Table of Contents

1. 1

   What Is Neuromorphic AI?Event-driven, sparse, temporal, energy-aware AI; and how to calibrate the “1000× efficiency” claim.
2. 2

   The Computational Problem: Energy, Time, and SparsityWhy memory movement dominates cost; dense vs. sparse computation; the efficiency case for event-driven processing.
3. 3

   The ANN-to-SNN ContinuumHow spiking networks relate to conventional deep learning; where the two frameworks converge and diverge.
4. 4

   Biological Neurons: What We Borrow and WhyAction potentials, synaptic transmission, dendritic computation; what neuroscience contributes and what it does not.
5. 5

   Time as a First-Class DimensionWhy timing matters: latency coding, temporal credit assignment, and the fundamental difference from frame-based AI.

1. 6

   Dynamical Systems for Neural ComputationFixed points, phase portraits, F-I curves, bifurcations, Wilson–Cowan equations, Fokker–Planck, and population-level models.

1. 7

   The Leaky Integrate-and-Fire NeuronMembrane potential dynamics, threshold, reset, refractory period; from equation to PyTorch cell.
2. 8

   Richer Neuron ModelsHodgkin–Huxley, Izhikevich, AdEx, multi-compartment; expressiveness vs. computational cost.
3. 9

   Synaptic DynamicsConductance-based synapses, short-term plasticity, gap junctions, and their SNN implementations.
4. 10

   Stochastic Spiking Networks and NoiseNoise sources, stochastic LIF, probabilistic synapses, and noise as a computational resource.
5. 11

   Spike-Train StatisticsISI distributions, Fano factor, correlation, variability; what makes a good spike-train representation.

1. 12

   How Spikes Represent InformationRate, temporal, phase, and population coding: the four main strategies and their accuracy–latency–energy tradeoffs.
2. 13

   Encoding Data into Spike TrainsRate encoding, latency encoding, delta modulation, and event-stream encoding for images, audio, and time-series data.
3. 14

   Decoding Spike TrainsReadout methods, population vector decoding, and connecting SNN output to downstream decisions.
4. 15

   Sparse Coding and Predictive CodingWhy sparsity is computationally efficient; local learning rules from predictive coding; bridge to Part VI.

1. 16

   Feedforward Spiking Neural NetworksLayer-by-layer SNN classifiers; time-step unrolling; batch normalization for spikes.
2. 17

   Recurrent Spiking Neural NetworksLSTM-analogs, leaky integrators with feedback, memory in spike sequences, and exploding gradient anatomy.
3. 18

   Convolutional Spiking Neural NetworksSpiking CNNs for event-vision; weight sharing across time; efficient inference on DVS data.
4. 19

   Reservoir Computing and Echo State NetworksFixed random recurrent dynamics; liquid state machines; why reservoirs suit hardware-constrained deployment.
5. 20

   Generative and Unsupervised Spiking ModelsSpiking RBMs, Spiking VAEs, energy-based models with spike representations.
6. 21

   Graph Spiking Neural NetworksMessage passing with spike timing; event-based graph processing for point clouds and sensor networks.
7. 64

   Hyperdimensional Computing and Vector-Symbolic ArchitecturesHigh-dimensional binary/bipolar representations; binding and bundling; HDC as a lightweight SNN alternative.

1. 22

   Why Training SNNs Is DifficultNon-differentiability of spikes, vanishing gradient through time, dead neuron problem, and the training landscape.
2. 23

   Spike-Timing-Dependent Plasticity (STDP)Hebbian learning in time; STDP variants; unsupervised feature learning; limits of pure STDP for classification.
3. 24

   Supervised Spike-Timing MethodsSpikeProp, Tempotron, ReSuMe: learning with exact spike times.
4. 25

   Surrogate Gradient LearningStraight-through estimator, sigmoid surrogates, ArcTan, Fast Sigmoid; BPTT through time-steps; the standard training approach.
5. 26

   Normalization in Spiking NetworksBatch normalization for SNNs, threshold-dependent BN (tdBN), layer norm: stabilizing temporal training dynamics.
6. 27

   ANN-to-SNN ConversionWeight rescaling, threshold balancing, conversion pipelines; accuracy–latency tradeoffs at various time-step budgets.
7. 28

   Direct Training: Scaling Depth and EfficiencyDeep surrogate-gradient training; TET baseline (83% DVS-CIFAR10); T-RevSNN reversible training (8.6× memory, 2× speed); ParaRevSNN.
8. 29

