Near Lossless Transfer Learning for Spiking Neural Networks

Authors: Zhanglu Yan, Jun Zhou, Weng-Fai Wong10577-10584

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We have implemented CQ training in CUDA-accelerated Py Torch version 1.6.0. The experiments were performed on a Intel Xeon E5-2680 server with 256GB DRAM and a Tesla P100 GPU... We tested our methods on different networks structures and datasets, and the results are summarized in Table 2.
Researcher Affiliation Academia Zhanglu Yan , Jun Zhou , Weng-Fai Wong Department of Computer Science, National University of Singapore {zhangluyan, zhoujun, wongwf}@comp.nus.edu.sg
Pseudocode Yes Algorithm 1 One iteration of CQ training. and Algorithm 2 Input encoding algorithm.
Open Source Code Yes The framework was developed in Py Torch and is publicly available.1 (Footnote 1: https://github.com/zhoujuncc1/shenjingcat)
Open Datasets Yes Using a 7 layer VGGand a 21 layer VGG-19, running on the CIFAR-10 dataset, we achieved 94.16% and 93.44% accuracy in the respective equivalent SNNs. MNIST consists of 60,000 28 28 grayscale images of handwritten digits from 0 to 9. The CIFAR-100 has the same structure as CIFAR-10 but with labels for 100 classes (Krizhevsky, Nair, and Hinton 2009).
Dataset Splits No MNIST consists of 60,000 28 28 grayscale images of handwritten digits from 0 to 9. 50,000 for training and 10,000 for testing. They are split to 50,000 training images and 10,000 test images. (No explicit mention of a validation set)
Hardware Specification Yes The experiments were performed on a Intel Xeon E5-2680 server with 256GB DRAM and a Tesla P100 GPU, running 64-bit Linux 4.15.0.
Software Dependencies Yes We have implemented CQ training in CUDA-accelerated Py Torch version 1.6.0.
Experiment Setup Yes We trained the networks using the Adam optimizer with an adaptive learning rate for 100 150 epochs until they converged. The length of the spike trains T were set between 100 to 1000 for Le Net*, VGG-11/*, VGG-13 and VGG-16/19 respectively.