\]\n  > Gerstner & Kistler's foundational textbook unifying IF, SRM, population, and plasticity theory.\n\n#### Neural Coding & Encoding Schemes\n\n- Spike-Based Strategies for Rapid Processing (**Neural Networks 2001**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002FS0893-6080(01)00083-1)\\]\n  > Thorpe et al.'s rank-order coding — firing *order* alone enables ultra-fast recognition.\n- Rapid Neural Coding in the Retina with Relative Spike Latencies (**Science 2008**). \\[[paper](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.1149639)\\]\n  > Biological evidence that relative first-spike latency carries robust, contrast-invariant information.\n- Neural Coding in Spiking Neural Networks: A Comparative Study for Robust Neuromorphic Systems (**Front. Neurosci. 2021**). \\[[paper](https:\u002F\u002Fwww.frontiersin.org\u002Farticles\u002F10.3389\u002Ffnins.2021.638474\u002Ffull)\\]\n  > Systematically benchmarks rate \u002F temporal \u002F phase \u002F burst codes for accuracy and robustness.\n- Deep Neural Networks with Weighted Spikes (**Neurocomputing 2018**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.neucom.2018.05.087)\\]\n  > Phase\u002Fweighted-spike coding assigns time-dependent weights to spikes, cutting latency and spike count.\n- Conversion of Analog to Spiking Neural Networks Using Sparse Temporal Coding (**IEEE ISCAS 2018**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FISCAS.2018.8351295)\\]\n  > Time-to-first-spike conversion that slashes operations versus rate coding at near-zero accuracy loss.\n- T2FSNN: Deep Spiking Neural Networks with Time-to-First-Spike Coding (**DAC 2020**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2003.11741)\\]\n  > Brings TTFS coding to deep SNNs with kernel thresholds and early-firing for low latency\u002Fenergy.\n- Temporal Coding in Spiking Neural Networks with Alpha Synaptic Function (**ICASSP 2020**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1907.13223)\\]\n  > Enables exact backprop through precise spike times using an alpha-shaped synaptic response.\n- Temporal-Coded Deep Spiking Neural Network with Easy Training and Robust Performance (**AAAI 2021**). \\[[paper](https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F17329)\\]\n  > Argues non-leaky single-spike temporal coding is best for directly-trainable, robust deep SNNs.\n- DIET-SNN: Direct Input Encoding with Leakage and Threshold Optimization (**IEEE TNNLS 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2008.03658)\\]\n  > Feeds analog pixels directly and learns leak\u002Fthreshold end-to-end, popularizing direct input encoding.\n- Supervised Learning Based on Temporal Coding in Spiking Neural Networks (**IEEE TNNLS 2018**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FTNNLS.2017.2726060)\\]\n  > Mostafa's exact gradient descent on the *time* of the first spike — a foundational temporal-coding training method.\n- Optimized Spiking Neurons Can Classify Images with High Accuracy through Temporal Coding with Two Spikes (FS neurons) (**Nature Machine Intelligence 2021**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs42256-021-00311-4)\\]\n  > Few-spike (FS) neurons emulate ANN activations with ~2 spikes, giving high-accuracy, ultra-sparse temporal coding.\n\n---\n\n### 2 · Neuron Models\n\n> **In one breath:** the neuron is the SNN's transistor. **LIF** (leaky integrate-and-fire) is the workhorse — cheap and good enough for deep learning. **Izhikevich** and **AdEx** buy richer spiking dynamics for little cost; **Hodgkin–Huxley** is biophysically exact but expensive. A modern trend is making neuron parameters (e.g., the membrane time constant) **learnable**, letting each neuron tune its own timescale.\n\n- Lapicque's Introduction of the Integrate-and-Fire Model Neuron (1907) (**Brain Res. Bull. 1999**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002FS0361-9230(99)00161-6)\\]\n  > Historical account crediting Lapicque (1907) with the original integrate-and-fire neuron.\n- Simple Model of Spiking Neurons (**IEEE TNN 2003**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FTNN.2003.820440)\\]\n  > Izhikevich's two-variable model reproducing rich cortical firing patterns at integrate-and-fire cost.\n- Which Model to Use for Cortical Spiking Neurons? (**IEEE TNN 2004**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FTNN.2004.832719)\\]\n  > The famous chart trading biological fidelity against compute cost across neuron models — a selection guide.\n- A Framework for Spiking Neuron Models: The Spike Response Model (**Handbook of Biol. Physics 2001**). \\[[paper](https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fpii\u002FS1383812101800154)\\]\n  > Formalizes the SRM, a kernel-based generalization of integrate-and-fire.\n- Adaptive Exponential Integrate-and-Fire Model as an Effective Description of Neuronal Activity (**J. Neurophysiology 2005**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1152\u002Fjn.00686.2005)\\]\n  > The AdEx model — an exponential spike term plus adaptation that accurately fits real neurons.\n- Generalized Leaky Integrate-and-Fire Models Classify Multiple Neuron Types (**Nature Communications 2018**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41467-017-02717-4)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FAllenInstitute\u002FGLIF_Teeter_et_al_2018)\\]\n  > Allen Institute's data-driven GLIF hierarchy fit to 645 real neurons across cell types.\n- Incorporating Learnable Membrane Time Constant to Enhance Learning of SNNs (PLIF) (**ICCV 2021**) ★. \\[[paper](https:\u002F\u002Fopenaccess.thecvf.com\u002Fcontent\u002FICCV2021\u002Fhtml\u002FFang_Incorporating_Learnable_Membrane_Time_Constant_To_Enhance_Learning_of_Spiking_ICCV_2021_paper.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Ffangwei123456\u002FParametric-Leaky-Integrate-and-Fire-Spiking-Neuron)\\]\n  > PLIF makes the membrane time constant learnable, boosting accuracy and easing initialization.\n- GLIF: A Unified Gated Leaky Integrate-and-Fire Neuron for Spiking Neural Networks (**NeurIPS 2022**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=UmFSx2c4ubT)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIkarosy\u002FGated-LIF)\\]\n  > Learnable gates fuse multiple bio-features per neuron, enlarging representational capacity.\n- KLIF: An Optimized Spiking Neuron Unit for Tuning Surrogate Gradient Slope and Membrane Potential (**arXiv 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.09238)\\]\n  > Adds a learnable scaling factor that dynamically shapes the surrogate-gradient curve during training.\n- Parallel Spiking Neurons with High Efficiency and Ability to Learn Long-Term Dependencies (PSN) (**NeurIPS 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.12760)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Ffangwei123456\u002FParallel-Spiking-Neuron)\\]\n  > Removes reset to reformulate neuronal dynamics for parallel (non-serial) simulation and long memory.\n- TC-LIF: A Two-Compartment Spiking Neuron Model for Long-Term Sequential Modelling (**AAAI 2024**). \\[[paper](https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F29625)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FZhangShimin1\u002FTC-LIF)\\]\n  > Soma–dendrite two-compartment neuron designed to propagate gradients over long temporal gaps.\n- CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Networks (**ICML 2024**). \\[[paper](https:\u002F\u002Fproceedings.mlr.press\u002Fv235\u002Fhuang24n.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FHuuYuLong\u002FComplementary-LIF)\\]\n  > A hyperparameter-free neuron that opens extra backprop paths to fight temporal vanishing gradients.