I2E framework converts images to events 300x faster for SNNs
New algorithm generates event-stream data from static images 300x faster, enabling high-accuracy SNNs
Spiking neural networks (SNNs) promise ultra-low-power computing but are held back by a lack of event-stream data—the natural input format for SNNs. Researchers from multiple institutions (Ruichen Ma, Liwei Meng, Guanchao Qiao, Ning Ning, Yang Liu, Shaogang Hu) present I2E, a framework that bridges this gap by converting everyday static images into realistic event streams. I2E mimics biological microsaccadic eye movements using a highly parallelized convolution approach, achieving a 300x speedup over previous conversion methods. This enables on-the-fly data augmentation and scalable dataset creation for SNN training.
I2E's effectiveness is demonstrated on large-scale benchmarks. Training an SNN on the synthetic I2E-ImageNet dataset achieves a state-of-the-art accuracy of 60.50%. Crucially, the framework supports a sim-to-real paradigm: pretraining on synthetic I2E data followed by fine-tuning on the real-world CIFAR10-DVS dataset yields an unprecedented 92.5% accuracy. This result proves synthetic event data can serve as a high-fidelity proxy for real sensor data, solving a long-standing challenge in neuromorphic engineering. The I2E algorithm and all generated datasets are open-source, providing a foundational toolkit for high-performance neuromorphic systems.
- I2E converts static images to event streams 300x faster than prior methods using parallelized microsaccade simulation
- SNN trained on I2E-ImageNet achieves 60.50% accuracy, and sim-to-real fine-tuning on CIFAR10-DVS hits 92.5% accuracy
- Open-source algorithm and datasets provide scalable solution to event-data scarcity for neuromorphic computing
Why It Matters
I2E removes the critical data bottleneck for energy-efficient SNNs, enabling practical neuromorphic systems with real-world accuracy.