RIPE-MambaSpike: Resolution-Independent Spiking-State-Space Interfaces for Parameter-Efficient Event-Based Vision

Md Muhiminul Islam; Shoaib Ahmed Dipu; Sayeed Shafayet Chowdhury

cs.CVSep 26, 2026

RIPE-MambaSpike: Resolution-Independent Spiking-State-Space Interfaces for Parameter-Efficient Event-Based Vision

Authors: Md Muhiminul Islam, Shoaib Ahmed Dipu, Sayeed Shafayet Chowdhury

Organizations: Independent Researcher · Indiana University

Abstract

Spiking-Mamba hybrids reach strong accuracy on event-based vision, but existing designs often require tens of millions of parameters. Much of that cost comes from how the spiking front-end is connected to the state-space backbone rather than from the hybrid architecture itself. In a representative model, a single resolution-dependent projection accounts for 33.55M of 36.25M parameters. To that end, we introduce RIPE-MambaSpike (Resolution-Independent, Parameter-Efficient), which replaces that projection with a hierarchical multi-resolution bridge of fixed channel width. Its deployed footprint is 0.870M parameters, constant at fixed time steps and widths across a 43x range of input areas. Reparameterized spiking stages, temporal decoupled modulation, and a dynamic convex-hull-bounded dual-stream membrane-potential attention preserve accuracy under this compact design. Result-wise, RIPE-MambaSpike is pareto-optimal on CIFAR10-DVS, N-Caltech101, and DailyDVS-200. Notably, on the 200-class DailyDVS-200, a scaled 8.04M configuration achieves 45.7% top-1 accuracy, the best reported spiking result on that benchmark, and outperforms prior spiking methods with 3.0-15.1x fewer parameters than dense ANNs. Overall, our findings demonstrate that competitive event-based recognition does not require resolution-dependent parameter growth. Code is available at https://github.com/MuhiminOsim/RIPE-MambaSpike.

Figures & tables

Figure 1: Overview of RIPE-MambaSpike (DVS-Gesture). (a) Three RepSpikeStage blocks (TDM

+

reparameterizable convolutions

+

DS-MPA, stride-2) emit the final attention map

a_{3}

and firing rates. (b) Multi-scale rates/increments are fused (MR-S3A) into a fixed-width token sequence. (c) Three gated bidirectional Mamba blocks process the tokens, each transition pooling the spatial axis

\times 2

and doubling width.

Figure 2: The SNN

\to

SSM interface’s parameter cost is quadratic in sensor resolution; RIPE-MambaSpike’s is constant. Markers are instantiated and counted (Appendix C.1 ); the curve is Eq. ( 1 ), which reproduces all four exactly, dashed where extrapolated. At the

1280{\times}720

of an HD event sensor that one layer reaches

1.89

B parameters (

{\approx}7

GiB fp32 ); RIPE-MambaSpike is tested flat at

0.870

M from

34^{2}

224^{2}

Method	$T$	Params	Acc. (%)
DVS-Gesture ( $T{=}16$ )
TET Deng et al. (2022)	16	3.66 M	97.92 ∗
SpikFormer Zhou et al. (2023)	16	2.57 M	98.30 ∗
SDT Yao et al. (2023)	16	2.57 M	99.30 ∗
MS-ResNet Hu et al. (2024)	16	11.20 M	98.61 ∗
RTFormer Wang et al. (2024a)	16	—	98.61 ∗

Table 1: State-of-the-art comparison across all five event benchmarks (inference-time parameters). ∗ published; ‡ / † the released Mamba-Spike code under its own / our protocol; § no parameter count is attributable (Appendix C.3 ). Among general-purpose backbones, RIPE-MambaSpike is smallest in three of four and Pareto-optimal across datasets (action-specific SpikMamba Chen et al. (2024) is smaller at

0.18

M on DVS-Gesture), leading nowhere on accuracy: it concedes at most

2.9

pp to models

1.7

–

41\times

its size,

6.25

pp to the best DailyDVS-200 result. DailyDVS-200 baselines are frame-based video models with no spiking time-step, so their

T

column reads “–”. Best per block underlined .

