Keming Labs

2011 June 1

There was a huge response on “The Twitter” to the Kindle typography article I posted in April. Amazon has an SDK for their lil' e-reader, so I started a company to explore the possibilities. Drop us a line if you want a Kindle app or have other digital typography needs.

Webtrends
851 SW 6th Avenue
Floor 16
Portland OR 97204

Prote.cs: A novel protein fold search algorithm

2010 December 4

Prote.cs is an algorithm that assigns an uncharacterized protein structure to a fold family by answering the question

“Which fold family has members that can best sum to approximate this protein?”

More specifically, proteins are mapped into a vector space by their alpha carbon distance matrices and then assigned to a fold family according to an $l^1$-norm minimized linear regression on a basis derived from protein structures with known fold.

An accuracy of 95% was achieved on a set of 466 CATH fold families using just six-by-six pixel distance matrix thumbnails (21 dimensions).

Details and examples here, code on github

YALL1

2010 November 14

I’ve rewritten YALL1 (“Your ALgorithms for L1”) for Octave compatibility. YALL1 is a Matlab program by Yin Zhang, Junfeng Yang, and Wotao Yin (see original) that solves several $l^1$ minimization problems that have become notable in connection with compressed sensing. These problems are variations on the basis pursuit theme:

\[ \min \norm{\vec x}_1 \quad \mbox{such that} \quad \mat A \vec x = \vec b, \]

where the measurement matrix $\mat A \in \real^{m \times n}$ has $m \lt\lt n$, and the solution $\vec x$ is known to be (approximately) sparse. As it turns out, sparse reconstructions have a wide variety of applications in image compression, signal acquisition, and even classification/machine learning problems.

Compressed sensing details and examples here, code on github

Rliblinear

2010 November 5

I’ve forked an R interface to the LIBLINEAR C++ library, which solves classification and regression problems having millions of instances and features. These problems are variations on the theme:

\[ \min_{\mathbf w} \quad \tfrac{1}{2}\mathbf w^{\mathrm{T}} \mathbf w + C \sum_{i=1}^{l} \max\left(0, 1 - y_{i} \mathbf{w}^{\mathrm T}\mathbf x_{i} \right)^{2}\kern-6pt, \]

with \( \mathbf x_i \) a sample vector, \( y_i \in \{0, 1\} \) its class, and \( \mathbf w \) the vector that defines a decision hyperplane (problems with \(k\) classes are handled automatically by finding \(k\) one-vs-all hyperplanes).

Essentially, this is a very fast, large-scale support vector machine library that has no desire to play in fancy-pants kernel spaces. The R interface automatically expands categorical features (i.e. factors) with $c$ levels into $c$ binary dimensions.

Details and examples here, code on github