MACHINE LEARNING

0 posts, 1 voice

PART 2: AN INTERVIEW WITH UDI SCHLESSINGER

• 

  Here's the long-awaited part two of my interview with Dr. Udi Schlessinger -- this one is more focused on his professional career. Again, very interesting stuff!

  Ian: How has machine learning inspired your approach to your professional career?

  Udi: Although what I found was very exciting and great fun to do as well, it was leading me in a direction I didn't want to spend the rest of my life doing. In addition, I had to move to New York because of my wife's boss moving here (poor me ;) ), and then started hunting for jobs that require the technologies I mastered. I wasn't looking to create biologically-based agents anymore: I wanted to use this to solve real world problems.

  Since I was in New York, real world problems means financial problems. And both evolutionary algorithms and machine learning are often used in this domain as well. However, the recession just started when I finished my studies (October 2007), and hedge funds/investment banks largely keep their algorithms secret. So I did not even know where to apply. I ended up working in a big investment bank as a developer, with the goal that one day I'll move to a team that utilizes these algorithms (when I was on the inside, I found several such teams in my first day on the job). However, I wasn't happy working in an IB, the recession got worse, so I quit. Ironically, had my wife's boss moved, say, to San Francisco, I might be working now in the computer game industry since my research is very strongly related to artificial intelligence in games (my project was very-game like - I recorded hundreds of videos).

  I'm still passionate about utilizing machine learning and evolutionary algorithms, and intend to create tools that incorporate those for either trading or affiliate marketing (the readers are probably not familiar with the term: think of it as trading & advertising combined). I've also started working on a game with a science fiction author friend, but had to stop since this requires way too much effort (and is not funded by anyone). I see plenty of opportunities.

  Ian: What current development in machine learning are you most excited about?

  Udi: To be honest, I'm less in touch with ML than I am with EC. One thing I never got to do (even though I really wanted to) is use development-like techniques to create neural networks. The human body is created from around 30,000 genes. These genes do not specify each and every trait, but are more like a recipe describing how to create a human being. The problem with evolutionary algorithms and machine learning is that it is really hard to create super-large neural networks (think millions or billions of nodes), however, using development this might be possible. If biology does it, so can we!

  I really wanted to get involved with this but never got the opportunity. Many people have tried, though so far no one has been able to obtain results better than standard ML or EC results (it's possible I'm not up-to-date though - although I doubt it).

  Ian: Has the machine learning field influenced the way you think about the world? If so, how?

  Udi: My Ph.D. project has strongly influenced the way I see the world. I can literally envision how the world started, how primitive organisms became increasingly complex, how the ecosystems have changed... because I have seen it in my own simulations.

  Here's a quick anecdote: after my simulations ran for a while they always reached a stable steady state, that is, no more changes. Often this would take around 8 hours, but sometimes less and sometimes more (a lot depends on the problem I posed it and the settings I used).

  In one test I was examining the effect of predator/prey on visual systems. You probably don't know, but animals that are prey in real world have eyes aligned to the side (so they would be able to notice potential predators), while predators have eyes pointed to the front (so they could chase their prey - we have our eyes aligned to the front). My supervisors & I were curious whether my agents would get this too should we introduce predator/prey interactions, and whether specific visual systems would evolve for predators and for prey as well as additional neural constructs within the networks.

  In these simulations, there would be two possible scenarios: an environment where no predators evolved. And an environment where slowly reproducing predators evolved that were able to eat enough prey but not too much that they would get their food source extinct. I got many variations of these two themes, but still, overall, two themes (later, btw, I incorporated many additional features that greatly, greatly affected the possible environments, but this is not relevant for this discussion).

  As said, once a steady state reached, it was always stable. I used to run numerous such simulations, one day, almost by accident, I saw that from some reason in a really stable environment, the predators started eating more and more of the prey. I've seen it before, but usually it led for predators dwindling in number, and prey growing back to their previous number. However, in this case it was not happening. I was staring at it in puzzlement: what's going to happen? Surely the predators... aren't going to eat all the prey? Then they would die. This never happened! But in front of my eyes, I saw the predators hunt down the remaining prey, one by one, and then they were left alone. After a while, they all died. And the environment was empty.

  It might sound silly, but I was shocked. I've never seen it (it's the only time this ever happened - at least that I witnessed it). And perhaps it might be over ambitious to draw conclusions from what is basically a simple simulation, it made me think that we human take our world for granted. We've destroyed the ecosystem, wiped out countless species, melted the glaciers (because of the greenhouse effect), and yet we expect it will always be fine because so far - it has. Well, one day, it won't. Just like in my simulation the inconceivable happened, there's a very real chance this could happen to us - but by then it will be too late. This kind of drove the point in that's always discussed in the media because *I've seen it happen*.

  Ian: What's your favorite machine-learning-inspired character/story in film?

  Udi: Hmmm, that's a problematic question, because even though I'm a big fan of AI (and sci-fi) in movies, it's almost always completely and utterly not based on science. Sometimes it's even ridiculously stupid (no offense to the Matrix lovers - I loved that movie too - but the science behind it is completely absurd from too many different reasons to count). I loved Wall-E, and Data, and many many other AI characters.. but the one I found most realistic (that I can think of at the moment) is probably in a single episode of Stargate, where an alien AI lodged itself into the Stargate mainframe, and kept finding ways of surviving despite their best attempts to wipe it out (when they formatted all machines, it copied itself to a mobile machine, etc). This, in my opinion, is true AI.

  Posted by Ian Ma 5 months ago

Reply to this thread

You are replying anonymously to this thread. Please login if you have an account.

Name

Your Reply

Captcha

Send me updates about this topic » « Don't send me updates on this topic

Your E-Mail Address

Saving..