Researchers have introduced a hybrid pipeline that combines artificial neural networks (ANNs) with spiking neural networks (SNNs) to achieve high performance while leveraging local, biologically inspired learning. The work, detailed in a paper on arXiv, proposes a method that effectively uses pretrained ANN embeddings to enable high-performance SNNs.
Architecture: EfficientNet Encoder and CoLaNET Classifier
The proposed architecture couples a pretrained EfficientNet encoder with a CoLaNET spiking classifier. The encoder's activations are converted into spike trains via rate-coding, a process that represents analog values as firing rates of spikes. This conversion allows the subsequent SNN classifier to process information in a manner closer to biological neural systems.
Training with Local Plasticity
A key innovation is the use of local, biologically inspired learning rules to train the SNN classifier. Unlike conventional deep networks that rely on end-to-end gradient propagation, this approach bypasses that requirement. The local plasticity mechanism enables the classifier to learn from the spike-encoded representations without backpropagating errors through the entire network, potentially offering energy efficiency and biological plausibility.
Performance Results
The hybrid ANN-SNN pipeline achieved 99.09% accuracy on a 64-class ImageNet benchmark. This performance is on par with conventional deep networks, demonstrating that the hybrid approach does not sacrifice accuracy while offering the benefits of spiking neural computation.
| Metric | Value |
|---|---|
| Accuracy | 99.09% |
| Classes | 64 |
| Dataset | ImageNet |
| Encoder | EfficientNet |
| Classifier | CoLaNET |
| Training | Local plasticity |
Implications and Context
The work presents a biologically plausible and efficient framework for adapting powerful pretrained encoders to downstream spiking neural network tasks. By decoupling the encoder from the classifier training, the method offers practical advantages for scenarios where end-to-end training is infeasible or where spike-based processing is desired. The authors note that the approach achieves performance on par with conventional deep networks, indicating that hybrid ANN-SNN systems can be viable alternatives for certain applications.
The paper is available on arXiv under a Creative Commons BY-NC-SA 4.0 license. The research was conducted by Larionov, Denis, Shtanchaev, Khairutin, Kiselev, Mikhail, Korovin, Tugoy, and Ivan. Further details can be accessed at the provided URL.