Figure 3: Accuracy vs. inference-time parameters (log scale). Panels plot exactly the Table 1 entries reporting a parameter count and label the frontier, so the DailyDVS-200 panel’s visible ceiling is EvMamba (

49.65

) not the benchmark best (

51.99

); its dotted line marks

50\%

top-1. RIPE-MambaSpike (filled star) is Pareto-optimal on CIFAR10-DVS, N-Caltech101 and DailyDVS-200, but not on DVS-Gesture, where SpikMamba is smaller and more accurate, nor on the saturated N-MNIST. Hollow stars are rescaled variants: off-frontier on CIFAR10-DVS, but on DailyDVS-200 the lite config is itself a frontier point.

#	Ablation	DVSG ( $\Delta$ )	$\Delta$ C10	$\Delta$ NCal
A1	– DS-MPA $\to$ id.	97.73 (–0.45)	–0.80	–3.16
A2	– RepConv $\to$ $3{\times}3$	97.73 (–0.45)	–0.60	–1.94
A3	– TDM ( $\lambda{=}0$ )	98.11 (–0.07)	–1.40	–2.43
A4	– MR-S3A $\to$ st.-3 only	98.11 (–0.07)	–1.10	–1.46
RIPE-MambaSpike (full)		$\mathbf{98.18\pm 0.31}$	81.40	85.54

Table 2: Component-wise ablation and cross-dataset transfer. DVSG

\Delta

uses the 5-seed mean; C10/NCal are single-seed.

Module	DVS-Gest. ( $\Delta$ )	N-Cal101 ( $\Delta$ )
Hier. Mamba (ours)	$\mathbf{98.18\pm 0.31}$	$\mathbf{85.49\pm 0.18}$
Transformer	$97.65\;(-0.53^{*})$	$82.75\;(-2.75^{***})$
GRU	$97.73\;(-0.45^{*})$	$82.48\;(-3.01^{***})$
Conv1D	$97.96\;(-0.22^{\textrm{n.s.}})$	$82.24\;(-3.26^{***})$
None	$97.20\;(-0.98^{***})$	$79.44\;(-6.05^{***})$

Table 3: Parameter-matched sequence-module ablation (5 seeds/cell, mean

\,\pm\,

s.d.). Each arm is matched to within

0.1\%

of hierarchical-Mamba (None is

23.7\%

smaller).

{}^{***}p{<}0.001

{}^{*}p{<}0.05

, n.s.

p{\geq}0.05

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Design axis	Mamba-Spike	RIPE-MambaSpike
Reparameterized convs	×	✓
Per-channel temporal filter	×	✓(TDM)
Linear membrane attention	×	✓(DS-MPA)
Fixed-width SNN $\to$ SSM bridge	×	✓(MR-S3A)
Multi-resolution SNN bridge	×	✓
Inference-fused $3{\times}3$ kernel	×	✓

Appendix

Table 4: Architectural comparison of RIPE-MambaSpike and Mamba-Spike.

Figure 4: RepSpikeStage internals. (a) Two TDM

\to

RepConv

\to

BNTT

\to

neuron units. (b) The seven-branch RepConv fuses to a single

3{\times}3

kernel (Prop. 1 ). (c) DS-MPA, a dynamic convex-hull bounded linear attention with

\phi{=}\mathrm{ELU}{+}1

, queries/keys from the membrane

U

, values from the previous spike tensor.

Figure 5: Dual-branch training with the SGC loss. Coral marks training-time-only components removed at deployment. During the continuous pass the BatchNorm momentum is frozen so running statistics are not updated twice per step.

Component	$P^{\text{train}}$	$P^{\text{infer}}$
RepSpikeStage backbone (6 RepConv units)	1.014 M	0.368 M
Hierarchical Mamba + MR-S3A bridge + classifier	0.508 M	0.508 M
Full model	1.522 M	0.875 M

Appendix

Table 5: Train-time vs. inference-time parameter count of RIPE-MambaSpike (hierarchical configuration, stage channels

[32,64,128]

, hierarchical Mamba level dimensions

(32,64,128)

). Structural reparameterization fuses the multi-branch spiking backbone into single

3{\times}3

kernels, reducing backbone parameters by

\sim

2.8\times

at inference and yielding a deployed model of 0.875 M parameters (Proposition 1 ).

