I attended a really nice talk by Arash Amini yesterday about detecting ‘communities’ in sparse graphs. The basic problem is: In a big graph (like the Facebook graph, or the graph of scientific papers citations) you have clusters of friends/associates, and you want to pick out those clusters. Dr. Amini and his collaborators have been working on methods to solve this problem in particularly noisy cases. The methods presented were developed for cases where you know the specific number of clusters you are seeking to find. It was a fascinating talk, and I came away wondering if there are good methods to use when you aren’t sure of how many clusters there are.
I’ve finally had a bit of time to finish up the code for the Rock-Paper-Scissors bot discussed in the previous posts; I’ve put the code in a GitHub repository here. Feel free to file issues if you would like it to be a bit more user friendly.
The bot which uses the Fourier transform on move probabilities to search for profitable patterns is named `diaconis`, after Persi Diaconis. It’s currently working just fine, but is a bit slow computationally. On startup, it generates all possible move patterns that it will investigate during the course of each game, as well as some character tables. After each play, it tests some of these patterns, and tries to determine if their predictive power is better than any of the patterns seen thus far. If so, it begins using that pattern to choose move probabilities.
This works fine as a proof-of-concept of the basic ideas. Additional improvements could be had by doing some code optimization to speed things up a bit, and keeping a list of good patterns and allowing a bit more dexterity in switching between the patterns used for prediction.
I’m teaching an intro calculus class this year (specifically, ‘Math for Life and Social Science’), and came a while ago to the section on optimization. It’s a really important subject, and yet the optimization problems one finds in Calculus books (even good ones) tend to be contrived examples which I refer to as ‘box factory problems.’ Things along the lines of ‘minimize the surface area of a rectangular box with volume ‘.
These are fine for a problem or two: There’s a useful skill in taking a real-sounding problem and translating it into the mathematics you’re learning. We use the constraints (in this case, on the volume) to reduce the number of dimensions, turn the problem into a one-variable calculus problem, and then solve. All well and good, but these problems somehow completely miss the impact of optimization on society at large. Largely because the optimization problems that occur most commonly in the wild have a slightly different flavour.
Problem: In Boston, we observe that the monthly rent for three one-bedroom apartments are $1300, $1150, and $950. Rent on three two-bedroom apartments were $1500, $1700, and $1200. Assuming that the cost of a 0-bedroom apartment is $500, find the best possible line describing the rent as a function of the number of bedrooms. Continue reading
In the last two posts, we’ve looked at using machine learning for playing iterated Roshambo. Specifically, we saw how to use Bayes’ theorem to try to detect and exploit patterns, and then saw how Fourier transforms can give us a concrete measurement of the randomness (and non-randomness) in our opponent’s play. Today’s post is about how we can use representation theory to improve our chances of finding interesting patterns.
In the last post, we looked at using an algorithm suggested by Bayes’ Theorem to learn patterns in an opponent’s play and exploit them. The game we’re playing is iterated rock-paper-scissors, with 1000 rounds of play per game. The opponent’s moves are a string of choices, ‘r’, ‘p’, or ‘s’, and if we can predict what they will play, we’ll be able to beat them. In trying to discover patterns automatically we’ll gain some general knowledge about detecting patterns in streams of characters, which has interesting applications ranging from biology (imagine ‘GATC’ instead of ‘rps’) to cryptography.
Fourier analysis is helpful in a wide variety of domains, ranging from music to image encoding. A great example suggested by ‘Building Machine Learning Algorithms with Python‘ is classifying pieces of music by genre. If we’re given a wave-form of a piece of music, automatically detecting its genre is difficult. But applying the Fourier transform breaks the music up into its component frequencies, which turn out to be quite useful in determining whether a song is (say) classical or metal.
I’ve recently been doing some reading on machine learning with a mind towards applying some of my prior knowledge of representation theory. The initial realization that representation theory might have some interesting applications in machine learning came from discussions with Chris Olah at the Toronto HackLab a couple months ago; you can get interesting new insights by exploring new spaces! Over winter break I’ve been reading Bishop’s ‘Pattern Recognition and Machine Learning‘ (slowly), alongside faster reads like ‘Building Machine Learning Systems with Python.‘ As I’ve read, I’ve realized that there is plenty of room for introducing group theory into machine learning in interesting ways. (Note: This is the first of a few posts on this topic.)
