Hierarchical Channel-spatial Encoding for Communication-efficient Collaborative Learning

Authors: Qihua ZHOU, Song Guo, YI LIU, Jie ZHANG, Jiewei Zhang, Tao GUO, Zhenda XU, Xun Liu, Zhihao Qu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that SGQ achieves a higher traffic reduction ratio by up to 15.97 and provides 9.22 image processing speedup over the uniform quantized training, while preserving adequate model accuracy as FP32 does, even using 4-bit quantization.
Researcher Affiliation Academia Qihua Zhou1, Song Guo1 , Yi Liu1, Jie Zhang1, Jiewei Zhang1, Tao Guo1, Zhenda Xu1, Xun Liu1, Zhihao Qu2 1Department of Computing, The Hong Kong Polytechnic University 2School of Computer and Information, Hohai University
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks (clearly labeled algorithm sections or code-like formatted procedures).
Open Source Code No Source codes will be shared at Github after the double-blind review.
Open Datasets Yes Our benchmarks are image classification tasks based on the training of Alex Net [18], VGG-11 [28], Res Net-18/34 [9], Shuffle Net-V2-1.0x/0.5x [21], and Mobile Net-V1 [10], with the CIFAR-10/100 (CF10/100) [17], Fashion MNIST (FM) [36] and mini-Image Net (MI) [33] datasets.
Dataset Splits No The paper states that experimental settings are discussed in 4.1, but section 4.1 only mentions batch sizes and optimizers, not specific train/validation/test dataset splits or their percentages/counts.
Hardware Specification Yes To match the edge environment, we evaluate SGQ on two types of devices: (1) NVIDIA Jetson Nano series [24], and (2) HUAWEI Atlas 200DK [13], both of which are connected to the NVIDIA RTX 2080Ti server through 10Gb E network.
Software Dependencies Yes All of these benchmarks are implemented via Py Torch-1.7.1 [25].
Experiment Setup Yes As to MI, the batch size is 32 with the SGD optimizer. As to CF and FM, the batch size is 100 with the Adam [16] optimizer.