Input	RIPE-MambaSpike	Mamba-Spike
resolution	deployed (M)	total (M)	input_proj (M)	share
$34\times 34$	0.870	4.79	2.10	43.8%
$48\times 48$	0.870	7.41	4.72	63.6%
$128\times 128$	0.870	36.25	33.55	92.6%
$224\times 224$	0.870	105.46	102.76	97.4%
growth	$\mathbf{1.00\times}$	$22.0\times$	$49.0\times$	—

Appendix

Table 6: The SNN

\to

SSM interface’s parameter cost scales with input resolution; RIPE-MambaSpike’s does not. Both models instantiated and counted directly at each resolution (RIPE-MambaSpike after fusion, i.e. the deployed graph); configuration held fixed down each column, so this isolates resolution alone.

Bridge	Total params	input_proj	Top-1
Flatten-then-project (released)	$36.25$ M	$33.55$ M ( $92.6\%$ )	$48.90$
Pool to $8{\times}8$ , then project	$\mathbf{4.80}$ M	$\mathbf{2.10}$ M ( $43.7\%$ )	$\mathbf{53.30}$

Appendix

Table 7: Interface swap on CIFAR10-DVS at

128^{2}

, released Mamba-Spike code under its own protocol. Parameters are counted from the trained checkpoints.

Model	Params	FP16 weights
RIPE-MambaSpike (small benchmarks)	$0.870$ M	$\mathbf{1.7}$ MB
RIPE-MambaSpike lite (DailyDVS-200)	$3.82$ M	$\mathbf{7.6}$ MB
RIPE-MambaSpike (DailyDVS-200)	$8.04$ M	$\mathbf{16.1}$ MB
TSM	$24.3$ M	$48.6$ MB
Swin-T	$27.8$ M	$55.6$ MB
Mamba-Spike	$36.25$ M	$72.5$ MB

Appendix

Table 8: Deployed weight storage at FP16 (2 bytes per parameter). Parameter counts are the deployed/fused values used throughout; baseline counts are as cited in the corresponding SOTA table. Storage is weights only; activations and the runtime are additional.

Hyperparameter	Default	DVSG	C10-DVS	NCal101	NMNIST	DDVS-200
Architecture
Stage channels $[c_{1},c_{2},c_{3}]$	$[32,64,128]$	—	—	—	—	$[64,128,256]$ (lite) / $[96,192,384]$
Hier. level dims	$(32,64,128)$	—	—	—	—	$(64,128,256)$ (lite) / $(96,192,384)$
Hier. pool ratios	$(1,2,2)$	—	—	—	—	—
$d_{\text{state}}$	$16$	—	—	—	—	—
Drop-path rate	$0.1$	—	$0.2$	$0.2$	—	$0.2$

Appendix

Table 9: Per-dataset hyperparameters for RIPE-MambaSpike. “—” means identical to “Default”.

Level	Model dim $d_{\text{m}}$	Tokens $k_{\ell}$	State $d_{\text{state}}$	Expansion
1	32	256	16	2
2	64	128	16	2
3	128	64	16	2

Appendix

Table 10: Per-level configuration of the hierarchical bidirectional Mamba backbone (DVS-Gesture,

H_{3}{=}W_{3}{=}16

). Each level is a forward

+

backward scan pair; “tokens” is the retained spatial-token count (preserved across all

T

Component removed	DVSG	CIFAR10-DVS	NCal101
DS-MPA	$1.5\sigma$	$7.3\sigma$	$17.4\sigma$
RepConv	$1.5\sigma$	$5.5\sigma$	$10.7\sigma$
TDM	$0.2\sigma$	$12.8\sigma$	$13.4\sigma$
MR-S3A	$0.2\sigma$	$10.0\sigma$	$8.0\sigma$

Appendix

Table 11: Ablation effect sizes in units of each dataset’s own five-seed

\sigma

. DVS-Gesture

\sigma{=}0.31

pp, CIFAR10-DVS

0.11

pp, N-Caltech101

0.18

pp. Effects below

{\approx}2\sigma

are not claimed.