\n- Temporal Dendritic Heterogeneity Incorporated with SNNs for Learning Multi-Timescale Dynamics (DH-LIF) (**Nature Communications 2024**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41467-023-44614-z)\\]\n  > Multi-branch dendritic neuron with learnable per-branch time constants for multi-timescale learning.\n- Neural Heterogeneity Promotes Robust Learning (**Nature Communications 2021**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41467-021-26022-3)\\]\n  > Making neuronal time-constants diverse and *learnable* improves accuracy and robustness — a principled case for heterogeneity.\n\n---\n\n\u003Cp align=\"right\">\u003Ca href=\"#table-of-contents\">↑ Back to top\u003C\u002Fa>\u003C\u002Fp>\n\n## Part II · Learning & Models\n\n### 3 · Training Methods\n\n> **In one breath — the field's central problem.** A spike is a non-differentiable step, so plain backprop fails. Three families answer this: **(1) Conversion** trains a normal ANN then maps it to an SNN (high accuracy, high latency); **(2) Surrogate-gradient direct training** pretends the spike has a smooth derivative and runs backprop-through-time (best accuracy-latency trade-off today); **(3) Bio-plausible local rules** like STDP learn from spike timing without global gradients (most brain-like, hardest to scale).\n\n#### 3.1 ANN-to-SNN Conversion\n\n> Train in the easy (ANN) world, deploy in the efficient (SNN) world. The art is matching an SNN's *firing rate* to an ANN's *activation* — via weight\u002Fthreshold normalization — so almost no accuracy is lost, ideally at low latency.\n\n- Spiking Deep Convolutional Neural Networks for Energy-Efficient Object Recognition (**IJCV 2015**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1007\u002Fs11263-014-0788-3)\\]\n  > The seminal work mapping a trained CNN to a spiking IF network via ReLU↔firing-rate correspondence — it launched the conversion paradigm.\n- Fast-Classifying, High-Accuracy Spiking Deep Networks Through Weight and Threshold Balancing (**IJCNN 2015**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FIJCNN.2015.7280696)\\]\n  > Introduced weight normalization \u002F threshold balancing that keeps firing rates in range, making conversion near-lossless.\n- Conversion of Continuous-Valued Deep Networks to Efficient Event-Driven Networks (**Front. Neurosci. 2017**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffnins.2017.00682)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FNeuromorphicProcessorProject\u002Fsnn_toolbox)\\]\n  > Spiking equivalents of BatchNorm\u002Fmax-pool\u002Fsoftmax\u002Fbias enabling accurate VGG\u002FInception conversion — ships the widely used SNN-Toolbox.\n- Going Deeper in Spiking Neural Networks: VGG and Residual Architectures (**Front. Neurosci. 2019**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffnins.2019.00095)\\]\n  > Scaled conversion to deep VGG-16 \u002F ResNet on ImageNet with a Spike-Norm scheme, proving SNNs can go deep.\n- Enabling Deep SNNs with Hybrid Conversion and Spike-Timing-Dependent Backpropagation (**ICLR 2020**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=B1xSperKvH)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fnitin-rathi\u002Fhybrid-snn-conversion)\\]\n  > Initialize from a converted SNN, then fine-tune with spike-based backprop — cutting inference timesteps by an order of magnitude.\n- RMP-SNN: Residual Membrane Potential Neuron for Deeper, High-Accuracy, Low-Latency SNNs (**CVPR 2020**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2003.01811)\\]\n  > \"Soft reset\" (residual membrane potential) removes a key source of conversion error for near-lossless deep SNNs.\n- Optimal Conversion of Conventional ANNs to SNNs (**ICLR 2021**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=FZ1oTwcXchK)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FJackn0\u002Fsnn_optimal_conversion_pipeline)\\]\n  > Decomposes conversion loss layer-wise and uses a rate-norm activation with optimal threshold\u002Fshift to shorten simulation length.\n- A Free Lunch From ANN: Towards Efficient, Accurate SNN Calibration (**ICML 2021**). \\[[paper](https:\u002F\u002Fproceedings.mlr.press\u002Fv139\u002Fli21d.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fyhhhli\u002FSNN_Calibration)\\]\n  > Light-weight layer-by-layer calibration on a handful of samples, scaling to MobileNet\u002FRegNet on ImageNet.\n- Optimal ANN-SNN Conversion for High-Accuracy and Ultra-Low-Latency SNNs (QCFS) (**ICLR 2022**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2303.04347)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fputshua\u002FANN_SNN_QCFS)\\]\n  > Trains the ANN with a quantization-clip-floor-shift activation matching SNN dynamics — high accuracy in as few as 4 timesteps.\n- Optimized Potential Initialization for Low-Latency Spiking Neural Networks (**AAAI 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.01440)\\]\n  > Setting the initial membrane potential to half-threshold minimizes conversion error, enabling accuracy under 32 timesteps.\n- Bridging the Gap Between ANNs and SNNs by Calibrating Offset Spikes (**ICLR 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.10685)\\]\n  > Identifies \"offset spikes\" as the dominant residual error and fixes it by shifting the initial membrane potential.\n- A Unified Optimization Framework of ANN-SNN Conversion (**ICML 2023**). \\[[paper](https:\u002F\u002Fproceedings.mlr.press\u002Fv202\u002Fjiang23a.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FHaiyanJiang\u002FSNN_Conversion_unified)\\]\n  > A SlipReLU unifies performance-loss and conversion-error views — first to reach usable accuracy at a *single* timestep.\n\n#### 3.2 Surrogate Gradient & Direct Training\n\n> Replace the spike's undefined derivative with a smooth \"surrogate\" and backpropagate through time. This is where most SOTA accuracy on hard datasets now comes from.\n\n- Error-Backpropagation in Temporally Encoded Networks of Spiking Neurons (SpikeProp) (**Neurocomputing 2002**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002FS0925-2312(01)00658-0)\\]\n  > The first backprop rule for temporally-coded spiking neurons — ancestor of all gradient-based SNN training.\n- Training Deep Spiking Neural Networks Using Backpropagation (**Front. Neurosci. 2016**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffnins.2016.00508)\\]\n  > Treats the membrane potential as a differentiable signal and spikes as noise, letting standard backprop train deep SNNs.\n- Spatio-Temporal Backpropagation for Training High-Performance SNNs (STBP) (**Front. Neurosci. 2018**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffnins.2018.00331)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fyjwu17\u002FSTBP-for-training-SpikingNN)\\]\n  > Unrolls the SNN in space *and* time (BPTT) with a surrogate derivative — the workhorse recipe for modern direct training.\n- Direct Training for Spiking Neural Networks: Faster, Larger, Better (NeuNorm) (**AAAI 2019**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1809.05793)\\]\n  > Adds neuron normalization (NeuNorm) and improved coding to scale STBP to larger nets and neuromorphic datasets.\n- SLAYER: Spike Layer Error Reassignment in Time (**NeurIPS 2018**) ★. \\[[paper](https:\u002F\u002Fpapers.nips.cc\u002Fpaper\u002F7415-slayer-spike-layer-error-reassignment-in-time)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fbamsumit\u002FslayerPytorch)\\]\n  > Back-propagates error through time with a temporal credit-assignment kernel, jointly learning weights and axonal delays.\n- SuperSpike: Supervised Learning in Multilayer Spiking Neural Networks (**Neural Computation 2018**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1162\u002Fneco_a_01086)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Ffzenke\u002Fpub2018superspike)\\]\n  > An online three-factor surrogate-gradient rule linking deep learning to biological plasticity.