Overview of Intelligent Systems

Yesterday’s Presenting Computer Science lecture was on intelligent systems, given by Petri Myllymäki. Professor Myllymäki would be a great person to interview, many things he said would make good soundbytes. That’s probably why this is the first lecture I’ve gotten around to noting here.

One of the course requirements is to write brief summaries of the lectures and to use the topics as starting points for our own thoughts on the matter. I don’t really care for this “study diary” method, but it’s fairly easy, and probably does help put the different fields of study in perspective. Most of the lectures have been fairly boring, but this isn’t really the fault of the lecturers. It’s a tall order to not only introduce an entire field of study, but also cover how it’s taught at the university, and go into specifics of the latest research�”all in an hour and a half.

Myllymäki is a senior research scientist at HIIT’s Complex Systems Computation Research Group (CoSCo).

Update: Myllymäki’s presentation is available online (pdf).

Computer science asks: what can we automate, and how? So artificial intelligence research asks: how can we automate intelligence? We don’t have to define intelligence. We’ll leave that to philosophers. For us, it’s suffice to say that intelligence is behaving intelligently. So what do we need for intelligent behavior? Learning�”adaptation�”for one thing. And for learning we need memory and the ability to generalize. The world doesn’t repeat itself. This is one reason why rules-based behavior works in a laboratory environment but fails in the real world.

In order to generalize, we need to make models. Models allow us to assess situations, and to predict outcomes. Models are also mathematically computable.

About ten years ago, the term artificial intelligence fell into disrepute. Mainly because it failed to meet the expectations and demands of business.

Historical AI research was focused on cases such as making a robotic arm pick up a green ball and place it on a red block. This worked in labs, but failed in the real world, where perfect geometric shapes are rare. Thus was born the “science of uncertainty”�”along with “intelligent systems,” “soft computing,” “real-world computing,” “complex systems computing,” and “deep computing,” which was named after a certain chess playing machine.

Research projects

Some projects and research that Myllymäki talked about:

  • B-course, a free data analysis server. B-course is based on Bayesian and simplistic causal networks, and features no user-defined variables. This is notable because most neural network tools compete by how many hundreds of variables they allow to be toggled. This, of course, makes them extremely difficult to learn and use. According to Myllymäki, B-course has had over 10 000 users, so CoSCo’s approach has been successful.

  • Next Generation Information Retrieval. Besides building a core search engine to compete with Google, CoSCo wants to use natural language processing and (probabilistic) statistical modeling (PROSE) to refine searching. CoSCo’s search projects include: Search-in-a-Box, topic-specific search, mobile search, and collaborative search. Rather than compete with Google’s server farms, CoSCo plans to build a superpeer-based network of servers from different universities and institutions (Alvis). Myllymäki had interesting things to say about Google, but more about that later.

  • Mobile device location. Rather than being based on traditional basestation triangulation, CoSCo’s system uses antenna signal strengths in low powered, consumer antenna equipped devices, such as mobile phones. Two field tests were conducted in Manhattan in 2001 and 2002, so the system provably works in high urban environments where tall buildings interfere with both GPS and traditional basestation triangulation positioning methods. The system has been commercialized by Ekahau, which uses it to locate devices on WLAN networks. “We’ve got millions of dollars from grannies from Florida, ” Myllymäki joked about Ekahau’s venture capital funding.

  • The MDL Research site. Minimum description language was founded by Jorma Rissanen, a HIIT Fellow and a professor emeritus at the Tampere University of Technology.

Unattributed quote, paraphrased: By 2010, there will be over one yottabyte (1024) of data online (on the Internet and in corporate intranets).

We need both mathematicians and great hackers

Researchers must be very careful when creating intelligent systems. When working with a small sample group or in a closed environment, there is the danger of making the system over-taught or over-learned (ylioppinut, in Finnish). An over-taught system will give you perfect results within its own test data, but will choke when given new [incongruous, as in real-world] data.

Intelligent systems research overlaps algorithm and information systems research. But one way that intelligent systems differs from other fields is that it can’t reformulate its problems. Many times research problems turn out to have complex and ugly solutions that can be made into beautiful and elegant formulas by restating the problem. While I like beautiful mathematical formulas, real-world problems can’t be jiggered.

For example, in the mobile device location problem, the interference of tall buildings could have been eliminated by increasing the power of the device’s antenna. This is a typical engineer’s answer. But mobile phones don’t have powerful enough batteries or antennas. We want to create solutions that work now.

At CoSCo, we need all kinds of people. Besides guys who like to prove new mathematical formulas, we also need great hackers. Most of the people at CoSCo have been there for over ten years, so there’s alot of long-term involvement.

Intelligent systems is fun field. But it’s not only fun and games, we are reliant on outside funding, so we need to achieve results that work.

A whiff of paranoia and a dash of recklessness

Information retrieval is so fundamental to networks that, as nice and useful as Google is, we can’t afford leave it rely solely on an American corporation.

Myllymäki’s slides had quite a striking collection of bullets on Google. I didn’t catch them all, but here’s the top of the list:

  • Google’s immortal cookie (expires in 2038)
  • Google records all it can
  • Google retains all data indefinitely
  • Google ignores privacy policy questions
  • Google hires spooks (NSA)

What exactly Myllymäki meant by these wasn’t all clear, but he did say that these issues are taken very seriously by CoSCo senior research scientist Wray Buntine, a former Google employee. There was also a link to Google Watch.

“We want to do to Google what Linux did to Microsoft.”

What if Google started giving preferencial treatment to American companies over European ones? The EU understood this: as soon as we asked them this question, the money started coming in. (The Alvis superpeer semantic search project is EU funded.)

While the Google stuff sounded almost paranoid, it constrasts sharply with what Myllymäki almost off-handedly said about the ethical considerations of their research.

We’re kind of like the scientists who developed the atom bomb. We don’t think about the ethical ramifications of how our research will be applied. I hope someone does.

---

There is 1 other entry posted on this day.