   Online Learning and Eligibility Traces / e-propThree-factor learning rules, eligibility traces, e-prop on SpiNNaker2; the online learning alternative to BPTT.
9. 30

   Reinforcement Learning with Spiking NetworksReward-modulated STDP, spike-based actor–critic, neuromorphic RL for robotics.
10. 31

    Continual Learning in Neuromorphic SystemsCatastrophic forgetting, elastic weight consolidation for SNNs, federated neuromorphic learning.
11. 65

    Biologically Plausible Learning Without BackpropagationForward-forward algorithm, target propagation, contrastive Hebbian learning; what “no backprop” costs in accuracy.
12. 66

    SNN Model Compression: Pruning and QuantizationStructured and unstructured pruning for spiking models; binary and ternary weights; hardware-aware compression.
13. 67

    SNN Deployment Pipeline: From Trained Model to HardwareExport via NIR, chip mapping, simulation-to-hardware gap, latency and energy profiling on real chips.

1. 32

   Why Neuromorphic Hardware ExistsVon Neumann bottleneck, in-memory computing, event-driven circuits, and the energy argument.
2. 33

   Digital Neuromorphic ChipsIntel Loihi 2 (1M neurons, 120 MSOPS/mW); Hala Point (1.15B neurons, 128B synapses, 140K cores); SpiNNaker2 and SpiNNcloud (Sandia + Leipzig, 2025); Innatera Pulsar; BrainChip Akida 2; IBM NorthPole; Lava SDK deprecation.
3. 34

   Analog and Mixed-Signal Neuromorphic SystemsBrainScaleS-2, DYNAP-SE2, SynSense Xylo/Speck; the mismatch problem and calibration strategies.
4. 35

   Memristive and Non-Volatile Memory DevicesPCM, RRAM, OTS synapses; in-memory learning; materials and reliability challenges.
5. 36

   Address-Event Representation (AER)AER bus protocol, routing, spike routing in multi-chip systems, and on-chip network topologies.
6. 37

   FPGA and Custom ASIC ImplementationsSNN on FPGAs (resource mapping, timing), neuromorphic ASICs, open-source hardware flows.
7. 38

   Hardware–Software Co-DesignMapping SNN topology to chip constraints, routing-aware training, NIR as the co-design interface.

1. 39

   Event Cameras and Dynamic Vision SensorsDVS operating principle; Prophesee GenX320 and IMX636; OpenEB toolchain; SynSense Speck in-sensor SNN compute.
2. 40

   Event-Vision AlgorithmsDetection and tracking: RVT → SAST → SMamba/PMRVT lineage; event representations; self-supervised pretraining as open problem.
3. 41

   Event-Based Stereo Vision and 3D PerceptionStereo event cameras; DERD-Net (NeurIPS 2025, ≥42% MAE reduction); depth estimation on MVSEC/DSEC.
4. 42

   Neuromorphic Audio and SpeechSilicon cochlea, event-based audio processing, Xylo audio SoC, keyword spotting and speech recognition with SNNs.
5. 43

   Tactile and Multimodal Neuromorphic SensingEvent-based tactile sensors, sensor fusion across event modalities, and multimodal SNN architectures.

1. 44

   Neuromorphic Edge AIKeyword spotting, anomaly detection, always-on sensing; the watt–milliwatt–microwatt hierarchy; TinyML vs. neuromorphic.
2. 45

   Robotics and Autonomous SystemsEvent-camera SLAM, spiking motor control, reflex arcs, and neuromorphic perception–action loops.
3. 46

   Biomedical Monitoring and Neural InterfacesUltra-low-power biosignal processing, implantable SNN inference, brain–computer interfaces.
4. 47

   Industrial IoT and Predictive MaintenanceVibration and acoustic anomaly detection; always-on edge inference; event-based industrial sensing.
5. 48

   Autonomous Vehicles and Drone NavigationHigh-speed obstacle avoidance with event cameras; low-latency optic-flow estimation; real deployments.
6. 68

   Neuromorphic Computing for Combinatorial OptimizationIsing machines, QUBO mapping, simulated annealing on neuromorphic hardware, benchmark comparison.