\n- The Remarkable Robustness of Surrogate Gradient Learning for Instilling Complex Function in SNNs (**Neural Computation 2021**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1162\u002Fneco_a_01367)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Ffzenke\u002Frandman)\\]\n  > Learning is robust to surrogate *shape* but sensitive to its *scale* — practical design guidance.\n- Temporal Spike Sequence Learning via Backpropagation for Deep SNNs (TSSL-BP) (**NeurIPS 2020**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2002.10085)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fstonezwr\u002FTSSL-BP)\\]\n  > Splits credit assignment into inter- and intra-neuron dependencies for precise temporal learning in very few steps.\n- Differentiable Spike: Rethinking Gradient-Descent for Training SNNs (Dspike) (**NeurIPS 2021**). \\[[paper](https:\u002F\u002Fproceedings.neurips.cc\u002Fpaper\u002F2021\u002Fhash\u002Fc4ca4238a0b923820dcc509a6f75849b-Abstract.html)\\]\n  > An adaptively-tunable family of differentiable surrogates that minimizes gradient mismatch.\n- Training Feedback SNNs by Implicit Differentiation on the Equilibrium State (IDE) (**NeurIPS 2021**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2109.14247)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fpkuxmq\u002FIDE-FSNN)\\]\n  > Trains feedback SNNs via implicit differentiation of their equilibrium state — memory cost independent of timesteps.\n- Sparse Spiking Gradient Descent (**NeurIPS 2021**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2105.08810)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fnpvoid\u002FSparseSpikingBackprop)\\]\n  > Exploits spatiotemporal sparsity in the backward pass for up to 150× faster, 85% lower-memory training.\n- Temporal Efficient Training of SNNs via Gradient Re-weighting (TET) (**ICLR 2022**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2202.11946)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FGus-Lab\u002Ftemporal_efficient_training)\\]\n  > A per-timestep loss compensating surrogate-gradient momentum loss — flatter minima, better generalization; now a default trick.\n- Training High-Performance Low-Latency SNNs by Differentiation on Spike Representation (DSR) (**CVPR 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.00459)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fqymeng94\u002FDSR)\\]\n  > Treats the firing-rate representation as a sub-differentiable map and trains through it, sidestepping non-differentiability.\n- Online Training Through Time for Spiking Neural Networks (OTTT) (**NeurIPS 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.04195)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fpkuxmq\u002FOTTT-SNN)\\]\n  > A constant-memory online alternative to BPTT with a three-factor Hebbian form suited to on-chip learning.\n- RecDis-SNN: Rectifying Membrane Potential Distribution for Directly Training SNNs (**CVPR 2022**). \\[[paper](https:\u002F\u002Fopenaccess.thecvf.com\u002Fcontent\u002FCVPR2022\u002Fhtml\u002FGuo_RecDis-SNN_Rectifying_Membrane_Potential_Distribution_for_Directly_Training_Spiking_Neural_CVPR_2022_paper.html)\\]\n  > A membrane-potential distribution loss that mitigates degeneration, saturation, and gradient mismatch.\n- IM-Loss: Information Maximization Loss for Spiking Neural Networks (**NeurIPS 2022**). \\[[paper](https:\u002F\u002Fproceedings.neurips.cc\u002Fpaper_files\u002Fpaper\u002F2022\u002Fhash\u002F010c5ba0cafc743fece8be02e7adb8dd-Abstract-Conference.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fyfguo91\u002FIM-Loss-Information-Maximization-Loss-for-Spiking-Neural-Networks)\\]\n  > Maximizes activation information entropy to counter the loss caused by 0\u002F1 spike quantization.\n- RMP-Loss: Regularizing Membrane Potential Distribution for Spiking Neural Networks (**ICCV 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.06787)\\]\n  > A loss pulling membrane potentials toward spike values, directly shrinking quantization error.\n- Real Spike: Learning Real-valued Spikes for Spiking Neural Networks (**ECCV 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2210.06686)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fyfguo91\u002FReal-Spike)\\]\n  > Learns real-valued spikes during training, re-parameterized back to binary at inference — capacity for free.\n- Surrogate Module Learning: Reduce Gradient Error Accumulation in Training SNNs (**ICML 2023**). \\[[paper](https:\u002F\u002Fproceedings.mlr.press\u002Fv202\u002Fdeng23d.html)\\]\n  > Surrogate modules create a shortcut path for more accurate gradients, curbing layer-wise gradient-error buildup.\n- Towards Memory- and Time-Efficient Backpropagation for Training SNNs (SLTT) (**ICCV 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.14311)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fqymeng94\u002FSLTT)\\]\n  > Shows temporal backprop contributes little; dropping those routes cuts memory >70% and training time >50%.\n- A Tandem Learning Rule for Effective Training and Rapid Inference of Deep SNNs (**IEEE TNNLS 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1907.01167)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fdeepspike\u002Ftandem_learning)\\]\n  > Couples an ANN and SNN via weight sharing — the ANN carries gradients while the SNN counts spikes.\n- Advancing Spatiotemporal Representations in SNNs via Parametric Invertible Transformation (PIT) (**ICLR 2026**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=3JwNXQzxll)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FYinsongYan\u002FICLR26)\\]\n  > Conjugates an invertible transform with neuron dynamics and corrects surrogate-gradient mismatch, expanding the usable binary-spike representation space.\n- Accurate and Efficient Time-Domain Classification with Adaptive Spiking Recurrent Neural Networks (**Nature Machine Intelligence 2021**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs42256-021-00397-w)\\]\n  > Adaptive spiking neurons + surrogate-gradient RSNNs reach RNN-level accuracy on speech\u002Fgesture at far less compute.\n\n#### 3.3 Biologically-Plausible \u002F Local Learning\n\n> **STDP** and friends: synapses strengthen or weaken based purely on the *relative timing* of pre- and post-synaptic spikes — local, unsupervised, and hardware-friendly, but historically hard to push to ImageNet scale.\n\n- Synaptic Modifications in Cultured Hippocampal Neurons: Dependence on Spike Timing... (**J. Neuroscience 1998**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1523\u002FJNEUROSCI.18-24-10464.1998)\\]\n  > Bi & Poo's landmark experiment quantifying STDP — the biological basis of local SNN learning rules.\n- Unsupervised Learning of Digit Recognition Using STDP (**Front. Comput. Neurosci. 2015**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffncom.2015.00099)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fpeter-u-diehl\u002Fstdp-mnist)\\]\n  > Diehl & Cook's STDP + lateral-inhibition network learns MNIST unsupervised (~95%) — the canonical bio-plausible baseline.\n- STDP-Based Spiking Deep Convolutional Neural Networks for Object Recognition (**Neural Networks 2018**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1016\u002Fj.neunet.2017.12.005)\\]\n  > Stacks STDP-trained conv layers with latency coding, showing local plasticity can learn deep hierarchical features.\n- Bio-Inspired Digit Recognition Using Reward-Modulated STDP in Deep Conv Networks (**Pattern Recognition 2019**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1804.00227)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fmiladmozafari\u002FSpykeTorch)\\]\n  > Combines unsupervised STDP with reward-modulated STDP (a three-factor rule) for reinforcement-driven feature learning.