Cybenko/Santos Groups Meetings

George Cybenko is the lead of Human Terrain research at Dartmouth, and Eugene Santos works on modeling human intent from Bayesian Knowledge Bases for more than 20 years. Although our groups share a grad student lab and George and Eugene have offices in nearby, we don't collaborate much officially (while students talk and take classes together). Today we had a meeting of both complete groups and went through overviews of each other's activities. Gene's approach is to create a lot of small BKB submodels, called fragments, and have experts estimate the parameters (conditional probabilities) from the context information. An example, work by JT, was modeling South Caroline primaries, with the intercept of the declining Clinton's popularity graph by the rising one of Obama. The fragments then are fused together into that high-level model, providing overall insights as the intercept. The expert is needed in the loop to estimate the relative links, but several experts can be used and their reliability, and indeed the variability of each parameter, can be evaluated w.r.t. changing the outcome.

David Fleet at Dartmouth CS

David Fleet is a professor at the University of Toronto, studying human gait as recognized in video. His group builds dynamic models of humans walking, of their legs and joints moving, even across the occlusions. They also can identify people from their gaits, but not the groups such as racial/age, etc.

AAAI '09, Dieter Fox

The second day keynote was by Dieter Fox of UW. I have fond memories of a Dieter's lecture being the first thing which had come up on a local cable TV channel when I moved to Seattle and stayed at Amazon's corporate housing. I though, wow, this is really a city of high-tech! Dieter was talking about SLAM, and it was captivating -- little robots swarming an apartment and discovering where everything is, together. He showed more cool stuff in his keynote, some of which I saw at his Dartmouth CS talk, so now could learn more details and hear more sense. I think the most important algorithmic contribution is the hierarchical CRF model, with Lin Liao. It naturally progresses bottom up, from basic behavior such as mode of travel -- walking, driving, riding a bus -- to places, to goals, and predictive advice ("knock-knock"). The last part had a visualization of a USSR-like mob of people and a little robot trying to learn his way through it, cutting across at the last possible moment and zigzagging to safety. The focus here is for robots to learn to "flow" with people, instead of treating them as mere obstacles. That in turn requires learning about what people are doing, and when their flow changes. (Reminds me of the "detect change" item of the Thayer agenda.)

AAAI '09 Spring Symposia Workshop on HBM

The first AAAI Spring Symposia workshop on HBM was held at Stanford from March 23-25, 2009. The organizers were Professors Tanzeem Choudhury of Dartmouth, Henry Kautz of Rochester, and Ashish Kpoor of Microsoft Research. They did a great job in putting together a full program and top keynote speakers! I was there and took extensive notes; a few of those will be here, and the rest will percolate through the blog slowly as relevant!

The program of the workshop

I review the first keynote here, and will blog more on the workshop in the subsequent posts.

The first keynote was given by Eric Horvitz, the Microsoft Research Area Manager in Adaptive Systems, and the President of the AAAI. I was familiar with his work from before, notably the Predestination paper -- but now learned much more about it, its context, and the future of where it's going -- the Mobile Commodities. Imagine you drive somewhere and need gas, snacks, SD cards... And there're merchants willing to provide their products and services at low prices. If there's a platform where they can bid for your gas money and the like, you could see a prompt to stop by a certain gas station for less. Furthermore, there can be a say $5 incentive to exit from the highway and go a bit further, to diverge or even go back a bit, for a specific promotion.

I can further envision this service extending to inviting other people to share the road with. Say you're visiting Portland, and want biking buddies. The platform could match you with some like-minded folks, or maybe just cute ones.

I also liked the robotic concierge at MSR, which can identify non-MSers as those dressed in suits. As someone who visited MS campus extensively, I can affirm that's true and this is how real human receptionists focus on you and attune their tone of voice and welcome!

I had a chance to have lunch with Eric, and he had very interesting stories about the early Internet times in the USSR -- it turns out he was one of the originators of GlasNet, one of the earliest Internet providers in then-USSR! So it thanks to him and other GlasNet folks that one of my emails found its way to Dartmouth. Next time in Moscow I hope to trace more of the Russian Internet history -- especially given that my parents worked at the Kurchatov Institute, the founder of Relcom; and I had early experience with Demos, as well as attending the first GNU conference in Moscow (with a paper) and meeting Richard Stallman there, who brought a backpack full of food and magnetic tapes with GNU software to hide in case US submerges.

More on the workshop in the next post!

Human Behavior Modeling Blog Begins!

Welcome to the HBMB, the Human Behavior Modeling Blog. I am Alexy Khrabrov, a computer scientist currently at Thayer School of Engineering, Dartmouth College. You can read more about me at my hub website.

Human Behavior Modeling (HBM) is a new and exciting area of Computer Science. The main factor enabling it is the emergence of mobile devices and sensor networks. Now we can gather and analyze massive amounts of data about people going about their life, moving their bodies, traveling around, talking to each other, browsing the Internet, making purchasing decisions and choosing to spend their time and attention in a certain way. Several centers of knowledge were tracking these abilities for a while, notably the MIT Media Lab, specifically the group of Sandy Pentland; University of Washington, notably Dieter Fox, Henry Kautz and their students, and now Dartmouth, where Tanzeem Choudhury is advancing the area in Computer Science, and George Cybenko is a pioneer of intelligent agents and process query at Thayer School.

The goal of this blog is to track the field and serve as a resource for the researchers and the practitioners. I will strive to survey the new papers and developments as they occur, and report from the relevant conferences I attend.