7 Highlights from CODS-COMADS 2017
The co-located CODS and COMAD 2017 conferences by ACM, were held at IIT-Madras from 9–11 March 2017. Three members of Next Tech Lab (Arjun Bahuguna, Sourav Sharan, Sree Harsha Nelaturu) were awarded travel grants (of upto ₹4,000 each) by the conference chair to attend the conference.
Every morning at 9:00AM, the conference began with the invited talk. A round of paper presentations followed. First, usually those previously presented at conferences like NIPS and KDD, and then papers that were submitted to CODS-COMADS directly. Coffee break at 11:00AM was followed by another session of paper presentations till 12:30PM, after which we break-off for lunch till 2:00PM. Post-lunch we had tutorials or talks, which had some of the best moments of the day. We broke off at around 4:30–5:30PM each day for poster-presentations over coffee and continued till 6:00PM after which everyone left for home, still buzzing with the ideas from the day. The detailed scheduling for each day, along with slides to some keynotes, can be found here.
All through the four days, we had enriching discussions with some of the brightest minds in data science and enjoyed ingenious papers from different domains of data science, from learning opinion dynamics (a paper published in NIPS 2016) to estimating accuracy of knowledge graphs. In the barrage of brilliant talks, papers, and ideas presented at the conference some moments definitely stood out as highlights of the conference, and were especially memorable. Some of the highlights of the conference were:
- Lise Getoor’s keynote[slides from the talk]: Lise Getoor is a professor of Computer Science at University of California, Santa Cruz. Her brilliant talk was a true reaffirmation that complex ideas needn’t always be complex. Lise’s talk was a well-structured, well-presented graphical-model-analysis 101 session. Divided into introduction, tools, applications, she provided a broad introduction to graph-model based analysis, followed by an introduction to Probabilistic Soft Logic (PSL), a framework developed by her PhD students and her for data analysis using probabilistic graphical models, that scales better than existing frameworks for large networks. She ended her talk by showing case-studies of how PSL had been applied to problems as varied as drug-discovery, to social science problems like debate-stance classification of online comments. We were fortunate enough to get to talk to her later during lunch, and she had excellent ideas for applying PSL to their current projects at the lab which deal with computational biology and anamoly detection.
- IBM India Research Director’s industry keynote: The wise decision Sriram Raghavan made with his keynote talk was to focus on questions rather than answers. A barrage of interesting ideas for research topics, his talk was a reminder of the countless interesting problems industry has to offer to academic research. His talk on “Trust, Security, and Compliance in a Cognitive Era”, was divided into three key project-proposals for researchers. The first one was a complete NLP solution for tracking and reasoning about legalese, especially in a QnA environment where a user could query if what they’re planning is permitted given the current regulations and what’s the best way forward; a legal expert system of sorts. The second was addressing the problem of data-security and attacks on machine learning models to extract data. He listed some methods of attack like data-theft and data-poisoning which could adversely affect the trust people place in our models. The third was the idea of utilizing ‘trust-free’ technologies like blockchain for data storage and analysis, a direction already being pursued by companies like gem.co. This would be especially instrumental in doing analysis on sensitive data like healthcare, and hopefully convince hospitals and patients to make their data public without loss of integrity.
- Director of AI Research at Apple’s talk: Ruslan Salakhutdinov, Director of AI research at Apple, Associate professor at CMU, and Sloan fellow, is a name right next to Geoffery Hinton in the history of deep learning who happens to be his doctoral adviser. Having co-authored some of the most legendary papers of the field with his doctoral adviser Hinton, he is responsible for developing Bayesian Program Learning, and publishing fundamental results on popular models like Deep Belief Networks and Autoencoders. He also famously co-authored the paper detailing the essential training hack dropout. He truly is a mammoth of the field. And what could be better than learning about his research from Russ himself. On the last day of the workshop, Russ gave a talk on recent advances in deep learning that started from his early work with Hinton and ended at the cutting edge of current research, with new research project ideas bursting from the slides. It was one of the high points of conference, and was really educative.
- Visualization Seminar: An entertaining seminar on visualization was conducted by Kathirmani Sukumar from Gramener, fueled by visualizations available on their website. It was a great reminder on how essential visualizations are to good data science, and how sometimes explaining visualizations can sometimes the most fun part of the analysis. Kudos to Kathirmani for an entertaining and informative talk.
- U Rajesh in concert and Banquet at Westin: After an exhausting day of brilliant talks and ideas, the greatest rejuvenation was provided by an hour long session of carnatic music performed by the brilliant U Rajesh on electric mandolin, with other greats from carnatic music. U Rajesh, brother of the great U Srinivas, has played alongside legendary jazz-fusionist John McLaughlin on the album Floating Album. We were lucky enough to secure autographs of the maestro. The banquet that followed was lavish and plenty, and fuelled some really interesting conversations with eminent personalities on the data science scene.
- This paper: There were tens of interesting papers and results presented at the conference, but the one that really caught our attention was the above linked paper. Transfer Learning is always an interesting topic, since it hints at the golden goal of a more general artificial intelligence. This fascinating papers presented at the conference did two interesting things. They learnt common representations of multimodal data (they call it views, we call it media for the lack of a better word. It refers to captions, videos, audios, other types of ‘media’ data) and they learnt common representations of multilingual data. The simple idea was to learn one vector that carried the information of video, audio, text and image data in one representation. Then use a ‘pivot’ view to learn between two languages/media, and to reconstruct the original information in different vectors. Doing this allowed them to ‘transfer learn’ across media and languages effectively, and their results can be seen here. Apart from the simple yet elegant idea presented, we would also like to compliment the clarity of the presentation given by IBM Research’s Mitesh Khapra. It took the paper a long way in making it our favourite from the conference.
- The BoFs proposal: PK from IIIT-H, along with other professors from IIT-Madras, had an interesting proposal. They wanted to create an online community for researchers working with user generated content, especially from social networks, where they could come together to collaborate, share resources and ideas. It’s an idea in its infancy, but an exciting one nonetheless. Those interested in joining such a community, though still unformed, would have to email pk@iiitd.ac.in with the subject #UGCcommunity. If you are interested and familiar with UGC research, or have some ideas to share, I’m sure he would like to hear from you.
All three of us loved the experience, and look forward to attending the conference again next year, hopefully this time as presenters and not members of the audience. The tentative dates this time are early January, and the tentative location is Goa. We’d like to thank the chair again for awarding us with travel grants, and the entire organizing committee for another brilliant installment of CODS & COMADS. We look forward to seeing you all there next year.
Article by Arjun Bahuguna, Member and Researcher in Minsky Lab at Next Tech.