Configuration (Stage 1, 2, 3)	Acc. (%)	Params (M)	SR-var
CSiLIF, SiLIF, SiLIF (ours)	98.48	0.88	0.0005
SiLIF, SiLIF, CSiLIF	98.11	0.88	0.0004
SiLIF, CSiLIF, CSiLIF	98.11	0.88	0.0003
LIF, SiLIF, SiLIF	97.73	0.88	0.0003
LIF, LIF, LIF	97.73	0.88	0.0009
SiLIF, SiLIF, SiLIF	97.35	0.88	0.0002

Appendix

Table 12: Heterogeneous neuron-mix ablation on DVS-Gesture. Each row assigns a neuron family

\{

LIF, SiLIF, CSiLIF

\}

to each of the three RepSpikeStages. “SR-var” is the cross-epoch standard deviation of the mean spike rate during the last 50 epochs (lower = more stable). All rows are measured single-seed runs; the deployed configuration ( bold ) attains the highest accuracy, while training is comparably stable across all mixes (SR-var

\leq 10^{-3}

). Rows sorted by accuracy.

$k$	32	64	128	256 (full, deployed)
Tokens / sample	512	1024	2048	4096
Acc. (%)	97.73	98.48	98.48	98.48
$\Delta$ (pp)	–0.75	0.00	0.00	—

Appendix

Table 13: Why the deployed model prunes no tokens. An earlier version of this architecture selected a top-

k

subset of spatial tokens before the scan; this sweep is why we removed it.

k

is the number of spatial tokens retained per timestep and the full sequence is

N{=}H_{3}W_{3}{=}256

, so

k{=}256

is the unpruned model. Accuracy is identical for

k\in\{64,128,256\}

: pruning to a quarter of the tokens bought no accuracy. It also bought no parameters — the resolution invariance of Table 6 holds with pruning disabled, since Mamba blocks are

\mathcal{O}(d_{\text{m}}^{2})

and independent of sequence length — and little compute, an

8\times

change in sequence length moving total FLOPs by

11\%

(Table 20 ). The deployed model therefore scans every token. Caveat: the DVS-Gesture split has

0.38

pp granularity per clip, so the tied entries are indistinguishable at this sample size rather than provably equal;

k{=}32

costing

0.75

pp is two clips.

$T$	4	8	12	16	20
Acc. (%)	94.32	96.59	97.73	98.48	98.48
MACs (M)	430.5	861.0	1291.4	1721.9	2152.4

Appendix

Table 14: Sensitivity to time bins

T

on DVS-Gesture. All measured single-seed. MACs rather than SOPs because TDM makes backbone convolutions dense (§ 5 ).

Figure 8: Sensitivity to the spike-rate L1 coefficient

\lambda_{\text{L1}}

on DVS-Gesture. The default

10^{-4}

gives the highest accuracy at a low spike rate; accuracy and

\bar{r}

are otherwise nearly flat across four orders of magnitude.

Figure 9: Per-stage spike rate evolution (N-Caltech101; the same depth-dependent pattern holds on DVS-Gesture): stage 1 stays near

\bar{r}\approx 0.2

, deeper stages settle well below it.

Figure 10: Stage-3 DS-MPA attention $A_{t}$ (mean over channels) at $t\in\{2,6,10,14\}$ on a DVS-Gesture “hand-clap” sample. Attention is broadly distributed and strongest at the mid-gesture motion peak ( $t{=}6$ ), then localizes as the hands meet ( $t{=}14$ ); values are dynamic convex-hull bounded (Prop. 2 , strictly within $[0,1]$ for unprojected binary spikes).

Class	Acc. (%)
hand-clap, R/L-hand wave, R/L-arm cw, R/L-arm ccw, arm-roll, other	100.00
air-drums	91.67
air-guitar	91.67
Mean	98.48

Appendix

Table 15: Per-class top-1 accuracy of RIPE-MambaSpike on DVS-Gesture.

