Research map / Architectures and efficiency / Spiking networks
EventRPG: Event Data Augmentation with Relevance Propagation Guidance
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 SLTRP and SLRP, layer-time-wise and layer-wise relevance propagation rules for spiking neural networks, and uses them to guide event dropping and mixing (EventRPG) for data augmentation on event-based classification.
Problem
Overfitting in SNN event-based classification, and the lack of saliency/CAM extraction methods for SNNs needed to guide saliency-based data augmentation.
Approach
Extends layer-wise relevance propagation to spiking layers by propagating relevance backward through time (SLTRP) or with time-averaged contributions (SLRP), then uses the resulting saliency maps to guide event dropping (RPGDrop) and a CutMix-style event mixing (RPGMix) with label mixing weighted by bounding-box area.
Key results
- EventRPG reaches 85.62% on N-Caltech101 and 85.55% on CIFAR10-DVS (Spike-VGG11), improving over identity by 9.92% and 6.7%.
- EventRPG (Saliency Map, SLRP) reaches 91.59% on SL-Animals 4 sets with SEW Resnet18, a 3.82% gain over the second-best augmentation method.
- SLRP/SLTRP saliency maps achieve top A.I./A.D. on N-Caltech101 and CIFAR10-DVS among compared CAM methods, though Grad-CAM++ wins on N-Cars.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace no augmentation with EventRPG using saliency-map guidance | Spike-VGG11 · N-Caltech101 | accuracy: 75.70 -> 85.62 | Helped |
| Replace no augmentation with EventRPG using CAM guidance | Spike-VGG11 · CIFAR10-DVS | accuracy: 78.85 -> 85.55 | Helped |
| Replace no augmentation with EventRPG using saliency-map guidance | Spike-VGG11 · N-Cars | accuracy: 94.92 -> 96.00 | Helped |
| Use EventRPG with saliency-map guidance instead of EventMix | Spike-VGG11 and Pre-Act Resnet18 · N-Cars | accuracy: 96.29 -> 96.00 | Hurt |
| Use EventRPG with saliency-map guidance instead of no augmentation | SEW Resnet18 · SL-Animals (4 sets) | accuracy: 85.42 -> 91.59 | Helped |
| Use EventRPG with SLTRP saliency-map guidance instead of no augmentation | SEW Resnet18 · SL-Animals (3 sets) | accuracy: 89.09 -> 93.75 | Helped |
| Use EventRPG with SLTRP saliency-map guidance instead of no augmentation | SEW Resnet18 · DVSGesture | accuracy: 94.33 -> 96.53 | Helped |
| Use EventRPG instead of NDA | SEW Resnet18 · mini N-ImageNet | Top-1 accuracy: 35.84 -> 40.90 | Helped |
| Use EventRPG instead of NDA | SEW Resnet18 · mini N-ImageNet | Top-5 accuracy: 63.64 -> 67.74 | Helped |
| Use RPGMix instead of Identity | Spike-VGG11 · N-Caltech101 | accuracy: 75.70 -> 81.75 | Helped |
| Use Saliency Mix instead of Identity | Spike-VGG11 · N-Caltech101 | accuracy: 75.70 -> 72.63 | Hurt |
| Use RPGMix instead of Puzzle Mix (mask only) | SEW Resnet18 · SL-Animals (4 sets) | accuracy: 85.68 -> 88.67 | Helped |
Surprising findings
- Grad-CAM++ outperforms the proposed relevance methods on N-Cars, attributed to its binary classification nature.
- Saliency Mix and Puzzle Mix sometimes underperform the identity (no augmentation) baseline on SNNs.
Limitations
- Only applicable to classification tasks.
- Less effective on N-Cars, a binary dataset where background samples lack label-related objects to localize.
- SLTRP-based saliency is costlier than SLRP since it propagates through all layers and time steps.
Experimental setup
Datasets: N-Caltech101 CIFAR10-DVS N-Cars mini N-ImageNet DVS128 Gesture SL-Animals-DVS
Benchmarks: N-Caltech101 CIFAR10-DVS N-Cars DVSGesture SL-Animals mini N-ImageNet
Metrics: accuracy Average Increase (A.I.) Average Drop (A.D.) time cost (s) Top-1 accuracy Top-5 accuracy
Baselines: SAM Grad-CAM Grad-CAM++ EventDrop NDA EventMix SALT TET Saliency Mix Puzzle Mix
Builds on: layer-wise relevance propagation alpha-beta rule contrastive LRP EventDrop NDA TET
Closest papers by meaning
- Spiking Neural Network as Adaptive Event Stream Slicer (NeurIPS 2024 · similarity 0.87)
- SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.87)
- SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.86)
- EventZoom: A Progressive Approach to Event-Based Data Augmentation for Enhanced Neuromorphic Vision (AAAI 2025 · similarity 0.85)
- Leveraging Asynchronous Spiking Neural Networks for Ultra Efficient Event-Based Visual Processing (AAAI 2025 · similarity 0.85)
- Efficient Event-Based Object Detection: A Hybrid Neural Network with Spatial and Temporal Attention (CVPR 2025 · similarity 0.85)
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.