2nd/10_Wiki/Topics/AI_and_ML/Spiking-Neural-Networks-SNNs.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-spiking-neural-networks-snns	Spiking Neural Networks (SNNs)	10_Wiki/Topics	verified	self	snn neuromorphic-nn event-driven-nn third-gen-nn	none	A	0.9	applied	neuromorphic snn deep-learning brain-inspired low-power		2026-05-10	pending	language framework python snntorch-norse-lava

Spiking Neural Networks (SNNs)

매 한 줄

"매 neurons fire discrete spikes — 매 event-driven, time-coded compute". SNN 은 매 biological neuron 의 spike train (LIF, Izhikevich) 을 모방한 third-generation neural network — 매 event-driven sparsity 의 ultra-low-power potential. 매 2026 의 Intel Loihi 2, IBM NorthPole, BrainScaleS-2 의 hardware 가 mature 단계.

매 핵심

매 vs ANN

• ANN: continuous activations, synchronous, dense compute.
• SNN: binary spikes, asynchronous events, sparse compute.
• Time as a dimension: 매 information 의 spike timing / rate 에 인코딩.
• Power: 매 event-driven 의 mW-scale (Loihi 2 의 ~74mW).

매 neuron model

• LIF (Leaky Integrate-and-Fire): 매 simplest, hardware-friendly.
• Izhikevich: 매 cortical regular spiking 의 biologically richer.
• Hodgkin-Huxley: 매 ion-channel 수준 — 매 expensive.
• AdEx: adaptive exponential — 매 sweet spot.

매 encoding

• Rate coding: 매 spike count / unit time = real value.
• Temporal coding: 매 first-spike timing.
• Phase / population coding: 매 spike pattern.

매 training

• STDP (Spike-Timing-Dependent Plasticity): 매 biologically plausible, unsupervised.
• ANN-to-SNN conversion: 매 trained ANN 의 weights 의 SNN 으로 transfer.
• Surrogate gradient: 매 backprop 의 의 spike non-differentiability 의 우회 — 매 2026 의 standard.
• e-prop / online learning: 매 BPTT 대체.

매 응용

1. Edge AI — 매 hearing aid, IoT sensor, drone (mW power).
2. Event-camera vision — 매 DVS / Prophesee sensor 의 native fit.
3. Robotics — 매 reflex / closed-loop control.
4. BCI — 매 spike-train 의 native processing.
5. Neuromorphic computing — 매 Loihi 2, NorthPole, SpiNNaker 2.

💻 패턴

LIF neuron (snnTorch, surrogate gradient)

import torch, torch.nn as nn, snntorch as snn

class SpikingMLP(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(784, 256)
        self.lif1 = snn.Leaky(beta=0.95, spike_grad=snn.surrogate.fast_sigmoid())
        self.fc2 = nn.Linear(256, 10)
        self.lif2 = snn.Leaky(beta=0.95, spike_grad=snn.surrogate.fast_sigmoid())

    def forward(self, x, T=25):
        m1 = self.lif1.init_leaky(); m2 = self.lif2.init_leaky()
        spk_rec = []
        for t in range(T):
            cur1 = self.fc1(x); spk1, m1 = self.lif1(cur1, m1)
            cur2 = self.fc2(spk1); spk2, m2 = self.lif2(cur2, m2)
            spk_rec.append(spk2)
        return torch.stack(spk_rec)  # [T, B, 10]

Rate coding

def rate_encode(image, T=25):  # image in [0,1]
    return (torch.rand(T, *image.shape) < image.unsqueeze(0)).float()

Surrogate gradient (custom)

class ATan(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x):
        ctx.save_for_backward(x); return (x > 0).float()
    @staticmethod
    def backward(ctx, grad_out):
        x, = ctx.saved_tensors
        return grad_out * (1 / (1 + (math.pi * x) ** 2))

ANN→SNN conversion (rate-based)

# train ANN with ReLU normally, then replace ReLU with IF neurons
# scale weights so max activation == firing threshold
def convert(ann_model, calibration_loader):
    for layer in ann_model.modules():
        if isinstance(layer, nn.ReLU):
            replace_with_IF(layer, threshold=compute_max_act(layer, calibration_loader))

Event camera + SNN (Prophesee)

from metavision_core.event_io import EventDatReader
events = EventDatReader("recording.dat")
# events: (x, y, polarity, timestamp_us)
spike_tensor = events_to_spike_tensor(events, H=720, W=1280, T=25, dt=10000)
output = snn_model(spike_tensor)

Loihi 2 deployment (Lava framework)

from lava.proc.lif.process import LIF
from lava.proc.dense.process import Dense
from lava.magma.core.run_configs import Loihi2HwCfg

lif = LIF(shape=(256,), du=0.5, dv=0.5, vth=10)
dense = Dense(weights=W)
dense.s_out.connect(lif.a_in)
lif.run(condition=RunSteps(num_steps=100), run_cfg=Loihi2HwCfg())

Norse (deep-learning style)

import norse.torch as norse
model = nn.Sequential(
    norse.LIFCell(),
    nn.Linear(256, 10),
    norse.LIFCell(),
)

매 결정 기준

상황	Approach
Edge inference (mW)	ANN→SNN convert + Loihi 2 / Akida
Event-camera	native SNN with surrogate gradient
GPU prototyping	snnTorch / Norse
Biological research	NEURON / Brian 2
Scale-out neuromorphic	SpiNNaker 2 / BrainScaleS-2

기본값: 매 prototype 은 snnTorch + surrogate gradient. 매 deployment 는 ANN→SNN convert 후 Loihi 2 / Akida AKD1500.

🔗 Graph

• 부모: Neural-Networks · Neuromorphic-Computing
• 응용: Edge-AI · Brain-Computer-Interface
• Adjacent: SpiNNaker

🤖 LLM 활용

언제: literature summarization, surrogate gradient choice 추천, snnTorch boilerplate 생성. 언제 X: hardware-specific tuning (Loihi 2 의 chip-level 의 quirk) — 매 vendor docs / community 의 expert.

❌ 안티패턴

• GPU-only mindset: SNN 의 GPU 의 simulation 은 ANN 보다 slow. 매 power 이득 은 neuromorphic HW 에서.
• Naive BPTT: 매 spike 의 non-differentiability 무시. Surrogate gradient mandatory.
• Too few timesteps: 매 T=5 의 rate coding 의 information 부족. T=25~100 typical.
• Ignoring conversion error: ANN→SNN 의 accuracy drop. Calibration / threshold balancing.
• Wrong neuron model: 매 LIF 만 가지고 의 모든 task fit. Task 별 적합 model 선택.

🧪 검증 / 중복

• Verified (Maass 1997, snnTorch docs 2025, Intel Loihi 2 papers, Tavanaei et al. 2019 review, Eshraghian 2023).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — SNN fundamentals + 2026 neuromorphic stack