\n- A Solution to the Learning Dilemma for Recurrent Networks of Spiking Neurons (e-prop) (**Nature Communications 2020**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fs41467-020-17236-y)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIGITUGraz\u002Feligibility_propagation)\\]\n  > Local eligibility traces + top-down learning signals approximate BPTT without backward-in-time — enabling on-chip learning.\n- Equilibrium Propagation: Bridging Energy-Based Models and Backpropagation (**Front. Comput. Neurosci. 2017**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffncom.2017.00024)\\]\n  > Computes exact gradients using only local, same-type computation across two phases — a biologically plausible backprop alternative.\n- Training Spiking Neural Networks via Augmented Direct Feedback Alignment (**NeurIPS 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.07776)\\]\n  > A gradient-free random-projection feedback-alignment rule avoiding weight transport, improving biological\u002Fhardware fit.\n- Backpropagation-Free Spiking Neural Networks with the Forward-Forward Algorithm (**arXiv 2025**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2502.20411)\\]\n  > Adapts Hinton's Forward-Forward (two contrastive forward passes, layer-local goodness) to spiking neurons.\n- The Tempotron: A Neuron That Learns Spike Timing-Based Decisions (**Nature Neuroscience 2006**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1038\u002Fnn1643)\\]\n  > Gütig & Sompolinsky's tempotron — a single neuron that learns to classify by the *timing* of its input spikes.\n- Unsupervised Learning of Visual Features through Spike-Timing-Dependent Plasticity (**PLoS Comput. Biol. 2007**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1371\u002Fjournal.pcbi.0030031)\\]\n  > Masquelier & Thorpe show STDP + latency coding self-organizes selective visual features — a landmark unsupervised result.\n\n#### 3.4 Efficiency: Pruning, Quantization, Distillation\n\n> Making an already-efficient model *more* efficient: fewer timesteps, fewer weights, lower precision, and distilling knowledge from ANN teachers.\n\n- Towards Ultra-Low-Latency SNNs for Vision and Sequential Tasks Using Temporal Pruning (**ECCV 2022**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1007\u002F978-3-031-20083-0_42)\\]\n  > Iteratively prunes timesteps during training, driving SNNs toward single-timestep inference.\n- Constructing Deep SNNs from ANNs with Knowledge Distillation (**CVPR 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2304.05627)\\]\n  > Uses an ANN teacher to distill feature\u002Fresponse knowledge into an SNN student, avoiding costly from-scratch training.\n- TP-Spikformer: Token Pruned Spiking Transformer (**ICLR 2026**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=L5llQD0nMf)\\]\n  > Prunes low-information tokens without retraining across several spiking-Transformer families, reducing storage and compute while retaining competitive accuracy.\n- Towards Lossless Memory-efficient Training of SNNs via Gradient Checkpointing and Spike Compression (**ICLR 2026**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=nrBJ0Uvj7c)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FAllenYolk\u002Fsnn-gradient-checkpointing)\\]\n  > Combines adaptive spatiotemporal checkpointing with lossless binary-spike compression for up to 8× lower training memory without accuracy loss.\n\n> Also relevant: several low-timestep \u002F low-memory training methods double as efficiency techniques — see **TET**, **SLTT**, **DSR** in [3.2](#32-surrogate-gradient--direct-training) and **DIET-SNN** in [Neural Coding](#1--foundations--neural-coding).\n\n---\n\n### 4 · Architectures\n\n> **In one breath:** the network *shape*. The 2019–2021 breakthrough was getting **residual\u002FBN tricks** to work in the spiking domain so SNNs could go deep; the 2023+ wave brought **spiking Transformers**, redesigning self-attention to run on spikes.\n\n#### 4.1 Deep Spiking CNNs & ResNets\n\n> Residual connections and spike-aware normalization (tdBN, BNTT, TEBN) are what let SNNs scale past a few layers without spikes vanishing or exploding.\n\n- Spiking Deep Residual Networks (**IEEE TNNLS 2021**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1805.01352)\\]\n  > The original \"Spiking ResNet\" — scaled shortcuts + error compensation build the first >40-layer SNN matching ANN accuracy.\n- Deep Residual Learning in Spiking Neural Networks (SEW-ResNet) (**NeurIPS 2021**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2102.04159)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Ffangwei123456\u002FSpike-Element-Wise-ResNet)\\]\n  > The spike-element-wise (SEW) block enables identity mapping and solves vanishing\u002Fexploding gradients — first directly-trained 100+ layer SNNs.\n- Advancing Spiking Neural Networks Toward Deep Residual Learning (MS-ResNet) (**IEEE TNNLS 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2112.08954)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FAriande1\u002FMS-ResNet)\\]\n  > Membrane-potential (pre-activation) shortcuts preserve spike-driven computation and gradient-norm equality, scaling to 482 layers.\n- Going Deeper with Directly-Trained Larger SNNs (STBP-tdBN) (**AAAI 2021**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2011.05280)\\]\n  > Threshold-dependent BatchNorm (tdBN) balances firing rates across time, extending directly-trained SNNs from \u003C10 to 50 layers.\n- Revisiting Batch Normalization for Training Low-Latency Deep SNNs from Scratch (BNTT) (**Front. Neurosci. 2021**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2010.01729)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIntelligent-Computing-Lab-Panda\u002FBNTT-Batch-Normalization-Through-Time)\\]\n  > Decouples BN parameters along the time axis to capture spike temporal dynamics and enable low-latency training from scratch.\n- Temporal Effective Batch Normalization in Spiking Neural Networks (TEBN) (**NeurIPS 2022**). \\[[paper](https:\u002F\u002Fproceedings.neurips.cc\u002Fpaper_files\u002Fpaper\u002F2022\u002Fhash\u002Fde2ad3ed44ee4e675b3be42aa0b615d0-Abstract-Conference.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FChaotengDuan\u002FTEBN)\\]\n  > Rescales inputs with distinct learnable weights per time-step, smoothing temporal distributions and the optimization landscape.\n- Membrane Potential Batch Normalization for Spiking Neural Networks (MPBN) (**ICCV 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.08359)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fyfguo91\u002FMPBN)\\]\n  > A second BN on the membrane potential before firing, folded into the threshold by re-parameterization — zero inference cost.\n\n#### 4.2 Spiking Transformers & Attention\n\n> Self-attention, re-derived so that queries\u002Fkeys\u002Fvalues are spikes and the expensive softmax is replaced by spike-friendly operations — bringing Transformer-level accuracy to the spiking world.\n\n- Spikformer: When Spiking Neural Network Meets Transformer (**ICLR 2023**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.15425)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FZK-Zhou\u002Fspikformer)\\]\n  > Introduces softmax-free Spiking Self-Attention (SSA) with spike-form Q\u002FK\u002FV — the first vision Transformer built directly in the spiking domain.\n- Spike-driven Transformer (**NeurIPS 2023**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2307.01694)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FBICLab\u002FSpike-Driven-Transformer)\\]\n  > A purely spike-driven Transformer whose attention uses only mask + sparse addition (linear complexity), cutting attention energy up to 87×.