Research map / Architectures and efficiency / Spiking networks
Spiking Neural Network as Adaptive Event Stream Slicer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SpikeSlicer, a plug-and-play method where a small spiking neural network decides adaptively when to slice an event stream, trained with a membrane-potential-based SPA-Loss and a feedback-update scheme driven by downstream ANN loss.
Problem
Fixed event-count or fixed-duration slicing of event streams loses temporal information under diverse motion speeds and needs hyperparameters tuned per task, hurting event-based tracking and recognition.
Approach
Use an SNN LIF neuron as a learnable event trigger whose spike time defines the slice boundary; supervise its no-reset membrane potential to fire at a desired step (SPA-Loss), and set that desired step by a neighborhood search using the downstream ANN's loss as feedback, with the ANN then finetuned on the new slices.
Key results
- SpikeSlicer improves TransT on FE108 ALL RSR from 51.0 (fixed event) to 63.6 with SpikeSlicer-S, and reports up to 11.9% gain in tracking.
- On recognition, SpikeSlicer with ResNet-34 improves accuracy by 2.78% on DVS-Gesture and 6.46% on N-Caltech101 over fixed slicing (per the text).
- SpikeSlicer adds only 0.85 mJ energy per the efficiency table (0.32% increase) while raising TransT's RSR, with the paper reporting 22.3% RSR improvement against the no-SpikeSlicer baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace fixed event slicing with adaptive SpikeSlicer-S slicing | TransT · FE108 ALL | RSR: 51.0 -> 63.6 | Helped |
| replace fixed event slicing with adaptive SpikeSlicer slicing | ResNet-34 · DVS-Gesture | accuracy: 93.40 -> 96.18 | Helped |
| replace fixed event slicing with adaptive SpikeSlicer slicing | ResNet-34 · N-Caltech101 | accuracy: 76.08 -> 82.54 | Helped |
| replace fixed event slicing with adaptive SpikeSlicer slicing | ResNet-34 · DVS-CIFAR10 | accuracy: 79.39 -> 82.23 | Helped |
| replace fixed event slicing with adaptive SpikeSlicer slicing | ResNet-18 · SL-Animals | accuracy: 83.93 -> 88.39 | Helped |
| add LA-Loss to Mem-Loss (ablation) | SpikeSlicer with ResNet-18 · SL-Animals | accuracy: 87.50 (+3.57) -> 88.39 (+4.46) | Helped |
| add LA-Loss to Mem-Loss (ablation) | SpikeSlicer with ResNet-34 · SL-Animals | accuracy: 88.52 (+1.02) -> 89.73 (+2.23) | Helped |
| use the small-size rather than base-size slicer network (ablation) | SpikeSlicer with PrDiMP · FE108 | RSR: 59.24 -> 60.88 | Helped |
| add SpikeSlicer | TransT · FE108 | Energy: 259.26 -> 260.11 | Hurt |
| add SpikeSlicer | TransT · FE108 | Speed (s / img): 0.045 -> 0.060 | Hurt |
| add SpikeSlicer | TransT · FE108 | Performance: 51.0 -> 62.4 | Helped |
Surprising findings
- Fixed-slice sensitivity experiments show large variance in accuracy across fixed event counts and durations on N-Caltech101 (e.g., ResNet-34 fixed-count variance 2.90).
- The SNN trigger is fixed at 0.42M parameters for the Small variant, yet it matches or beats the 45.11M Base variant on tracking.
Limitations
- Multi-stage SNN-ANN training leaves room for improvement in the slicing strategy.
- Recognition uses single-frame representations rather than multi-frame ones; experiments run on GPUs rather than neuromorphic hardware, so inference is slower.
Experimental setup
Datasets: FE108 DVS-Gesture N-Caltech101 DVS-CIFAR10 SL-Animals
Benchmarks: FE108 DVS-Gesture N-Caltech101 DVS-CIFAR10 SL-Animals
Metrics: RSR RPR OP0.50 OP0.75 accuracy
Baselines: SiamFC++ KYS CLNet DiMP PrDiMP TaMOs TransT ResNet-18
Builds on: TransT DiMP voxel grid leaky integrate-and-fire
Closest papers by meaning
- SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.89)
- Leveraging Asynchronous Spiking Neural Networks for Ultra Efficient Event-Based Visual Processing (AAAI 2025 · similarity 0.88)
- Fully Spiking Neural Networks for Unified Frame-Event Object Tracking (NeurIPS 2025 · similarity 0.88)
- SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.88)
- CREST: An Efficient Conjointly-trained Spike-driven Framework for Event-based Object Detection Exploiting Spatiotemporal Dynamics (AAAI 2025 · similarity 0.88)
- EventRPG: Event Data Augmentation with Relevance Propagation Guidance (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.