TY - JOUR

T1 - Winning Solutions and Post-Challenge Analyses of the ChaLearn AutoDL Challenge 2019

AU - Liu, Zhengying

AU - Pavao, Adrien

AU - Xu, Zhen

AU - Escalera, Sergio

AU - Ferreira, Fabio

AU - Guyon, Isabelle

AU - Hong, Sirui

AU - Hutter, Frank

AU - Ji, Rongrong

AU - Junior, Julio C. S. Jacques

AU - Li, Ge

AU - Lindauer, Marius

AU - Luo, Zhipeng

AU - Madadi, Meysam

AU - Nierhoff, Thomas

AU - Niu, Kangning

AU - Pan, Chunguang

AU - Stoll, Danny

AU - Treguer, Sebastien

AU - Wang, Jin

AU - Wang, Peng

AU - Wu, Chenglin

AU - Xiong, Youcheng

AU - Zela, Arber

AU - Zhang, Yang

N1 - Funding Information: The authors would like to thank NVIDIA Corporation for the donation of the GPU used for this research. This work was supported by in part by the Google Research (Zu€rich), in part by the 4Paradigm, in part by the Amazon, in part by the Microsoft, in part by the ICREA through the ICREA Academia Programme. The work of team automl_freiburg was supported in part by the European Research Council (ERC) through the European Union’s Horizon 2020 Research and Innovation Programme under Grant 716721, in part by the Robert Bosch GmbH, in part by the institutions of the co-authors, and in part by the Spanish project PID2019-105093GB-I00. The authors would also like to thank Olivier Bousquet and André Elisseeff at Google for their help with the design of the challenge and the countless hours that André spent engineering the data format. The special version of the CodaLab platform we used was implemented by Tyler Thomas, with the help of Eric Carmichael, CK Collab, LLC, USA. Many people contributed time to help formatting datasets, prepare baseline results, and facilitate the logistics. The authors would also like to thank Stephane Ayache (AMU, France), Hubert Jacob Banville (INRIA, France), Mahsa Behzadi (Google, Switzerland), Kristin Bennett (RPI, New York, USA), Hugo Jair Escalante (IANOE, Mexico and ChaLearn, USA), Gavin Cawley (U. East Anglia, UK), Baiyu Chen (UC Berkeley, USA), Albert Clapes i Sintes (U. Barcelona, Spain), Bram van Ginneken (Radboud U. Nijmegen, The Netherlands), Alexandre Gramfort (U. Paris-Saclay; INRIA, France), Yi-Qi Hu (4paradigm, China), Tatiana Merku-lova (Google, Switzerland), Shangeth Rajaa (BITS Pilani, India), Herilalaina Rakotoarison (U. Paris-Saclay, INRIA, France), Lukasz Romaszko (The University of Edinburgh, UK), Mehreen Saeed (FAST Nat. U. Lahore, Pakistan), Marc Schoenauer (U. Paris-Saclay, INRIA, France), Michele Sebag (U. Paris-Saclay; CNRS, France), Danny Silver (Acadia University, Canada), Lisheng Sun (U. Paris-Saclay; UPSud, France), Wei-Wei Tu (4paradigm, China), Fengfu Li (4paradigm, China), Lichuan Xiang (4paradigm, China), Jun Wan (Chinese

PY - 2021/9/1

Y1 - 2021/9/1

N2 - This paper reports the results and post-challenge analyses of ChaLearn's AutoDL challenge series, which helped sorting out a profusion of AutoML solutions for Deep Learning (DL) that had been introduced in a variety of settings, but lacked fair comparisons. All input data modalities (time series, images, videos, text, tabular) were formatted as tensors and all tasks were multi-label classification problems. Code submissions were executed on hidden tasks, with limited time and computational resources, pushing solutions that get results quickly. In this setting, DL methods dominated, though popular Neural Architecture Search (NAS) was impractical. Solutions relied on fine-tuned pre-trained networks, with architectures matching data modality. Post-challenge tests did not reveal improvements beyond the imposed time limit. While no component is particularly original or novel, a high level modular organization emerged featuring a 'meta-learner', 'data ingestor', 'model selector', 'model/learner', and 'evaluator'. This modularity enabled ablation studies, which revealed the importance of (off-platform) meta-learning, ensembling, and efficient data management. Experiments on heterogeneous module combinations further confirm the (local) optimality of the winning solutions. Our challenge legacy includes an ever-lasting benchmark (http://autodl.chalearn.org), the open-sourced code of the winners, and a free 'AutoDL self-service'.

AB - This paper reports the results and post-challenge analyses of ChaLearn's AutoDL challenge series, which helped sorting out a profusion of AutoML solutions for Deep Learning (DL) that had been introduced in a variety of settings, but lacked fair comparisons. All input data modalities (time series, images, videos, text, tabular) were formatted as tensors and all tasks were multi-label classification problems. Code submissions were executed on hidden tasks, with limited time and computational resources, pushing solutions that get results quickly. In this setting, DL methods dominated, though popular Neural Architecture Search (NAS) was impractical. Solutions relied on fine-tuned pre-trained networks, with architectures matching data modality. Post-challenge tests did not reveal improvements beyond the imposed time limit. While no component is particularly original or novel, a high level modular organization emerged featuring a 'meta-learner', 'data ingestor', 'model selector', 'model/learner', and 'evaluator'. This modularity enabled ablation studies, which revealed the importance of (off-platform) meta-learning, ensembling, and efficient data management. Experiments on heterogeneous module combinations further confirm the (local) optimality of the winning solutions. Our challenge legacy includes an ever-lasting benchmark (http://autodl.chalearn.org), the open-sourced code of the winners, and a free 'AutoDL self-service'.

KW - AutoML

KW - deep learning

KW - hyperparameter optimization

KW - meta-learning

KW - model selection

KW - neural architecture search

UR - http://www.scopus.com/inward/record.url?scp=85104569330&partnerID=8YFLogxK

U2 - 10.48550/arXiv.2201.03801

DO - 10.48550/arXiv.2201.03801

M3 - Article

C2 - 33891549

AN - SCOPUS:85104569330

VL - 43

SP - 3108

EP - 3125

JO - IEEE Transactions on Pattern Analysis and Machine Intelligence

JF - IEEE Transactions on Pattern Analysis and Machine Intelligence

SN - 0162-8828

IS - 9

M1 - 9415128

ER -