\n- Spike-driven Transformer V2 (Meta-SpikeFormer) (**ICLR 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2404.03663)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FBICLab\u002FSpike-Driven-Transformer-V2)\\]\n  > A meta spiking backbone unifying classification, detection, and segmentation, guiding next-gen neuromorphic chip design (80% ImageNet).\n- Scaling Spike-driven Transformer with Efficient Spike Firing Approximation (V3) (**IEEE TPAMI 2025**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2411.16061)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FBICLab\u002FSpike-Driven-Transformer-V3)\\]\n  > Integer training + spike-driven inference plus a spike masked autoencoder scale SNNs to 86.2% on ImageNet.\n- Spikformer V2: Join the High-Accuracy Club on ImageNet with an SNN Ticket (**arXiv 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.02020)\\]\n  > A Spiking Convolutional Stem + self-supervised pretraining — among the first SNNs past 80% top-1 on ImageNet.\n- QKFormer: Hierarchical Spiking Transformer using Q-K Attention (**NeurIPS 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.16552)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fzhouchenlin2096\u002FQKFormer)\\]\n  > Linear-complexity binary Q-K attention + hierarchical pyramid — first directly-trained SNN past 85% top-1 on ImageNet.\n- Spikingformer: A Key Foundation Model for Spiking Neural Networks (**AAAI 2026**). \\[[paper](https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F37207)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FTheBrainLab\u002FSpikingformer)\\]\n  > Replaces Spikformer's non-spike residuals with a fully spike-driven design, removing integer-float multiplications for hardware friendliness.\n- SpikingResformer: Bridging ResNet and Vision Transformer in SNNs (**CVPR 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2403.14302)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fxyshi2000\u002FSpikingResformer)\\]\n  > Combines a ResNet-style multi-stage backbone with Dual Spike Self-Attention (DSSA) for high accuracy at fewer params\u002Fenergy.\n- Masked Spiking Transformer (**ICCV 2023**). \\[[paper](https:\u002F\u002Fopenaccess.thecvf.com\u002Fcontent\u002FICCV2023\u002Fhtml\u002FWang_Masked_Spiking_Transformer_ICCV_2023_paper.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fbic-L\u002FMasked-Spiking-Transformer)\\]\n  > An ANN-to-SNN converted Transformer with Random Spike Masking that prunes redundant spikes to cut energy without accuracy loss.\n- Spiking Transformer with Spatial-Temporal Attention (STAtten) (**CVPR 2025**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.19764)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIntelligent-Computing-Lab-Panda\u002FSTAtten)\\]\n  > Block-wise attention jointly integrating spatial and temporal information at the cost of spatial-only spiking attention.\n- Neural Dynamics Self-Attention for Spiking Transformers (**ICLR 2026**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=jJedqisfOt)\\]\n  > Adds a local receptive-field bias and realizes attention through charge–fire–reset dynamics, avoiding explicit attention-matrix storage at inference.\n- TIM: An Efficient Temporal Interaction Module for Spiking Transformer (**IJCAI 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.11687)\\]\n  > A lightweight convolutional module injecting previous-timestep information into the attention matrix, strengthening temporal modeling.\n- Temporal-wise Attention Spiking Neural Networks for Event Streams Classification (TA-SNN) (**ICCV 2021**). \\[[paper](https:\u002F\u002Fopenaccess.thecvf.com\u002Fcontent\u002FICCV2021\u002Fhtml\u002FYao_Temporal-Wise_Attention_Spiking_Neural_Networks_for_Event_Streams_Classification_ICCV_2021_paper.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FBICLab\u002FTA-SNN)\\]\n  > Temporal-wise attention weights event frames and discards noisy ones — a landmark attention-SNN for event data.\n- Attention Spiking Neural Networks (MA-SNN) (**IEEE TPAMI 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2209.13929)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FBICLab\u002FAttention-SNN)\\]\n  > Multi-dimensional (temporal\u002Fchannel\u002Fspatial) attention modulates membrane potentials, yielding sparser firing and higher accuracy.\n\n#### 4.3 Recurrent, Reservoir & Other\n\n> Recurrent SNNs (LSNN), liquid state machines, spiking GNNs\u002Fautoencoders\u002FGANs, and neural-architecture-searched SNNs.\n\n**Recurrent & Reservoir**\n- Long Short-Term Memory and Learning-to-Learn in Networks of Spiking Neurons (LSNN) (**NeurIPS 2018**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1803.09574)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIGITUGraz\u002FLSNN-official)\\]\n  > Adaptive-threshold (ALIF) neurons + BPTT learning-to-learn first matched LSTM-level temporal computing power.\n- Real-Time Computing Without Stable States (Liquid State Machine) (**Neural Computation 2002**) ★. \\[[paper](https:\u002F\u002Fdirect.mit.edu\u002Fneco\u002Farticle\u002F14\u002F11\u002F2531\u002F6650)\\]\n  > The foundational Liquid State Machine \u002F reservoir-computing model — a recurrent spiking circuit projects inputs into a high-dimensional readable state.\n\n**Spiking Graph Neural Networks**\n- Spiking Graph Convolutional Networks (SpikingGCN) (**IJCAI 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2205.02767)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FZulunZhu\u002FSpikingGCN)\\]\n  > Encodes graph convolution into spike trains end-to-end, bringing energy-efficient SNN inference to node\u002Fgraph tasks.\n- Scaling Up Dynamic Graph Representation Learning via Spiking Neural Networks (SpikeNet) (**AAAI 2023**). \\[[paper](https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F26034)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FEdisonLeeeee\u002FSpikeNet)\\]\n  > Replaces RNNs with spiking neurons to scale temporal graph learning to millions of nodes at low compute.\n- A Graph is Worth 1-bit Spikes: Graph Contrastive Learning Meets SNNs (SpikeGCL) (**ICLR 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.19306)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FEdisonLeeeee\u002FSpikeGCL)\\]\n  > Learns binarized 1-bit graph representations via spiking contrastive learning — ~32× storage compression at comparable accuracy.\n- Dynamic Spiking Graph Neural Networks (Dy-SIGN) (**AAAI 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2401.05373)\\]\n  > Tackles information loss in dynamic spiking GNNs with implicit differentiation and information compensation.\n\n**Spiking Generative Models**\n- Spiking-GAN: A Spiking Generative Adversarial Network Using Time-To-First-Spike Coding (**IJCNN 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2106.15420)\\]\n  > The first spike-based GAN, trained with temporal coding and BPTT for ultra-low-energy generation.\n- Spiking Denoising Diffusion Probabilistic Models (SDDPM) (**WACV 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2306.17046)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FSageCao1125\u002FSDDPM)\\]\n  > Brings diffusion models into SNNs with a Spiking U-Net backbone, matching\u002Fbeating ANN DDPM on FID.\n- SDiT: Spiking Diffusion Model with Transformer (**arXiv 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.11588)\\]\n  > Replaces the diffusion U-Net with a spiking Transformer backbone for higher-quality, lower-cost SNN image generation.\n- Spiking Generative Adversarial Network with Attention Scoring Decoding (**Neural Networks 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2305.10246)\\]\n  > An attention-scoring decoder generates higher-fidelity images from spike features.