Dynamic Sparse Graph for Efficient Deep Learning

Authors: Liu Liu, Lei Deng, Xing Hu, Maohua Zhu, Guoqi Li, Yufei Ding, Yuan Xie

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show significant memory saving (1.7-4.5x) and operation reduction (2.3-4.4x) with little accuracy loss on various benchmarks.
Researcher Affiliation Academia 1Department of Electrical and Computer Engineering, University of California, Santa Barbara 2Department of Computer Science, University of California, Santa Barbara 3Center for Brain Inspired Computing Research, Department of Precision Instrument, Tsinghua University
Pseudocode Yes Algorithm 1: DSG training
Open Source Code No The paper does not provide an explicit statement or link for the open-source code for the described methodology.
Open Datasets Yes Regarding the evaluation network models, we use Le Net (Le Cun et al., 1998) and a multi-layered perceptron (MLP) on small-scale FASHION dataset (Xiao et al., 2017), VGG8... on medium-scale CIFAR10 dataset (Krizhevsky & Hinton, 2009), VGG8/WRN8-2 on another medium-scale CIFAR100 dataset (Krizhevsky & Hinton, 2009), and Alex Net... on large-scale Image Net dataset (Deng et al., 2009) as workloads.
Dataset Splits No The paper mentions 'validation set' and 'validation accuracy' but does not explicitly provide the specific dataset split percentages or methodologies for creating these splits.
Hardware Specification Yes The programming framework is Py Torch and the training platform is based on NVIDIA Titan Xp GPU.
Software Dependencies No The paper mentions 'Py Torch' as the programming framework and 'MKL compute library' but does not specify any version numbers for these software dependencies.
Experiment Setup Yes The projection matrices are fixed after a random initialization at the beginning of training. We just update the projected weights in the low-dimensional space every 50 iterations to reduce the projection overhead.