Channel-mixing operator	Params	Fwd (ms)	Fwd+bwd (ms)	Act. (MiB)
DS-MPA (ours)	$64{,}736$	$11.48$	$32.26$	$790.6$
$1{\times}1$ conv $+$ sigmoid gate	$21{,}728$	$3.50$	$9.75$	$336.0$
removed (identity)	$0$	$0.29$	$1.30$	$112.0$
does the $O(NC^{2})$ term worsen with width? DS-MPA at $B{=}1$ :
$[32,64,128]$	$64{,}736$	$3.11$	$8.87$	$197.7$
$[64,128,256]$	$258{,}496$	$5.56$	$16.12$	$397.6$

Appendix

Table 16: DS-MPA cost, isolated. Measured at the exact

(B,T,C,H,W)

each DS-MPA instance receives in the deployed DVS-Gesture model (

B{=}4

— the batch the recipe trains at —

T{=}16

128^{2}

input), summed over the three stages on one RTX 3050; CUDA-event timing, 30 reps after warm-up, activation = peak allocated during one forward+backward. removed returns

X

unchanged with

A{=}0

, so the downstream

A_{t}

projection is inert. Against a full training step of

405.8

ms, DS-MPA is

7.9\%

of step time and

4.3\%

of parameters; substituting the cheap gate recovers only

5.5\%

of step time while forfeiting the

3.16

pp DS-MPA contributes on N-Caltech101 (Table 2 ) — the largest single-component effect we measure. Two properties make this more than a single-GPU anecdote. First, the DS-MPA/gate ratio is

3.30/3.33/3.31/3.28

B{=}1/2/4/8

, stable to

\pm 0.8\%

across an

8\times

change in utilisation, and cost scales linearly with

B

from

B{=}1

, so the operator is compute-bound rather than launch-bound: the ratio reflects arithmetic, not this device’s idle time. Absolute latencies are of course hardware-specific; the parameter and activation figures are not. Second, the

O(NC^{2})

term does not grow with channel width here:

N

falls

16\times

C

rises

4\times

, so

NC^{2}

is constant (

4.19

M per stage) and measured forward time decreases with depth.

Model	Weights	$B{=}1$ lat.	$B{=}8$ lat.	$B{=}8$ peak / act.
Mamba-Spike (baseline)	138.3 MiB	38.2 ms	51.5 ms	456 / 272 MiB
RIPE-MambaSpike, training graph	5.9 MiB	36.5 ms	159.3 ms	1073 / 1043 MiB
RIPE-MambaSpike, deployed	3.4 MiB	34.2 ms	103.8 ms	1070 / 1043 MiB

Appendix

Table 17: Measured deployment cost on DVS-Gesture (

T{=}16

128{\times}128

, fused; CUDA events, median of 30 runs after 10 warm-ups, RTX 3050). “Weights” is parameters

+

buffers; “act.” is peak allocation minus weights. RIPE-MambaSpike is

41\times

smaller in weights and its fused graph is

1.53\times

faster than its training graph, but its activations are larger than the baseline’s — the trade is weight memory for activation memory and batched throughput.

DS-MPA parameterization	Params	vs. dep.	Bound	Deployed total
deployed : $3\times(C\!\to\!C)$	$64{,}736$	$1.00\times$	✓	$\mathbf{875{,}122}$
low-rank: $3\times(C\!\to\!\tfrac{C}{4}\!\to\!C)$	$32{,}480$	$0.50\times$	✓	$842{,}866$ ( $-3.7\%$ )
grouped: $3\times(C\!\to\!C)$ , $g{=}4$	$16{,}352$	$0.25\times$	✓	$826{,}738$ ( $-5.5\%$ )

Appendix

Table 18: Cheaper DS-MPA parameterizations are available and preserve the bound, but we do not adopt them. The bound is structural:

\phi\geq 0

and the normalizer make

A_{i,\cdot}

a convex combination of the rows of

V

(Prop. 2 ), which holds for any factorization of

W_{Q},W_{K},W_{V}

— verified at every stage width by checking each

A_{i,c}

against the componentwise range of

V_{\cdot,c}

that variant produces (

9/9

). So the

O(C^{2})

projection cost can be reduced without theoretical loss. We nevertheless deploy the full operator, for a reason the last column makes explicit: DS-MPA is only

7.4\%

of the deployed

875{,}122

parameters, so even the

4\times

reduction moves the headline figure by

5.5\%

, while removing DS-MPA altogether costs

3.16

pp on N-Caltech101 (Table 2 ) — the largest single-component effect we measure. We are not willing to risk a measured

3.16

pp for an unmeasured

5.5\%

of weight storage, and we have not trained these variants, so we make no accuracy claim for them. They are reported as a sound design option, not as a result.

