HYBRID ANN-SNN PIPELINE WITH LOCAL PLASTICITY

This paper addresses the challenge of training Spiking Neural Networks (SNNs) for complex image classification tasks. While SNNs offer potential for energy efficiency and event-driven processing, they are notoriously difficult to train using standard gradient-based methods. The authors propose a hybrid pipeline that decouples feature extraction from classification: a pretrained EfficientNet-B3 acts as a fixed feature extractor, while a CoLaNET (Columnar Layered Network) spiking classifier handles the final decision-making using local, biologically inspired learning rules.

The architecture converts high-level semantic embeddings from the EfficientNet encoder into spike trains using rate-coding. These spikes are then processed by the CoLaNET, which employs anti-Hebbian and dopamine-modulated plasticity to update synaptic weights. Crucially, the network operates in a single-pass, online learning regime, meaning it adapts to new data without revisiting previously seen examples or requiring multiple training epochs. The authors optimized the network's hyperparameters using a genetic algorithm to navigate the non-smooth optimization landscape inherent to spiking systems.

The hybrid system achieved 99.09% accuracy on a 64-class subset of ImageNet. The authors demonstrate that this performance is competitive with conventional single-layer Artificial Neural Networks (ANNs) of similar size. By using an ensemble of CoLaNET classifiers, the system maintains high accuracy while keeping the decision-making stage computationally efficient and biologically plausible. The study highlights that this asymmetric design—a powerful, frozen ANN encoder paired with a lightweight, plastic spiking decoder—is a viable strategy for deploying SNNs in resource-constrained environments.

This work provides a practical path toward integrating SNNs into modern deep learning workflows. By bypassing the need for end-to-end backpropagation, the proposed pipeline offers a framework for continuous, online learning that is more compatible with neuromorphic hardware. It demonstrates that spiking networks can achieve high-level performance on complex visual tasks without sacrificing the biological plausibility that makes them attractive for edge computing and robotics.