\n- Fully Spiking Variational Autoencoder (FSVAE) (**AAAI 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2110.00375)\\]\n  > The first fully-spiking VAE — samples latent variables as autoregressive Bernoulli spike trains, generating images end-to-end in spikes.\n\n**Neural Architecture Search for SNNs**\n- Neural Architecture Search for Spiking Neural Networks (SNASNet) (**ECCV 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.10355)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIntelligent-Computing-Lab-Panda\u002FNeural-Architecture-Search-for-Spiking-Neural-Networks)\\]\n  > Training-free NAS selecting architectures by spike-activation diversity at initialization, plus temporal backward connections.\n- AutoSNN: Towards Energy-Efficient Spiking Neural Networks (**ICML 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2201.12738)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fnabk89\u002FAutoSNN)\\]\n  > Spike-aware architecture search whose fitness jointly optimizes accuracy and spike count.\n- Differentiable Hierarchical and Surrogate Gradient Search for SNNs (SpikeDHS) (**NeurIPS 2022**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=Lr2Z85cdvB)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FHuawei-BIC\u002FSpikeDHS)\\]\n  > Jointly and differentiably searches cell\u002Flayer architecture and the surrogate-gradient function.\n\n---\n\n### 5 · Spiking Large Models & LLMs\n\n> **In one breath:** the fastest-moving frontier — spiking **language models** and **multimodal** models that bring Transformer\u002FLLM-scale capability into the event-driven, energy-efficient spiking world. (Vision spiking Transformers live in [§4.2](#42-spiking-transformers--attention).)\n\n- SpikeGPT: Generative Pre-trained Language Model with Spiking Neural Networks (**TMLR 2024**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2302.13939)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fridgerchu\u002FSpikeGPT)\\]\n  > The first large generative spiking language model (up to 260M params), linearizing attention for ~20× fewer operations.\n- Spiking Convolutional Neural Networks for Text Classification (**ICLR 2023**). \\[[paper](https:\u002F\u002Fopenreview.net\u002Fforum?id=pgU3k7QXuz0)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FLvchangze\u002Fsnn)\\]\n  > A conversion-plus-fine-tuning recipe encoding word embeddings as spikes, matching ANN text classifiers with more robustness.\n- SpikingBERT: Distilling BERT to Train Spiking Language Models Using Implicit Differentiation (**AAAI 2024**). \\[[paper](https:\u002F\u002Fojs.aaai.org\u002Findex.php\u002FAAAI\u002Farticle\u002Fview\u002F28975)\\]\n  > A spiking BERT trained via implicit-differentiation equilibrium + ANN-to-SNN distillation, bringing SNNs to NLU tasks.\n- SpikeBERT: A Language Spikformer Learned from BERT with Knowledge Distillation (**arXiv 2023**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2308.15122)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FLvchangze\u002FSpikeBERT)\\]\n  > Two-stage distillation from BERT into a Spikformer, matching BERT on English\u002FChinese text classification at far lower energy.\n- SpikeLM: Towards General Spike-Driven Language Modeling via Elastic Bi-Spiking Mechanisms (**ICML 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.03287)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FXingrun-Xing\u002FSpikeLM)\\]\n  > A fully spike-driven mechanism handling both discriminative and generative language tasks with elastic bidirectional spikes.\n- SpikeLLM: Scaling up Spiking Neural Networks to Large Language Models via Saliency-Based Spiking (**arXiv 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2407.04752)\\]\n  > The first spiking LLM scaled to 7–70B params, using generalized IF neurons and saliency-based spiking to beat quantization baselines.\n- SpikeZIP-TF: Conversion is All You Need for Transformer-based SNN (**ICML 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2406.03470)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FIntelligent-Computing-Research-Group\u002FSpikeZIP-TF)\\]\n  > Losslessly converts quantized Transformers into SNNs, closing the accuracy gap with ANN Transformers on vision and language.\n- SpikingBrain: Spiking Brain-inspired Large Models (**arXiv 2025**) ★. \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2509.05276)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FBICLab\u002FSpikingBrain-7B)\\]\n  > CAS (Li Guoqi & Bo Xu) 7B-linear \u002F 76B-MoE spiking LLMs with adaptive spike coding — >100× TTFT speedup on 4M-token contexts, trained on domestic (MetaX) GPUs.\n- SpikingBrain2.0: Brain-Inspired Foundation Models for Efficient Long-Context and Cross-Platform Inference (**arXiv 2026**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.22575)\\]\n  > Extends the family to 5B language and vision-language models with dual sparse attention and INT8-spiking \u002F FP8 paths, supporting 10M+ tokens and GPU or neuromorphic inference.\n- Sorbet: A Neuromorphic Hardware-Compatible Transformer-Based Spiking Language Model (**ICML 2025**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2409.15298)\\]\n  > Replaces softmax and LayerNorm with shift-based PTsoftmax\u002FBSPN for a hardware-friendly spiking LM — ~27× energy savings vs BERT on GLUE.\n- SpikeCLIP: A Contrastive Language-Image Pretrained Spiking Neural Network (**Neural Networks 2025**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2310.06488)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FLvchangze\u002FSpikeCLIP)\\]\n  > A CLIP-style *multimodal* spiking model aligning image and text in spike space — extends SNNs to vision-language.\n\n---\n\n\u003Cp align=\"right\">\u003Ca href=\"#table-of-contents\">↑ Back to top\u003C\u002Fa>\u003C\u002Fp>\n\n## Part III · Hardware & Systems\n\n### 6 · Neuromorphic Hardware\n\n> **In one breath:** SNNs only pay off when the *chip* is event-driven too. **Digital** platforms (TrueNorth, Loihi, SpiNNaker, Tianjic) route spikes as packets and idle between them; **analog\u002Fin-memory** designs (BrainScaleS, memristor\u002FRRAM crossbars) compute inside the memory to kill the von-Neumann data-movement cost. **Event cameras** (DVS) are the matching sensor.\n\n#### Foundations of Neuromorphic Engineering\n\n- Neuromorphic Electronic Systems (**Proceedings of the IEEE 1990**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002F5.58356)\\]\n  > Carver Mead's founding manifesto — analog VLSI that mimics neural computation. The paper that named the field.\n- A Silicon Neuron (**Nature 1991**) ★. \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002F354515a0)\\]\n  > Mahowald & Douglas's analog VLSI neuron reproducing real spiking dynamics — the first silicon neuron.\n- Point-to-Point Connectivity Between Neuromorphic Chips Using Address Events (**IEEE TCAS-II 2000**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002F82.842110)\\]\n  > Boahen formalizes Address-Event Representation (AER) — the spike-as-packet protocol every neuromorphic chip now uses.\n\n#### Digital Neuromorphic Chips\n\n- A Million Spiking-Neuron Integrated Circuit with a Scalable Communication Network (TrueNorth) (**Science 2014**) ★. \\[[paper](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.1254642)\\]\n  > IBM's TrueNorth packs 1M neurons + 256M synapses into a 65 mW event-driven non-von-Neumann chip — the landmark large-scale digital neuromorphic silicon.\n- Convolutional Networks for Fast, Energy-Efficient Neuromorphic Computing (TrueNorth) (**PNAS 2016**). \\[[paper](https:\u002F\u002Fwww.pnas.org\u002Fdoi\u002F10.1073\u002Fpnas.1604850113)\\]\n  > Maps deep convnets onto TrueNorth at near-SOTA accuracy, 1,200–2,600 fps and tens of mW.