Input resolution	Deployed (M)	Train (samp/s)	Infer (samp/s)	Peak (GiB)
$34\times 34$	0.875	$25.4$	$102.0$	$0.27$
$48\times 48$	0.875	$25.4$	$83.4$	$0.45$
$128\times 128$	0.875	$10.3$	$50.1$	$2.67$
$224\times 224$	0.875	$0.8^{\dagger}$	$18.2$	$8.00^{\dagger}$

Appendix

Table 19: What resolution-independence does and does not buy. Complements Table 6 : deployed parameters are constant across a

43\times

range in pixel count, but activation memory and latency are not — they scale with resolution as in any convolutional model. Measured on one RTX 3050,

T{=}16

, batch

4

, RIPE-MambaSpike’s DVS-Gesture configuration; parameters counted after fusion. † At

224^{2}

the training step reaches this card’s

8

GiB capacity, so the training throughput there is allocator-limited rather than a clean measurement (inference, which needs no gradients, still runs at

18.2

samples/s). This is the trade the paper claims and no more: weight storage is decoupled from resolution, total memory is not.

Configuration	Params (M)	GFLOPs	GMACs
deployed ( $T{=}16$ , $128^{2}$ )	$\mathbf{0.875}$	$\mathbf{9.00}$	$\mathbf{4.50}$
SSM sequence length (tokens entering the scan; $T{=}16$ , $128^{2}$ )
$L/T{=}32$	$0.875$	$8.11$	$4.05$
$L/T{=}128$	$0.875$	$8.49$	$4.25$
$L/T{=}256$	$0.875$	$9.00$	$4.50$
time bins $T$ ( $128^{2}$ )

Appendix

Table 20: Compute cost and how it scales with

k

T

C

and resolution. FLOPs are measured with PyTorch’s flop counter on the fused inference graph (torch’s convention: one multiply–accumulate is

2

FLOPs; MACs

=

FLOPs

/2

), so the selective scan’s matmuls and the DS-MPA contraction are counted as executed rather than derived from a formula. Batch

1

; the deployed DVS-Gesture point is

[32,64,128]

T{=}16

128^{2}

, all tokens scanned. Three things are worth reading off this table. First , an

8\times

change in the SSM sequence length (

32\!\to\!256

tokens per timestep) costs only

11\%

more compute, because the convolutional front-end dominates and not the scan — so the scan’s length is not the lever that matters here. Second , cost is exactly linear in

T

and exactly quadratic in

C

(

4\times

per width doubling). Third , deployed parameters are flat across resolution while compute is not: this is the same trade Table 19 makes for memory. The small parameter variation down the

T

block is BNTT, whose affine terms are per-timestep. This is a compute count and not an energy claim: TDM precedes each backbone convolution, so accumulate-only accounting does not apply (§ 5 ), and arithmetic counts say nothing about memory traffic.

Event-Based Vision

Event-based vision uses asynchronous camera measurements to perceive motion and scene changes with high temporal resolution and dynamic range.

See all 232 papers

Apr 28, 2026

Vision SmolMamba: Spike-Guided Token Pruning for Energy-Efficient Spiking State-Space Vision Models

Spiking Transformers have shown strong potential for long-range visual modeling through spike-driven self-attention. However, their quadratic token interactions remain fundamentally misaligned with the sparse and event-driven nature of spiking neural computation. To address this limitation, we propose Vision SmolMamba, an energy-efficient spiking state-space architecture that integrates spike-driven dynamics with linear-time selective recurrence. The key idea is a Spike-Guided Spatio-Temporal Token Pruner (SST-TP), which estimates token importance using both spike activation strength and first-spike latency. This mechanism progressively removes redundant tokens while preserving salient spatio-temporal information, enabling efficient scaling with token sparsity. Based on this mechanism, the proposed SmolMamba block incorporates spike events directly into bidirectional state-space recurrence, forming a spiking state-space vision backbone for efficient long-range modeling. Extensive experiments on both static and event-based benchmarks, including ImageNet-1K, CIFAR10/100, CIFAR10-DVS, and DVS128 Gesture, demonstrate that Vision SmolMamba consistently achieves superior accuracy-efficiency trade-offs. In particular, it reduces the estimated energy cost by at least 1.5x compared with prior spiking Transformer baselines and a Spiking Mamba variant while maintaining competitive or improved accuracy. These results demonstrate that combining spike-guided token sparsity with state-space modeling offers a scalable and energy-efficient paradigm for spiking vision systems.

