Research map / Architectures and efficiency / Spiking networks

Leveraging Asynchronous Spiking Neural Networks for Ultra Efficient Event-Based Visual Processing

AAAI 2025 · Architectures and efficiency › Spiking networks · approach: Event-camera spiking networks · 3 citations (OpenAlex)

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 ASGCN, an asynchronous spiking graph convolutional network that models event streams as dynamic graphs and processes them event-by-event with SNN inference for ultra-low-energy event-based object recognition.

Problem

Event cameras offer low latency and low power, but most event-based methods convert events into frames, losing sparsity and asynchrony; graph-based methods are dense and costly, and prior SNN methods also convert events to frames.

Approach

Treat each event as a graph node, extract K-hop ego-subgraphs for asynchronous node updates, and convert a trained ANN graph convolution into a spiking network (BN folded, threshold set by max activation, soft reset) so that only new events trigger sparse spike message passing.

Key results

• On N-Cars, ASGCN reaches 0.939 accuracy with 0.036 MJ energy versus 1050 MJ for EST and 98.9 MJ for AsyNet, i.e. over 10,000x lower energy than the compared models.
• Replacing the modified BN with vanilla BN drops N-Cars accuracy from 0.939 to 0.867.
• Accuracy reaches about 90% after roughly 2000 events on N-Cars, and the 6-layer network's firing rates stay below 0.1.

Measured effects reported in the paper

Surprising findings

• Energy savings of over 10,000x are obtained relative to frame-based methods while accuracy stays close (0.939 vs 0.925 to 0.945).
• Vanilla BN substitution causes a substantial accuracy drop, showing that BN must be modified for spike-based inference.

Limitations

• Energy figures are simulated from operation counts and firing rates rather than measured on neuromorphic hardware.
• Accuracy is slightly below the best frame-based methods on some datasets (e.g., 0.939 vs 0.945 for SSC on N-Cars; 0.977 on N-MNIST lower than some baselines).

Experimental setup

Datasets: N-MNIST CIFAR10-DVS N-Caltech101 N-Cars

Benchmarks: N-MNIST CIFAR10-DVS N-Caltech101 N-Cars

Metrics: top-1 accuracy energy (MJ) MOP (operations) storage consumption firing rate

Baselines: HOTS HATS DART YOLE AsyNet EST SSC NVS-S AEGNN HFirst Gabor-SNN TA-SNN

Builds on: Graph-based asynchronous event processing (NVS-S) AEGNN SpiKe-free ANN-SNN conversion (Rueckauer et al. 2017) SplineCNN / B-Spline kernel (Fey et al. 2018) Integrate-and-Fire neurons

Closest papers by meaning

• SpikePoint: An Efficient Point-based Spiking Neural Network for Event Cameras Action Recognition (ICLR 2024 · similarity 0.90)
• A Chaotic Dynamics Framework Inspired by Dorsal Stream for Event Signal Processing (ICML 2025 · similarity 0.90)
• Fully Spiking Neural Networks for Unified Frame-Event Object Tracking (NeurIPS 2025 · similarity 0.89)
• Graph Neural Network Combining Event Stream and Periodic Aggregation for Low-Latency Event-based Vision (CVPR 2025 · similarity 0.89)
• Spiking Neural Network as Adaptive Event Stream Slicer (NeurIPS 2024 · similarity 0.88)
• SFOD: Spiking Fusion Object Detector (CVPR 2024 · similarity 0.88)