Hybrid ANN-SNN Pipeline Achieves 99.09% Accuracy Using Only Local Plasticity—Why Researchers Are Paying Attention

In a new paper on arXiv, Denis Larionov and colleagues introduce a hybrid ANN-SNN architecture that marries the rich feature representations of pretrained artificial neural networks (ANNs) with the energy-efficient, event-driven computation of spiking neural networks (SNNs). The pipeline uses a pretrained EfficientNet as an encoder, converting its continuous activations into spike trains via rate-coding, then feeds them into a CoLaNET spiking classifier trained with local plasticity rules. This approach completely bypasses end-to-end gradient propagation, relying instead on biologically plausible, local learning signals.

The resulting model achieves 99.09% accuracy on a 64-class ImageNet benchmark—performance on par with full deep networks but with the potential for far lower power consumption when deployed on neuromorphic hardware. The work opens a practical pathway for adapting powerful pretrained vision encoders to spiking networks without the need for specialized training pipelines. Combined with local plasticity, this hybrid method could accelerate the development of efficient, brain-inspired AI systems for edge and embedded applications.