jump to navigation

Vision without categories? November 17, 2010

Posted by Sarah in Uncategorized.
Tags: , ,
add a comment

I’d just like to mention that I’ve come across Tomasz Malisiewicz’s blog on machine learning and computer vision, and I’m hooked. You should be too.

There’s the usual panoply of links to interesting papers, but then there’s also Tomasz’s radical idea for reimagining computer vision using a memex instead of a set of categories. He thinks that the “vision problem” will be solved by something much closer to actual AI than is generally considered necessary today. His ideas are informed by Wittgenstein and Vannevar Bush as well as contemporary research. It sounds interesting, to say the least. Then there’s also Tomasz’s stirring (if somewhat intimidating) advice to students and researchers to be Renaissance men and look beyond the A+ and the well-received publication. All in all, very worth reading.

(Sensible Sarah says: “Hey, wait a minute! I thought I was a math student — what’s up with all this vision stuff? And I have a qual to pass in a month!” Sensible Sarah throws up her hands in dismay.)


Church: a language for probabilistic computation October 23, 2010

Posted by Sarah in Uncategorized.
Tags: ,

As a novice with some interest in applied areas, it happens all too often that I fall in love with an idea about perception/cognition/data analysis, and find out only later “Oh wait, that wasn’t actually math! And I’d have to do so much background reading in a completely different field to understand it!” I’ve experienced “Oh wait, I’m not an electrical engineer!”, “Oh wait, I’m not a neuroscientist!”, and “Oh wait, I’m not a statistician!”

Well, today is the day I say “Oh wait, I’m not a computer scientist!”

The idea that caught my attention is probabilistic computing. We often want to build a machine that can make predictions and build models (a computer that can diagnose medical symptoms, or predict which creditors will default, or even, dare we whisper, an AI). This is essentially a Bayesian task: given some data, which probability functions best explain it? The trouble is, computers are bad at this. For the most part, they’re built to do the opposite task: given probability distributions and models, simulate some data. Generating probability functions and finding the best one can be prohibitively expensive, because the space of probability functions is so large. Also, while the computational complexity of evaluating f(g(x)) is just f + g, the computational complexity of composing two conditional probability distributions B|A and C|B is

ΣB P(C, B|A)

whose computational time will grow exponentially rather than linearly as we compose more distributions.

Church, a language developed at MIT is an attempt to solve this problem. (Apparently it’s a practical attempt, because the founders have already started a company, Navia Systems, using this structure to build probabilistic computers.) The idea is, instead of describing a probability distribution as a deterministic procedure that evaluates the probabilities of different events, represent them in terms of probabilistic procedures for generating samples from them. That is, a random variable is actually a random variable. This means that repeating a computation will not give the same result each time, because evaluating a random variable doesn’t give the same result each time. There’s a computational advantage here because it’s possible to compose random variables without summing over all possible values.

The nice thing about Church (which is based on Lisp, and named after Alonzo Church) is that it allows you to compose practically any query without significantly increasing runtime. “What is the probability of A and B or C and not D given that X and Y and not Z?” and so on.

The PhD dissertation that introduced Church is where I started, but it’s actually much clearer if you learn about it from the interactive tutorial. The tutorial is really a beautiful thing, well-written to the point of addictiveness. It is completely accessible to just about anyone.

Current reading: neuroscience September 30, 2010

Posted by Sarah in Uncategorized.
Tags: ,

Right now my “interests” are supposed to be limited to passing quals. Fair enough, but I’m also persistently trying to get a sense of what math can tell us about how humans think and how to model it. Except that I don’t actually know any neuroscience. So I’ve been remedying that.

Here’s one overview paper that goes over the state of the field, in terms of brain architecture and hierarchical organization. Neurons literally form circuits, and, in rough outline, we know where those circuits are. We can look at the responses of those circuits in vivo to observe the ways in which the brain clusters and organizes content: even to the point of constructing a proto-grammar based on a tree of responses to different sentences. I hadn’t realized that so much was known already — the brain is mysterious, of course, but it’s less mysterious than I had imagined.

Then here’s an overview paper by Yale’s Steve Zucker about image detection using differential geometry. In his model, detection of edges and textures is based on the tangent bundle. Apparently, unlike some approaches in computational vision, this differential geometry approach has neurological correlates in the structure of the connections in the visual cortex. The visual cortex is arranged in a set of columns; the hypothesis is that these represent \mathbb{R} \times S^1, with the column representing position and the slices at different heights of the columns representing orientation.