\n- Loihi: A Neuromorphic Manycore Processor with On-Chip Learning (**IEEE Micro 2018**) ★. \\[[paper](https:\u002F\u002Fieeexplore.ieee.org\u002Fdocument\u002F8259423)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Flava-nc\u002Flava)\\]\n  > Intel's 14 nm 128-core chip with programmable synaptic learning, dendritic compartments and delays — the leading on-chip-learning research platform.\n- Taking Neuromorphic Computing to the Next Level with Loihi 2 (**Intel Tech Brief 2021**). \\[[paper](https:\u002F\u002Fwww.intel.com\u002Fcontent\u002Fwww\u002Fus\u002Fen\u002Fresearch\u002Fneuromorphic-computing-loihi-2-technology-brief.html)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Flava-nc\u002Flava)\\]\n  > Loihi 2 adds graded spikes, programmable neuron microcode and up to 1M neurons in 7 nm, paired with the open-source Lava framework.\n- The SpiNNaker Project (**Proc. IEEE 2014**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FJPROC.2014.2304638)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FSpiNNakerManchester)\\]\n  > Manchester's massively-parallel ARM machine models spiking networks in biological real time via brain-inspired packet routing.\n- SpiNNaker: A 1-W 18-Core System-on-Chip for Massively-Parallel Neural Simulation (**IEEE JSSC 2013**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FJSSC.2013.2259038)\\]\n  > The 18-ARM-core GALS chip (100M transistors, 1 W) — physical building block of the million-core SpiNNaker machine.\n- SpiNNaker 2: A 10-Million-Core Processor System for Brain Simulation and Machine Learning (**arXiv 2019**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1911.02385)\\]\n  > 22 nm FDSOI successor adding numerical accelerators and adaptive power management for both brain simulation and ML.\n- Towards Artificial General Intelligence with Hybrid Tianjic Chip Architecture (**Nature 2019**) ★. \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-019-1424-8)\\]\n  > Tsinghua's Tianjic unifies ANN and SNN paradigms on one reconfigurable many-core chip — famously demoed driving an autonomous bicycle.\n- Darwin: A Neuromorphic Hardware Co-Processor Based on Spiking Neural Networks (**J. Systems Architecture 2017**). \\[[paper](https:\u002F\u002Fwww.sciencedirect.com\u002Fscience\u002Farticle\u002Fabs\u002Fpii\u002FS1383762117300231)\\]\n  > China's Darwin NPU (180 nm, 2,048 neurons, configurable delays) — an early low-power embedded SNN co-processor.\n- Darwin3: A Large-Scale Neuromorphic Chip with a Novel ISA and On-Chip Learning (**National Science Review 2024**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2312.17582)\\]\n  > Zhejiang Univ. (Gang Pan) digital chip with a dedicated 10-instruction ISA and on-chip learning — up to ~2.35M neurons per chip.\n- ODIN: A 0.086 mm² 12.7 pJ\u002FSOP 64k-Synapse 256-Neuron Online-Learning Digital SNN Processor (**IEEE TBCAS 2019**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1804.07858)\\]\\[[code](https:\u002F\u002Fgithub.com\u002FChFrenkel\u002FODIN)\\]\n  > A tiny open-source 28 nm chip with SDSP on-chip learning and Izhikevich-capable neurons, setting synaptic-density\u002Fenergy records.\n- MorphIC: A 65-nm Quad-Core Binary-Weight Digital Neuromorphic Processor with Stochastic Online Learning (**IEEE TBCAS 2019**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F1904.08513)\\]\n  > Frenkel's four-core follow-up to ODIN, maximizing synaptic density with binary weights and stochastic plasticity for edge learning.\n- μBrain: An Event-Driven and Fully Synthesizable Architecture for Spiking Neural Networks (**Front. Neurosci. 2021**). \\[[paper](https:\u002F\u002Fwww.frontiersin.org\u002Farticles\u002F10.3389\u002Ffnins.2021.664208\u002Ffull)\\]\n  > The first clockless, fully-synthesizable digital SNN chip (40 nm, sub-100 μW) for always-on near-sensor edge AI.\n- SENeCA: Building a Fully Digital Neuromorphic Processor (**Front. Neurosci. 2023**). \\[[paper](https:\u002F\u002Fwww.frontiersin.org\u002Farticles\u002F10.3389\u002Ffnins.2023.1187252\u002Ffull)\\]\n  > imec's flexible RISC-V + loop-buffer architecture balancing programmability and efficiency for diverse SNN + on-device learning.\n- A Digital Neurosynaptic Core Using Embedded Crossbar Memory with 45 pJ per Spike in 45 nm (**IEEE CICC 2011**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FCICC.2011.6055294)\\]\n  > Merolla et al.'s neurosynaptic core — the direct architectural precursor to IBM TrueNorth.\n- ReckOn: A 28 nm Sub-mm² Task-Agnostic Spiking Recurrent Neural Network Processor Enabling On-Chip Learning over Second-Long Timescales (**IEEE ISSCC 2022**). \\[[paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2208.09759)\\]\n  > Frenkel & Indiveri's chip does e-prop-style on-chip learning of temporal tasks over second-long horizons at sub-mW power.\n\n#### Analog & Mixed-Signal \u002F Sub-threshold\n\n- Neurogrid: A Mixed-Analog-Digital Multichip System for Large-Scale Neural Simulations (**Proc. IEEE 2014**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FJPROC.2014.2313565)\\]\n  > Boahen's 16-Neurocore board simulates a million subthreshold-analog neurons with billions of synapses in real time at just 3 W.\n- A Wafer-Scale Neuromorphic Hardware System for Large-Scale Neural Modeling (BrainScaleS-1) (**IEEE ISCAS 2010**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FISCAS.2010.5536970)\\]\n  > Heidelberg's HICANN wafer-scale system emulates ~200k neurons per wafer at 10,000× biological speed via accelerated analog dynamics.\n- The BrainScaleS-2 Accelerated Neuromorphic System with Hybrid Plasticity (**Front. Neurosci. 2022**). \\[[paper](https:\u002F\u002Fwww.frontiersin.org\u002Farticles\u002F10.3389\u002Ffnins.2022.795876\u002Ffull)\\]\n  > Couples continuous-time analog neuron\u002Fsynapse circuits with embedded SIMD processors for flexible on-chip hybrid plasticity.\n- DYNAP-SE: A Scalable Multicore Architecture with Heterogeneous Memory for Dynamic Neuromorphic Async Processors (**IEEE TBCAS 2018**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FTBCAS.2017.2759700)\\]\n  > Indiveri's subthreshold-analog chip introduces heterogeneous routing that solves the connectivity-scaling problem.\n- ROLLS: A Reconfigurable On-Line Learning Spiking Neuromorphic Processor (256 Neurons, 128k Synapses) (**Front. Neurosci. 2015**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.3389\u002Ffnins.2015.00141)\\]\n  > A mixed-signal chip emulating real neuron\u002Fsynapse physics with spike-based plasticity for fully on-chip online learning.\n- DYNAP-CNN & Speck: Event-Driven Convolutional Neuromorphic Vision Processors (SynSense) (**2020**). \\[[paper](https:\u002F\u002Fwww.synsense.ai\u002Fproducts\u002Fspeck-2\u002F)\\]\\[[code](https:\u002F\u002Fgithub.com\u002Fsynsense\u002Fsinabs)\\]\n  > Commercial sub-mW spiking-CNN processors (Speck integrates a DVS + DynapCNN on one SoC) for always-on event vision at μs latency.\n\n#### In-Memory \u002F Memristive \u002F RRAM & PCM Computing\n\n- Nanoscale Memristor Device as Synapse in Neuromorphic Systems (**Nano Letters 2010**) ★. \\[[paper](https:\u002F\u002Fpubs.acs.org\u002Fdoi\u002F10.1021\u002Fnl904092h)\\]\n  > Jo & Lu experimentally demonstrated STDP in a single nanoscale memristor — launching memristive-synapse computing.\n- Nanoelectronic Programmable Synapses Based on Phase-Change Materials (**Nano Letters 2011**). \\[[paper](https:\u002F\u002Fpubs.acs.org\u002Fdoi\u002F10.1021\u002Fnl201040y)\\]\n  > Continuous PCM resistance transitions emulate analog synaptic plasticity (STDP) at picojoule energy.