Bad Graphs!

One of the good things of the peer review process is that if you publish, you’re eventually going to have to review papers for conferences or journal in your (perceived) area of expertise. Sometimes you get pearls such as “the resulting results of algorithm X are resulted” (true story), or “the dynamics of the attorney of yes no plasmodium” (also true), but sometimes bad science comes from the bad presentation of results.

This is also a (essentially true) story. So I’m reviewing a paper that proposes some kind of method for predicting the value of (some) parameter that minimizes some error function. The method is fast, but not analytic. The graph in the paper looks something like:

Linear Interpolation (Interpolation, part I)

In a couple of different occasions I discussed the topic of interpolation, without really going into the details. Lately, I had to interpolate data and that got me interested (again) in interpolation; and I think I should share some of the things I learned.

In this first post (of a series), let us begin by the simplest interpolation of all (after constant interpolation): linear interpolation.

New Way of Computing Square Roots?

I sometimes have quite nerdy readings. As of late, I’ve been reading Le fabuleux destin de (The Fabulous Destiny of , one might translate) by Benoît Rittaud. This book is a treasure trove of facts, and one caught my eye more than the others so far: an iterative method to compute square roots of any (positive) number.

When the method is first presented, he leaves to the reader to find a demonstration (though he gives one much later on, several chapters later), but let’s see what we can find.

Introducing Theano & PyLearn

Today I am going to talk about my day job a bit. Contrary to previous jobs, a good part (but not all) of what I do now is either public domain or open-source. Two the projects I joined recently are Theano and PyLearn.

Theano is a mathematical expression compiler that maps expressions described in Python to machine-efficient code, either targeting the CPU or the GPU. PyLearn is a work in progress that aims to provide a comprehensive machine-learning framework for Theano.

Surrogate Functions

In some optimizations problems, the objective-function is just too complicated to evaluate directly at every iteration, and in this case, we use surrogate functions, functions that mimic most of the properties of the true objective-function, but that is much simpler analytically and/or computationally.

Lately, I have done some thinking about what properties a surrogate function (or surrogate model) should have to be practical.

Log Books

The other day, I was discussing with a friend about my log books, and it seems that, while it’s fairly common with scientists of all sorts, it’s not a generalized practice amongst computer scientists and programmers. But it should: the log book is not only for chemists.

First, the log book serves as… a log. A written trace of your activity during the day. While this sounds silly, it may be useful in retrospect when it is needed to assess time spent on a particular (class of) task(s), to get a good idea of were you are spending your time at work.

Epigraphs in LaTeX

There are times when part of the message, the gist, must be communicated to the reader in an out-of-band fashion, so to speak. One way of doing this is to use an epigraph to open a chapter or section, carefully chosen to convey the intended message but in the voice of another author (self-epigraphs are of very bad taste in my opinion).

is the preferred document preparation system of computer scientists, physicists, and mathematicians and if you intend to follow a career into the academia, it’s pretty much unavoidable. One day, you’ll have to learn . The thing is, is pretty much like C++: it can do just about anything, but it’s not going to help you do it. You have to rely on the innumerable packages or, if you really can’t find what you need, you can code it yourself. Let us have a look on how to code an epigraph macro in .

Lost+Found: Lego Antikythera Mechanism

POCS and Color Spaces

If you’re doing image processing, you’ve probably had to transform your image from one color space to another. In video coding, for example, the RGB image is transformed into YPrPb or YCrCb so that most of the visually relevant information is packed into the Y component which is essentially brightness. Subsampling the chroma bands (Cr and Cb) provides additional means for compression with little perceptual quality loss. While the human eye is very good at detecting brightness variation, it’s not very good at detecting subtle changes in chroma, either saturation or hue.

The thing is that very often, there are color space transformation matrices found in text book but they’re not, due to rounding (and other possible errors), always exactly inverses of each other. This week, I will discuss how we can use projections onto convex sets (POCS) to make sure that reduced precision matrices are exactly (within a given precision) reversible.

Rediscovering The Past

Progress, as conceived by most people, consists in replacing older objects, techniques, or philosophies by newer, better, ones. Sometimes indeed the change is for the better, but sometimes it is just change for change—ever had an older device of some sort that was perfectly adequate for your usage, yet you still replaced it with a newer version with no net gain? Unfortunately, the same happens with ideas, especially with mathematics and computer science.

But there are lessons to be learnt from the past. I’m not talking about fables and cautionary tales; I’m talking about the huge body of science that was left behind, forgotten, superseded by modern techniques.