It's been a while since I've last posted, which was mostly due to my lack of energy or willpower to maintain an active blog. But I thought I'd get back into it.
This summer, I attended two conferences, ICML and AAAI. While the two conferences overlap significantly in content (in the sense that the field of machine learning is more or less a sub-field of artificial intelligence), they differ dramatically in style and points of emphasis.
First, let's talk about the venues. While undistinguished or "boring" venues rarely hurt the experience, exotic or scenic venus can be a big plus. This year ICML was located in Edinburgh, which was definitely a huge positive. I greatly enjoyed walking around the iconic city center, as well as exploring the beautiful countryside.
AAAI was located in Toronto, which was a decent location, though nowhere near as scenic. One benefit of Toronto is its large Chinese population, and thus its excellent Chinese cuisine.
This was only my first time attending AAAI, so my views on the conference may be quite biased due to this small sample size. However, I have attended several ICMLs so I feel like I have a good grasp on things there.
AAAI covers an incredibly broad range of topics, including fields such as robotics, satisfiability, computational sustainability, game theory, natural language processing, and, of course, machine learning. While ICML often produces papers that touch on these topics, at its core those papers are all focused on machine learning.
In general, I found that the machine learning papers at AAAI were not as strong as their ICML counterparts. I suspect this is primarily due to researchers wanting to send their best machine learning work to conferences such as ICML (or KDD or NIPS), rather than a broad and diffuse conference such as AAAI. I'll discuss more detailed thoughts on ICML in a future post.
On the flip side, the sheer scope and breadth of applications at AAAI were at times inspiring and at times overwhelming. I was particularly interested in the computational sustainability track, in part due to having a paper in that track, and in part due to the numerous interesting applications that fall under the umbrella of computational sustainability. Applications include conservation planning, forest patrolling, environmental monitoring, and many more.
I'm quite fascinated with the concept of computational sustainability as a field. As with any new field, there's a lot of uncertainty in terms of what counts as sustainability and what doesn't. In some sense, almost all engineering projects count, since anything that makes smarter or more efficient use of existing resources is contributing towards sustainability. On the other hand, having such a broad definition makes the definition itself less meaningful. I suspect that, as time passes, various topics will wax and wane in popularity due to a combination of technical novelty and practical relevance.
With regards to the plenary sessions, AAAI (and the co-located IAAI conference) had hands-down the most wonderful collection of invited speakers I've ever seen (for a computer science research conference). The hits include Judea Pearl's Turing Lecture on causality, Josh Tenenbaum's talk on understanding the human mind, Luis von Ahn's talk on Duolingo, and (my favorite) Sebastian Thrun's talk on self-driving cars.
I had not realized how far the state-of-the-art on autonomous urban driving had advanced since the last DARPA Grand Challenge. In his talk, Sebastian Thrun demonstrated a car that can navigate the streets and highways of California at the speed limit while avoiding dangerous situations and obeying the law. Sebastian commented that he uses the autonomous Prius to drive him to Tahoe on the weekends -- a four hour drive! It is entirely possible that this technology will be consumer-ready within the next ten years.
Sebastian Thrun's talk closed on some thought provoking notes. It turns out that the infrastructure supporting automobiles is quite expensive and wasteful. Most cars are in use only a small fraction of the time, and are parked the vast majority of the time. Any public or commercial building must dedicate a significant amount of property (perhaps the majority of the entire compound) to parking lots and garages. Two reasons why car sharing hasn't taken off en masse are the logistical complexity of finding/reserving a shared car, and the inconvenience of traveling to a centralized car repository. Both issues can be resolved via self-driving cars. In such a scenario, cars can be organized via a centralized scheduler, which can then tell your reserved car to show up at your doorsteps. That perhaps is not exactly what the future will look like with regards to personalized modes of transportation, but freeing up all that inefficiency seems like an important problem to address nonetheless.