\n- Training and Operation of an Integrated Neuromorphic Network Based on Metal-Oxide Memristors (**Nature 2015**) ★. \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnature14441)\\]\n  > The first transistor-free memristor crossbar perceptron trained in situ — proof of integrated memristive neural networks.\n- Stochastic Phase-Change Neurons (**Nature Nanotechnology 2016**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnnano.2016.70)\\]\n  > IBM realizes integrate-and-fire neurons in PCM devices whose intrinsic stochasticity enables population coding.\n- Memristors with Diffusive Dynamics as Synaptic Emulators (**Nature Materials 2017**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnmat4756)\\]\n  > Ag-nanoparticle diffusive memristors reproduce Ca²⁺-like short-term synaptic dynamics — a biologically faithful analog synapse.\n- Fully Memristive Neural Networks for Pattern Classification with Unsupervised Learning (**Nature Electronics 2018**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41928-018-0023-2)\\]\n  > Integrates diffusive-memristor LIF neurons with nonvolatile memristor synapses into an all-memristive unsupervised network.\n- Equivalent-Accuracy Accelerated Neural-Network Training Using Analogue Memory (**Nature 2018**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-018-0180-5)\\]\n  > IBM's PCM+capacitor analog synapse hits software-equivalent training accuracy at ~100× better energy efficiency than GPUs.\n- Fully Hardware-Implemented Memristor Convolutional Neural Network (**Nature 2020**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs41586-020-1942-4)\\]\n  > Tsinghua integrates eight memristor crossbars into a complete CNN with hybrid training — >100× more energy-efficient than GPUs.\n- The Missing Memristor Found (**Nature 2008**) ★. \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnature06932)\\]\n  > HP Labs' physical realization of Chua's memristor — the device that launched the entire memristive-synapse field.\n- Experimental Demonstration and Tolerancing of a Large-Scale Neural Network (165,000 Synapses) Using Phase-Change Memory (**IEEE TED 2015**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FTED.2015.2439635)\\]\n  > IBM's Burr et al. train a 165k-synapse network on real PCM hardware — a landmark large-scale in-memory demonstration.\n- Neuromorphic Computing with Nanoscale Spintronic Oscillators (**Nature 2017**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fnature23011)\\]\n  > Uses a single spin-torque nano-oscillator to classify spoken digits — opening spintronics as a neuromorphic substrate.\n- Deep Learning Incorporating Biologically Inspired Neural Dynamics and In-Memory Computing (**Nature Machine Intelligence 2020**). \\[[paper](https:\u002F\u002Fwww.nature.com\u002Farticles\u002Fs42256-020-0187-0)\\]\n  > Woźniak et al.'s spiking neural units (SNUs) bring LIF dynamics into deep-learning layers deployable on in-memory hardware.\n\n#### Event Cameras & Neuromorphic Sensors\n\n- A 128×128 120 dB 15 μs Latency Asynchronous Temporal Contrast Vision Sensor (DVS) (**IEEE JSSC 2008**) ★. \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FJSSC.2007.914337)\\]\n  > Lichtsteiner, Posch & Delbruck's Dynamic Vision Sensor — pixels asynchronously emit spikes on brightness change, founding event-based vision.\n- A QVGA 143 dB Dynamic Range Frame-Free PWM Image Sensor (ATIS) (**IEEE JSSC 2011**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FJSSC.2010.2085952)\\]\n  > Combines event-based change detection with per-pixel PWM absolute-intensity encoding, adding grayscale to event vision.\n- A 240×180 130 dB 3 μs Latency Global-Shutter Spatiotemporal Vision Sensor (DAVIS) (**IEEE JSSC 2014**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FJSSC.2014.2342715)\\]\n  > Outputs asynchronous DVS events + synchronous APS frames from a shared photodiode — the most widely used event-camera format.\n- A 1280×720 Back-Illuminated Stacked Temporal-Contrast Event-Based Vision Sensor (**IEEE ISSCC 2020**). \\[[paper](https:\u002F\u002Fieeexplore.ieee.org\u002Fdocument\u002F9063149)\\]\n  > Sony\u002FProphesee scale event cameras to HD with the industry's smallest pixels and >124 dB HDR, enabling commercial deployment.\n- AER EAR: A Matched Silicon Cochlea Pair with Address-Event Representation Interface (**IEEE TCAS-I 2007**). \\[[paper](https:\u002F\u002Fdoi.org\u002F10.1109\u002FTCSI.2006.887979)\\]\n  > A binaural silicon cochlea modeling basilar-membrane filtering and emitting AER spikes — the auditory counterpart to the DVS.\n\n#### Commercial & Large-Scale Platforms\n\n- Intel Hala Point — 1.15B-neuron Loihi 2 system (**Intel 2024**). \\[[link](https:\u002F\u002Fnewsroom.intel.com\u002Fartificial-intelligence\u002Fintel-builds-worlds-largest-neuromorphic-system)\\]\n  > 1,152 Loihi 2 chips → 1.15B neurons \u002F 128B synapses across 140k cores at ~2.6 kW; Intel announced it in 2024 as the world's largest neuromorphic system.\n- IBM NorthPole: Neural Inference at the Frontier of Energy, Space, and Time (**Science 2023**). \\[[paper](https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.adh1174)\\]\n  > A compute-in-memory brain-inspired inference chip that eliminates off-chip memory (not strictly spiking, but a landmark neuromorphic-adjacent design).\n- **Lynxi KA200 \u002F HP-300** — Commercial heterogeneous SNN\u002FANN brain-inspired chip and compute cards\u002Fservers supporting large-scale brain simulation. \\[[info](https:\u002F\u002Fwww.lynxi.com\u002F)\\]\\[[sdk](https:\u002F\u002Fgithub.com\u002FLynxiTech\u002FBIDL)\\]\n  > 灵汐科技 — a commercial neuromorphic platform from China.\n- **BrainChip Akida (AKD1000\u002F1500)** — Commercial event-based neuromorphic SoC running fully on-chip spiking networks at sub-watt power. \\[[info](https:\u002F\u002Fbrainchip.com\u002Fai-for-iot-akida-spiking-neural-network-accelerators\u002F)\\]\n- **Innatera Spiking Neural Processor T1 \u002F Pulsar** — Ultra-low-power analog\u002Fmixed-signal spiking neural MCU for always-on sensor-edge inference. \\[[info](https:\u002F\u002Fopen-neuromorphic.org\u002Fneuromorphic-computing\u002Fhardware\u002Fsnp-by-innatera\u002F)\\]\n- **SynSense Xylo** — Ultra-low-power (hundreds of μW) digital LIF-based spiking inference chip for audio and bio-signal sensing. \\[[info](https:\u002F\u002Fwww.synsense.ai\u002Fproducts\u002Fxylo\u002F)\\]\n- **GrAI Matter GrAIOne \u002F GrAI VIP (NeuronFlow)** — Sparsity\u002Fevent-driven neuromorphic edge SoCs exploiting temporal sparsity (acquired by Snap, 2023). \\[[info](https:\u002F\u002Fwww.forbes.com\u002Fsites\u002Fkarlfreund\u002F2022\u002F05\u002F27\u002Fgrai-matter-labs-brain-inspired-ai-for-the-edge\u002F)\\]\n\n---\n\n\u003Cp align=\"right\">\u003Ca href=\"#table-of-contents\">↑ Back to top\u003C\u002Fa>\u003C\u002Fp>\n\n## Part IV · Applications\n\n### 7 · Applications\n\n> **In one breath:** SNNs shine wherever **power and latency** dominate and data is **naturally temporal\u002Fsparse** — event-camera vision, always-on audio, robotics\u002Fcontrol, and increasingly language models.\n\n#### Event-Based Vision — Recognition & 3D\n\n- A Low Power, Fully Event-Based Gesture Recognition System (**CVPR 2017**) ★. \\[[paper](https:\u002F\u002Fopenaccess.thecvf.com\u002Fcontent_cvpr_2017\u002Fpapers\u002FAmir_A_Low_Power_CVPR_2017_paper.pdf)\\]\n  > IBM's end-to-end DVS + TrueNorth system recognizes gestures at \u003C200 mW, and released the DVS128 Gesture dataset.\n- Spiking PointNet: Spiking Neural Networks for","这是一个面向脉冲神经网络（SNN）领域的综合性双语资源导航库，系统整理了340余篇论文、40余项奠基性工作、42个研究团队、45+开源项目及类脑硬件、数据集与工具。项目以深度标注方式组织内容，覆盖SNN基础理论、训练方法、神经形态硬件、事件相机应用与能效\u002F鲁棒性等关键方向，并提供中英文对照支持。适用于从事类脑计算、低功耗边缘AI、神经形态工程及生物启发式机器学习的研究者与工程师快速检索权威资源与技术脉络。",2,"2026-07-27 02:30:05","CREATED_QUERY"]