There’s a strong tradition of statisticians using group theory, perhaps most famously Persi Diaconis, who used representation theory of the symmetric group to find the mixing time for card shuffling. His notes ‘Group Representations in Probability and Statistics‘ are an excellent place to pick up the background material with a strong eye towards applications. Over the next few posts I’ll make a case for the use of representation theory in machine learning, emphasizing automatic factor selection and Bayesian methods.
First, an extremely brief overview of what machine learning is about, and an introduction to using a Bayesian approach to play RoShamBo, or rock paper scissors. In the second post, I’ll motivate the viewpoint of representation by exploring the Fourier transform and how to use it to beat repetitive opponents. Finally, in the third post I’ll look at how we can use representations to select factors for our Bayesian algorithm by examining likelihood functions as functions on a group.
Recently I’ve been playing with building a regression model for the brightness of images produced with the Raspberry Pi’s camera board. Essentially, I want to quickly figure out – hopefully from a single image – what shutter speed and ISO to choose to get an image of a given brightness.
This is a pretty standard regression problem: We some data, extract some information from it, and use that information to make a prediction. To get a better handle on the algorithms involved, I wrote my own code to perform the regression, using NumPy for fast linear algebra operations. You always learn something from re-inventing the wheel, after all.
I’ve done quite a bit to improve my Raspberry Pi-based timelapser. I have a pretty good system together now, thanks in part to recent firmware updates to the Pi which allow me to directly control the shutter speed and ISO of the camera sensor instead of relying on auto-exposure settings. The auto-exposure is pretty unreliable from one shot to the next: the camera makes a different decision each time it snaps a picture, which leads to quite a lot of flicker. Previously, I was dealing with this flicker by manipulating the images in post-production, but I’ve now written some code to get the camera to try to maintain a constant image brightness across a long shoot.
The code now consists of a ‘timelapser’ class, which keeps track of its current shutterspeed and ISO (SS/ISO henceforth), and the brightness of the last few images taken. It then adjusts SS/ISO to try to get the image brightness to 100. By keeping track of the last few images, it is a bit less susceptible to being upset by one strange image (like, say, if I put my hand over the lens for one shot, producing a black image), or more standard movement within the frame. On the other hand, it takes a while longer to settle down to the ‘right’ SS/ISO. So it’s currently set up with an initialization step, where it finds a good SS/ISO pretty quickly, and then transitions to actually taking pictures. The result is very little flicker as the timelapse goes on, and a pretty constant level of image brightness when light levels gradually change: like when we watch dawn or dusk. (If you’re interested in playing around with the code, I’ve set up a github repository here.)
As an example, this is a video that we shot over about three days on my friends Ketan and Ananya’s balcony. They have a great view over Toronto, from the CN Tower to Honest Ed’s.
Almost immediately after my first foray with the Raspberry Pi timelapse I was contacted by Cookie Roscoe, who coordinates a weekly farmer’s market for The Stop. The Stop is a community food center here in Toronto, providing lots of programming and emergency food access for lower-income people. My partner, Elizabeth, worked with them for about three years, so I was super happy to try doing a time-lapse of their market.
Here’s the results! Lots of technical stuff about what went into the video after the break!
My recent DIY electronics project has been putting together a Raspberry Pi-based camera. The Pi foundation sells a camera board which plugs into the Pi; it’s sold as ‘a bit better than the average camera in a mobile phone.’ But the Pi’s default Raspbian Linux installation comes with a couple programs for controlling the camera, and lets you take still pictures and videos easily.
In the interest of using the camera for something you can’t usually do with a store-bought digital camera, I wrote a short python script which takes a photo assigns it a file name based on the date and time taken. It then does some sampling of the picture, and only keeps pictures which aren’t ‘too dark.’ And then cron runs the photo script once every five minutes. In other words, the Pi is set up for long-form time-lapse photography. The resulting pictures are then easy to compile into a movie.