What I Learned Today

I recently attended the annual Conference on Artificial Intelligence, held this year in Chicago, IL. I took away several new ideas and thoughts. Here are some highlights:

• Alyosha Efros gave his invited talk on a variety of cool things you can do with a single image–from inferring 3D geometry to automatically filling in occluded parts to estimating the location where the image was taken, all of his work is exciting even if you’re not a computer vision researcher. (Bonus: how to digitally “shave”)
• Stuart Russell noted that the graduate Machine Learning course at Berkeley has the largest enrollment of any grad course in the entire university; enrollment has reached as high as 140 students. In his invited talk, he issued a call to arms motivating AI researchers to tackle bigger research problems. “Fruit flies can recognize digits!” he noted. He urged the development of agents with probabilistic first-order logic, that track their internal state, and have the ability to abstract both behavior and their lookahead abilities. Ambitious, of course!
• Two presentations discussed methods for training a learning algorithm with zero data. No, really! Both pointed out problems in which the class labels or identities themselves can be represented in feature space, so you don’t really need to further train on real data. This is pretty obvious–effectively, it’s for problems where you already have a class prototype or representative, which is one way to represent a learned model anyway–but I think the papers are still useful in that they cause us to stop and think about what we can leverage a priori, instead of just hunting for more and more training data.
• Explicit Semantic Analysis (ESA) is an interesting text analysis technique that is useful when analyzing short text snippets, like keywords or web search queries. There isn’t much content in a two-word phrase, but ESA can remedy that by searching for the phrase in wikipedia and then expanding the “semantic content” (or representation) of the phrase using the titles of all wikipedia articles that contain it. Pretty clever!

I also co-chaired the Colloquium on AI Education, which was very well attended, energizing, and fun. I may even talk myself into doing some teaching again this year. :)