After the fun, back to work
June 15, 2012
Posted by on
After the conference, I am staying with friends in Redmond for a few days.
They had to go to work. So I am just working in the living room.
I heard a series of musical notes behind me. It was the Roomba waking up. I had to erect a barrier to keep the robot away. It is persistent.
What I noticed is the combination of randomness with behavioural templates. The Roomba uses an ensemble of strategies to solve problems. This gave me some ideas for compiler optimization.
I’ve decided to rewrite the entire front-end and JIT compiler. Everything will change. It will only take a few days as the amount of code will be much less. I think I had to write it first in an ugly way to understand how to do it in a beautiful way. This may turn out to be important not only for correctness and performance. If the code is something anyone can understand and extend, then this project may go farther.
I won’t start on the rewrite until I get back in San Francisco. Today, I’m going to play with the Simon Funk (Brandyn Webb) stochastic gradient descent version of SVD (singular value decomposition). This works by minimizing the Frobenius norm. I also want to implement a small K-means for a collaborative filter. One of my hosts is a business guy and was talking about a recommender system. So the natural thing to do is try conventional approaches: factor and clustering filters.