1. 49

   ANN-SNN Hybrid ModelsANN front-end with SNN temporal backend; Neural ODEs and Liquid Time-Constant networks as continuous-time hybrids.
2. 50

   Spiking Transformers, Attention, and State Space ModelsSpikFormer → Spike-driven Transformer V3 (86.2% ImageNet, T-PAMI 2025); QKFormer; SpikingSSMs (AAAI 2025); P-SpikeSSM; diagonal SSM on Loihi 2.
3. 51

   Knowledge Distillation for SNNsSoft-target distillation from ANN teachers; feature-level alignment; closing the accuracy gap.
4. 52

   Neuromorphic AI and Foundation ModelsSpikeLLM (ICLR 2025, 7–70B params); SpikingBrain-7B/76B (>100× TTFT speedup); SpikeMLLM (Apr 2026, 25.8× power efficiency); NSLLM; BitNet b1.58 as quantization cousin.

1. 53

   Metrics for Neuromorphic AIAccuracy, latency, synaptic operations (SOP), energy–delay product; why single-metric comparison misleads.
2. 54

   Datasets and BenchmarksDVS-CIFAR10, N-MNIST, SHD, MVSEC, DSEC, ECMD, NSAVP; NeuroBench 2.3.0 (May 2026, Nature Comms 2025); Kudithipudi et al. Nature 637 roadmap.
3. 55

   Fair Energy MeasurementOn-chip vs. system-level power; idle power accounting; apples-to-apples comparison methodology.
4. 56

   Limitations of Neuromorphic AIThe SNN–ANN accuracy gap; training instability; hardware fragmentation; when not to use SNNs.
5. 69

   Explainability and Interpretability of SNNsSpike-based saliency, temporal attention visualization, causal analysis of spiking patterns.

1. 57

   On-Chip LearningLocal learning rules that run on hardware; on-chip e-prop; spike-based online gradient descent.
2. 58

   Scaling Neuromorphic SystemsMulti-chip interconnects, wafer-scale integration, and what “scaling laws” mean for SNNs.
3. 59

   Neuromorphic Computing and Computational NeuroscienceBrain simulation, large-scale cortical models, what neuroscience predicts for hardware design.
4. 60

   Security and Robustness of Neuromorphic SystemsAdversarial attacks on SNNs, hardware trojans, fault tolerance of spike-based computation.
5. 61

   Theoretical FoundationsCapacity of spiking networks, VC dimension for spike-timing codes, statistical learning theory for SNNs.
6. 62

   Neuromorphic PhotonicsOptical spiking neurons, photonic integrated circuits for neuromorphic computing, light-speed synapses.
7. 63

   The Future of Neuromorphic AIOpen research questions, convergence with large-model AI, the path from edge sensor to autonomous agent.

▶ Practitioner reading path: start here, then pull from earlier parts on demand.

1. 70

   The Neuromorphic Software Ecosystem and Framework SelectionsnnTorch, Norse, Tonic, Rockpool, Sinabs, Spyx, Nengo: 2026 status, use-case map, decision tree, and the “is this library dead?” checklist. Lava deprecation case study.
2. 71

   Cross-Framework Interoperability with NIRNeuromorphic Intermediate Representation (Nature Comms 2024): graph model, read/write across snnTorch/Norse/Rockpool/Sinabs; hardware targets Loihi 2, SpiNNaker2, Speck, Xylo, BrainScaleS-2; round-trip fidelity recipe.
3. 72

   Performance Engineering for SNN TrainingTime-axis vectorization, torch.compile, JAX scan, AMP, gradient checkpointing, vmap/functorch; T-RevSNN-style O(L)-memory multi-GPU recipe; profiling methodology.
4. 73

   Reproducible Research Engineeringuv/pixi lockfiles, Hydra/OmegaConf configs, W&B/MLflow/Aim tracking, DVC data versioning; gradcheck/gradgradcheck; Hypothesis property tests; NeurIPS reproducibility checklist.
5. 74

   AI-Assisted Development (“Vibe Coding”) for Neuromorphic ResearchLLM-assisted SNN coding workflow; reproducibility discipline (pin model+version, commit transcripts); LLM Guidelines for SE (Baltes et al. 2025); verifying AI-generated numerical code with gradcheck.
6. 75

   The Neuromorphic AI CookbookTen version-pinned runnable recipes: LIF cell, surrogate-gradient classifier, ANN→SNN conversion, NIR round-trip, event-camera pipeline, Speck/Xylo deployment, NeuroBench evaluation, and more.
7. 76

   Capstone: Building a Neuromorphic AI System End-to-EndFull pipeline: event sensor → spike encoding → SNN training (reproducible) → NIR export → hardware deployment → sim-to-hardware gap report.