Trend of AI Research

With a focus on papers at three AI conf.

Nov 6. 2017

Trends in papers presented at three major AI conferences — With a focus on keywords used in titles

Although advances in science and technology development are taking place in various places, there have been many cases in history where in-depth discussions on various topics at academic conferences would later lead to major milestones. For example, the conference which contributed the most to the development of physics in the 20th century is the 1927 Solvay Conference. In that conference, the best physicists of the time such as Albert Einstein, Niels Bohr, Maria Curie, Erwin Schrodinger and Wolfgang Pauli had heated discussions over quantum theories. Among the physicists who attended this historical conference, seventeen of them later received Nobel Prizes. It is expected that there will be many industry-leading conferences in AI research in the future. The NIPS (Neural Information Processing Systems) conference, which could become one of them, was introduced in Issue #1 of Kakao AI Report (https://brunch.co.kr/@kakao-it/51).
In this issue, we study the trends in papers presented at AI and machine learning conferences in 2017, with a focus on keywords used in titles. We hope that this report can be a useful reference to examine what kind of researches can be conducted in the field at present.

The conferences we are introducing in this issue are major AI and machine learning conferences held in 2017. The first one is ICLR (International Conference on Learning Representations) held in April 2017 in Toulon, France. 310 papers were accepted by ICLR 2017 (http://www.iclr.cc/doku.php?id=iclr2017:conference_posters), and top 5 most cited papers are as follows (as of Aug 23, 2017).

Energy-based Generative Adversarial Networks, cited 99 times (J Zhao, Y LeCun. et al.)

Dynamic Coattention Networks For Question Answering, cited 86 times (C Xiong, V Zhong. et al.)

Machine Comprehension Using Match-LSTM and Answer Pointer, cited 83 times (S Wang, J Jiang)

Adversarial Feature Learning, cited 83 times (J Donahue, P Krähenbühl. et al.)

Bidirectional Attention Flow for Machine Comprehension, cited 78 times (M Seo, A Kembhavi. et al.)

Keywords most often used in the titles of papers accepted by ICLR 2017

15 keywords which were most often used in paper titles

The second conference is CVPR (Computer Vision and Pattern Recognition) held in July 2017 in Honolulu, Hawaii. 783 papers were accepted by CVPR 2017 (http://openaccess.thecvf.com/CVPR2017.py), and top 5 most cited papers are as follows (as of Aug 23, 2017).

Densely Connected Convolutional Networks, cited 155 times (G Huang, Z Liu. et al.)

Image-To-Image Translation With Conditional Adversarial Networks, cited 144 times (P Isola, JY Zhu. et al.)

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, cited 121 times (C Ledig, L Theis. et al.)

YOLO9000: Better, Faster, Stronger, cited 48 times (J Redmon, A Farhadi)

Aggregated Residual Transformations for Deep Neural Networks, cited 44 times (S Xie, R Girshick. et al.)

Keywords most often used in the titles of papers accepted by CVPR 2017

15 keywords which were most often used in paper titles

The third conference is ICML (International Conference on Machine Learning) held in Aug 2017 in Sydney, Australia. A total of 434 papers were accepted by ICML 2017 (https://2017.icml.cc/Conferences/2017/Schedule?type=Poster), and top 5 most cited papers are as follows (as of Aug 23, 2017).