Hybrid EfficientNet-SNN Hits 99.09% on 64-Class ImageNet Without Backprop
June 19, 2026
Coupling a frozen EfficientNet encoder with a CoLaNET spiking classifier via rate-coded spike trains, this pipeline trains the SNN using local biologically inspired rules only, achieving 99.09% accuracy on a 64-class ImageNet subset. The approach sidesteps end-to-end gradient propagation while matching conventional deep network performance.
HOW THIS AFFECTS YOU
●
researcherDemonstrates a viable path to high-accuracy SNNs by decoupling pretrained ANN encoders from local-rule SNN classifiers, useful for neuromorphic hardware targeting.