Minerva: A Scalable and Highly Efficient Training Platform for Deep Learning

By Minjie Wang, Tianjun Xiao, Jianpeng Li, Jiaxing Zhang, Chuntao Hong, Zheng Zhang. 2014

28 citations recorded by the source corpus.

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In ML Systems & Hardware

Citations in the corpus

ImageNet Large Scale Visual Recognition Challenge (2014, neutral): “7 Peking University†, Microsoft Research‡, Shanghai Jiao Tong University∓, XiDian University§, Harbin Institute of Technologyς Tianjun Xiao†‡, Minjie Wang∓‡, Jianpeng Li§‡, Yalong Baiς‡, Jiaxing Zhang‡, Kuiyuan Yang‡, Chuntao Hong‡, Zheng Zhang‡ (Wang et al., 2014)”
MXNet: A Flexible and Efficient Machine Learning Library for Heterogeneous Distributed Systems (2015, background): “Different to most dataflow engines [14], our engine tracks mutation operations as an existing resource unit.”
Supporting Very Large Models using Automatic Dataflow Graph Partitioning (2018, background): “Systems like Minerva[45] and Strads[46] partition the parameter tensors to multiple GPUs.”
Making Contextual Decisions with Low Technical Debt (2016, methodology): “There are many systems designed for supervised machine learning such as CNTK [14], GraphLab [25], Parameter Server [35], ML-lib [45], TensorFlow [48, 49], Torch [51], Minerva [53] and Vowpal Wabbit [52] to name a few.”
Scalable deep learning on distributed GPUs with a GPU-specialized parameter server (2015, background): “Data-parallel approaches are limited by the desire to ﬁt the entire model in each worker’s memory, and, as observed by prior work [7, 13, 9, 28, 19, 25], this would seem to imply that GPU-based systems (with their limited GPU memory) are suited only for relatively small neural networks.”
GeePS: scalable deep learning on distributed GPUs with a GPU-specialized parameter server (2016, background): “Data-parallel approaches are limited by the desire to fit the entire model in each worker’s memory, and, as observed by prior work [7, 9, 14, 22, 30, 33], this would seem to imply that GPU-based systems (with their limited GPU memory) are suited only for relatively small neural networks.”
On parallelizability of stochastic gradient descent for speech DNNS (2014, methodology): “E.g., for our model size, ηmbs(256) ≈ 1.4 on a K20X GPU.”