A team from Inspur ranked in the Top 3 at the recent Auto Deep Learning Finals. Hosted by the NeurIPS, the AutoDL challenge is the world’s top competition in the field of automatic machine learning with participants from cutting-edge scientific research institutions including Carnegie Mellon University, Seoul National University, University of Freiburg, and Leibniz University Hannover, as well as leading AI technology companies, such as Google, Microsoft, Alibaba, and Inspur.
The decider of this challenge, the AutoDL 2019-2020 finals had a higher difficulty level than the previous challenges. This competition was not limited to any single technical field. Instead, it required contestants to design and develop an automatic deep learning system without any human intervention, and use the system to solve automatic multi-label classification problems for various field including images, videos, text, speech, and tabular data. The challenge was divided into a feedback stage and a finals stage. In the feedback stage, the contestants developed an AutoDL program offline based on 24 training data sets to realize training data processing, model structure design, parameter adjustment and other processes. In the finals stage, the contestants’ AutoDL program was required to automatically complete the model training of 10 private data sets without any human intervention. The winner of the finals was decided by the average ranking of the multiple rounds of evaluation in the finals stage. Inspur performed well in this competition, coming first place once, second place once, and third place three times in the 10 event items.
The full-process AutoDL solution developed by Inspur covered core components such as automatic data processing, automatic model construction and automatic hyper-parameter optimization. The AutoDL solution has three major characteristics: intelligent adaptation, which made it possible for the solution to constantly adapt to the various data and task generation models and increased the level of certainty amid uncertainty; high efficiency, which used incremental data to build high-performance data input pipelines to reduce the IO bottleneck of data transmission; and self-balancing, which made it possible to achieve a stable balance between time and performance in the model search and generation process. By leveraging the above-mentioned characteristics, the AutoDL solution developed by Inspur achieved an average accuracy improvement of nearly 20% over the baseline, average data reading efficiency improvement of 22%, and could realize the model establishment, search and generation processes within half an hour.
Inspur’s leading core technology used in this competition has been applied to Inspur AutoML Suite, an automatic machine learning AI algorithm platform product. AutoML Suite realizes a one-stop automatic generation model based on GPU cluster visualization operations. It has three major automation engines: modeling AutoNAS, hyper-parameter adjustment AutoTune, and model compression AutoPrune, to provide powerful support for computing power.