The 29th AAAI conference on Artificial Intelligence (AAAI 2015) was a totally new experience for me in many ways: (1) It was a very big and broad conference (2) Since it was broad, there were papers from many sub areas (3) They had many tracks and even keynotes in parallel (4) Met new students and professionals, and (5) had my paper accepted in the technical track :). Following is a brief descriptions on keynotes and papers I came across during AAAI 2015 that may help my colleagues in their research.
Invited talk by Oren Etzioni – Allen Institute for AI
Oren’s keynote was focused around GOFAI (Good Old Fashioned AI) using today’s AI and ML techniques. He mentioned that he and Allen Institute for AI do not focus on inventing or building completely new techniques for AI but utilize available resources in solving problems. He presented two systems: (1) A system that reads and analyzes text to answer questions – ARISTO (2) A system similar to Google Scholar that they refer to as Semantic Scholar. Both the systems look great. ARISTO can answer maths geometry questions by reading the text and also understanding the graphs or sketches which I though is very challenging and exciting. It can successfully answer grade 4 maths geometry questions (also showing good results in grades 9 and 11). The other is the Semantic Scholar which will be available to public this year. It will have more meta data than Google scholar and answer a lot of interesting queries and relationships between the papers.
Invited talk by Rayid Ghani
Rayid’s talk was about data science for social good. I recall that Pramod (one my colleagues) got this scholarship last year. Their effort is to help or make scientists with AI, ML, and data analysis skills to serve the common good. In doing so, they do not pick people who only does social good at the moment but at least showing interest in future. Throughout his speech he mentioned that, being able to pick which method or technique to solve the problem is very important.
Invited talk by Meinolf Sellmann – IBM TJ Watson
Meinolf is from IBM T. J. Watson group. His talk was about some Watson applications and systems using AI techniques. One system they have built is about selecting relevant parameters and algorithms given the data and environment. He mentioned that this is not the very first attempt that anyone who did it but they have improved upon previous efforts. He showed how important it is to automatically select parameters and algorithms when a computer sees completely different data (this in fact is useful for Watson). He also showed a short demo on automatic speech generation application using Watson. His takeaway message for this is that it is not a complex or advanced technology for Watson, but it makes great difference to customers. That is when IBM say that Watson (i.e., computer) can generate a speech for them given a topic using Wikipedia, it makes huge effect in impressing customers. There is a good lesson I learned from his presentation. That is, he used very simple and easy to understand graphics. I recall that his slides had “Disney Cars” pictures in explaining how to select parameters or algorithms in different context. I believe that everybody in the audience understood his points and to-the-point short demos during the presentation kept the audience alive.
Invited talk by Michael Bowling
This talk was about how to build computers to achieve John von Neumann’s dream of playing chess like humans – guessing the other players’ move, deception tactics, and bluffing. This is to behave like real life. The talk composed of how computers advanced in computing many decisions per second over time and their ability to reason. The talk wasn’t as interesting as I thought and I should have attended Lise Getoor’s talk (about statistics and semantics to solve big data problems).
Thoughts about papers and trend
The conference was very big and hence had broader theme for papers. Among the themes that match Kno.e.sis, social science was dominant. I found the trend in this direction now is to predict demographics or emotions for users using social data (including web site traffic). Most of them use machine learning techniques.
AAAI had some interesting new additions this time. They are, lunch with a fellow and job fair. The job fair was nice and it gave students and prospective graduates to talk with some of the leading companies (including Google, Microsoft, IBM, PARC, etc). I had some successful discussions with some of them during the fair.
Got to know some history of AI. Who knew A* search is a by-product of the first robot (Shakey).
AAAI15 had many sessions in parallel and even keynotes.
Many or at least considerable number of papers are about utilizing or improving current AI/ML techniques. There are less number of papers that introduced a completely new research theme.
AI and Web was the track I think had most number of papers.
Since they had many papers, some papers were given two minutes to present and a poster session.
I had good number of people coming to my poster and many seemed to be impressed :). At the end, I felt like loosing voice by explaining to that many people. But it was a very nice experience.