Research map / Graph learning / Knowledge and temporal graphs
A Dynamic GCN with Cross-Representation Distillation for Event-Based Learning
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 EDGCN, a dynamic graph CNN for event data whose neighborhoods use all vertex attributes, plus a frame-to-graph cross-representation distillation (CRD) framework that transfers knowledge from a frame-based CNN teacher without extra data.
Problem
Point-based graph models for event cameras underperform frame-based models because neighborhood definitions ignore semantic, spatial and temporal attributes jointly, and graph models lack robust pretrained weights and labeled data.
Approach
A dynamic aggregation layer (EDAL) builds neighborhoods from a fused joint representation of coordinate and semantic features, computes attention weights from coordinates only, and updates coordinates layer by layer; a hybrid frame-to-graph distillation (logit-level distillation plus NT-Xent feature contrast) transfers knowledge from a CNN teacher built on the same events.
Key results
- EDGCN with CRD reaches 0.835 on N-Caltech101 and 0.752 on CIFAR10-DVS classification, the best among graph-based methods compared, though still below pretrained MVF-Net (0.871 on N-Cal).
- On DVS128 Gesture, EDGCN achieves 0.985 accuracy with 0.72M parameters and 0.14 GFLOPs, versus 0.981 for the I3D-R teacher.
- On N-Caltech101 detection, EDGCN with CRD reaches 0.711 mAP versus 0.595 for AEGNN and 0.643 for Asynet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cross-representation distillation | EDGCN · N-Cal | accuracy: 0.801 -> 0.835 | Helped |
| Add cross-representation distillation | EDGCN · N-C | accuracy: 0.958 -> 0.963 | Helped |
| Add cross-representation distillation | EDGCN · CIF10 | accuracy: 0.716 -> 0.752 | Helped |
| Add cross-representation distillation | EDGCN · DVS128 | accuracy: 0.985 -> 0.983 | Hurt |
| Add cross-representation distillation | EDGCN · N-Cal object detection | mAP: 0.657 -> 0.711 | Helped |
| Define neighbors using joint representations of all vertex attributes (ablation) | EDGCN · N-Cal, CIF10, and DVS128 | accuracy: N-Cal: 0.793 -> 0.801; CIF10: 0.695 -> 0.716; DVS128: 0.954 -> 0.985 | Helped |
| Use coordinate attributes for attentive aggregation rather than fused joint representations (ablation) | EDGCN · N-Cal, CIF10, and DVS128 | accuracy: N-Cal: 0.783 -> 0.801; CIF10: 0.694 -> 0.716; DVS128: 0.972 -> 0.985 | Mixed |
| Add the coordinate attribute update module (ablation) | EDGCN · N-Cal, CIF10, and DVS128 | accuracy: N-Cal: 0.783 -> 0.801; CIF10: 0.694 -> 0.716; DVS128: 0.972 -> 0.985 | Helped |
| Add inference-level distillation (ablation) | EDGCN · N-Cal | accuracy: 0.801 -> 0.818 | Helped |
| Add inference-level distillation (ablation) | EDGCN · CIF10 | accuracy: 0.716 -> 0.724 | Helped |
| Add feature-level contrastive distillation (ablation) | EDGCN with inference-level distillation · N-Cal | accuracy: 0.818 -> 0.825 | Helped |
| Replace contrastive loss with hard feature constraints (ablation) | EDGCN with CRD · N-Cal and CIF10 | accuracy: N-Cal: 0.835 -> 0.814; CIF10: 0.752 -> 0.721 | Hurt |
Surprising findings
- The EDGCN-with-CRD student (0.985 on DVS128 without CRD is 0.985 vs 0.983 with CRD) does not improve over the no-distillation variant on DVS128, while the teacher scores 0.981.
- Adding a task loss to the student branch (variant B) hurts relative to variant C, which applies task loss only to the teacher, which the authors attribute to limited predictive power of intermediate features.
Limitations
- Still lags frame-based models with ImageNet pretraining (e.g., MVF-Net 0.871 on N-Caltech101 vs 0.835).
- CRD gain on DVS128 is negligible (0.985 without vs 0.983 with), and the reported DVS128 accuracy of the student is already above the teacher.
- Requires a frame-based teacher trained on the same event data, adding training cost.
Experimental setup
Datasets: N-Caltech101 CIFAR10-DVS DVS128 Gesture N-Cal detection
Benchmarks: N-Caltech101 N-Caltech101 detection CIFAR10-DVS DVS128 Gesture
Metrics: classification accuracy mAP (eleven-point) #Params GFLOPs inference time
Baselines: EST M-LSTM MVF-Net EventNet RG-CNNs EvS-S EV-VGCNN AEGNN VMV-GCN EV-Transformer LIAF-Net TA-SNN
Builds on: EV-VGCNN AEGNN FitNets Knowledge distillation (Hinton et al.) SimCLR-style NT-Xent
Closest papers by meaning
- AEGNN: Asynchronous Event-Based Graph Neural Networks (CVPR 2022 · similarity 0.88)
- L2-GCN: Layer-Wise and Learned Efficient Training of Graph Convolutional Networks (CVPR 2020 · similarity 0.87)
- A Voxel Graph CNN for Object Classification With Event Cameras (CVPR 2022 · similarity 0.87)
- Simplifying Graph Convolutional Networks (ICML 2019 · similarity 0.87)
- Early-Bird GCNs: Graph-Network Co-optimization towards More Efficient GCN Training and Inference via Drawing Early-Bird Lottery Tickets (AAAI 2022 · similarity 0.87)
- Deeper Insights Into Graph Convolutional Networks for Semi-Supervised Learning (AAAI 2018 · similarity 0.86)
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.