Dewei Bai, Hongxiang Peng, Yunyun Zeng +3

School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu 610054, China · College of Computer Science, Sichuan University, Chengdu 610065, China

Efficient Neural Network Inference Visual SSMs Spiking Neural Networks Visual Token Pruning State Space Models

Ask PDF

Oct 12, 2026

Removing the Compute Bottleneck for Event-based Visual Place Recognition with Spiking Neural Networks

Neuromorphic sensing and computing promise low latency and high energy efficiency for robotics, yet most event-based visual place recognition (VPR) pipelines reintroduce dense, frame-like computation downstream of an inherently sparse sensor, forfeiting the sensor's latency advantage. In this paper, we address the problem of resolving an event-based VPR query in tens of microseconds on a commodity CPU. We train a two-layer spiking neural network (SNN) with additive spike-timing-dependent plasticity (STDP), which drives the learned weight matrix towards a strongly bimodal distribution, amenable to aggressive magnitude pruning. At inference, our pipeline exploits this sparsity by reading only the pruned weights corresponding to the query's active pixels. As surviving weights share a near-uniform magnitude, each neuron's score reduces to an integer count of matching synapses instead of a weighted sum, lowering per-query latency. This computation is small enough to remain resident in a CPU's fast on-chip cache. On three event-based VPR benchmarks, our sparse SNN exceeds the recall of LENS, an existing event-based SNN, while reducing per-query latency by 5.5-6.2x to tens of microseconds. It matches the recall of a dense variant of our network and of the sum-of-absolute-differences (SAD) baseline, while reading 10x fewer weights than the dense variant. We further characterise how recall and latency scale with the number of active pixels and reference places. These results suggest that the sparsity of both event-based input and learned weights can already be exploited for low-latency inference on conventional processors.

Somayeh Hussaini, Tobias Fischer, Michael Milford

Visual Place Recognition Efficient Neural Network Inference Spiking Neural Networks

Ask PDF

Mar 24, 2026

Event-Driven Neuromorphic Vision Enables Energy-Efficient Visual Place Recognition

Reliable visual place recognition (VPR) under dynamic real-world conditions is critical for autonomous robots, yet conventional deep networks remain limited by high computational and energy demands. Inspired by the mammalian navigation system, we introduce SpikeVPR, a bio-inspired and neuromorphic approach combining event-based cameras with spiking neural networks (SNNs) to generate compact, invariant place descriptors from few exemplars, achieving robust recognition under extreme changes in illumination, viewpoint, and appearance. SpikeVPR is trained end-to-end using surrogate gradient learning and incorporates EventDilation, a novel augmentation strategy enhancing robustness to speed and temporal variations. Evaluated on two challenging benchmarks (Brisbane-Event-VPR and NSAVP), SpikeVPR achieves performance comparable to state-of-the-art deep networks while using 50 times fewer parameters and consuming 30 and 250 times less energy, enabling real-time deployment on mobile and neuromorphic platforms. These results demonstrate that spike-based coding offers an efficient pathway toward robust VPR in complex, changing environments.

Geoffroy Keime, Nicolas Cuperlier, Benoit R. Cottereau

IPAL, CNRS IRL 2955, Singapore · Univ Toulouse, CNRS, CerCo UMR 5549, Toulouse, France · Laboratoire ETIS UMR 8051, CY Cergy-Paris Université, ENSEA, CNRS, Cergy, France

Visual Place Recognition Neuromorphic Computing Spiking Neural Networks Efficient Neural Network Inference Event-Based Vision

Ask PDF

Abstract

Event-Based Vision

Explore similar work

Vision SmolMamba: Spike-Guided Token Pruning for Energy-Efficient Spiking State-Space Vision Models

Removing the Compute Bottleneck for Event-based Visual Place Recognition with Spiking Neural Networks

Event-Driven Neuromorphic Vision Enables Energy-Efficient Visual Place Recognition