1/28/15

1. Today was another fairly normal day! Highlights: volunteering at Rick's! I forgot that I had a shift until nearly an hour into the shift, but it ended up fine--the crowds didn't show up until I did. Learned a little bit of the backstories of my fellow volunteers Samuel, Andrea and Vivian, plus Catherine from the floor. 
2. After that, watched the first half of an episode of Newsroom over ice cream. This is the show my floormates have been watching, that I need to catch up on--it's really very good!
3. Rest of the day: math, hardware, neural, oh, a Lavin MentorConnect lunch with this venture capitalist. He manages the venture division of Paul Allen's private investment firm Vulcan Capital. Another dude with a crazy backstory--he tried entrepreneurship in Russia after the collapse, then returned to law school, tried to work up an Internet startup on the side, then happened to pick up a WSJ with a classified hiring a new VC associate. Met the VC's, became the Internet partner, pulled off a lucky first deal, and never looked back!
4. And come to think of it, I have an academic tidbit to share from neural. Today we talked about extracting "features" from arbitrary data. One method is known as Principal Component analysis--it takes N-dimensional data and projects it onto a smaller K-dimensional subspace that lies in the direction of maximum variance of the data. Applied to images of faces, it works like this--take the 10k vectors of pixel values for each of 10, 000 images, put them as the columns of a 10k X 10k square matrix. Now take the eigenvectors of this matrix. (This is an expensive operation.) Now take the K of these with the largest eigenvalues. These vectors are arrays of pixel values--they look kind of like faces, each with their own exaggerated features.  I'm not sure quite why these eigenvectors represent dimensions in "face space" that capture dimensions of large variance in the data--I need to study more linear! But in any case--this is great--these eigenvectors are called "eigenfaces", and you can use linear combinations of them to approximate any face! It's pretty great. 
5. Also, played badminton / basketball / soccer with Xin and finally got to eat dinner with him. It's been a while. 
6. I think that's it to report for the day. 'Night!