Projects and Software

This describes my recent major (or minor and beloved) projects, and software that I have released for wider use. Note that when I say "winter" I am referring to the first three months of the year, so Winter 2005 cannot mean December 2005.

SVMpython version 2

Wrote SVMpython version 2, a reimplementation of my original SVMpython. SVMpython is a Python embedded version of SVMstruct. One applies SVMstruct by modifying the svm_struct_api.c file and recompiling. SVMpython allows one to write these functions in Python instead: one applies SVMpython by creating a Python module (commonly just a .py file) with the appropriate methods. This module is loaded and specific methods called at runtime to support the structural learning algorithm.

One advantage of this Python embedding is this obviates many of the tedious non-learning-related maintenance procedures, allowing one to focus purely on the learning problem. The two most conspicuous examples are memory management, and reading and writing models to files (accomplished with Pickling). The primary disadvantage to using SVMpython is that it is slower than equivalent C code.

Spring 2007


Wrote PyGLPK, a Python module that encapsulates the functionality of the GNU Linear Programming Kit (GLPK). The GLPK allows one to specify linear programs (LPs) and mixed integer programs (MIPs), and to solve them with either simplex, interior-point, or branch-and-bound algorithms. The goal of PyGLPK is to give one access to all documented functionality of GLPK within a Pythonic (that is, non-SWIG) Python module.

Spring 2007


Wrote PyGraphcut, a Python module that encapsulates the functionality of some of the graph cut code from Vladimir Kolmogorov's page. My goal in writing the code is to support my own research, which required as a subcomponent an engine to solve quadratic pseudo-Boolean optimization problems, but it may be useful for other things. This software is less developed than PyGLPK.

Spring 2007

Nexus War Pathfinder

Wrote the Nexus War Pathfinder, a tool for finding the shortest distance between two points within the game Nexus War, a free browser based massively multipler game with which I was briefly obsessed in 2006. This game is played on a map, and often one wants to find the shortest path between two points.

This is, in some sense, an interesting instance of a serverless AJAX-like page, where instead of communicating with a remote server, the page communicates with a Java applet in page. Unfortunately browser support for this is somewhat "twitchy," so it only works in certain browsers.

Summer 2006

SVMlatte, A Super Secret Project

This project is so secretive I can't even talk about it. Either that or I can't be bothered to put anything sensible up yet. One of the two, that's for sure.

Fall 2005-present

SVMcluster, Supervised Clustering with Support Vector Machines (SVMs)

Worked with Thorsten Joachims on a supervised clustering algorithm trains over partitionings of sets to learn a similarity measure. This utilizes an existing framework SVMstruct, which learn functions with complex output spaces with SVMs. The algorithm learn to map a set of items to a partitioning over that set. Our current applications include document noun-phrase coreference, and clustering of related news stories. A paper describing this work appeared at ICML 2005 and won one of the distinguished student paper awards at that conference.

Clustering is often applied to problems where the user has a notion of the "correct" partitioning of a set. For example, in noun-phrase coreference, the goal is to group noun phrases so that the phrases that refer to the same entity are clustered together. This means that the user would like the clustering to reflect an equivalence relation with a particular semantic. Among all possible ways of partitioning a set of noun phrases (e.g. by morphological orgin, genus, etc.), a successful clustering algorithm should find clusterings that follow the user intended semantic.

Given training examples of sets (e.g. noun phrases) and their desired clusterings, the goal is to learn a function that predicts the clustering for new sets. Most previous approaches reduce this problem to a learning task over pairs of points. The idea is to learn a pair-wise similarity measure that induces the desired clustering. However, this approach decouples learning and clustering so that the learned similarity measure does not necessarily optimize to clustering effectiveness. We overcome this problem by directly optimizing to clustering effectiveness.

Winter 2004-Spring 2005


Wrote SVMpython, an extension to the existing SVMstruct package. SVMstruct allows users to write interface functions in C to enable machine learning of functions with complex outputs. SVMpython embeds Python, allowing users to write the functions in Python instead of C. This enables more rapid prototyping of new ideas for structured learning.

The primary advantages are that Python tends to be easier and faster to code than C, less resistant to change and code reorganization, tends to be many times more compact, there's no explicit memory management, and Python's object oriented-ness means that some tedious tasks in SVMstruct can be easily replaced with default built in behavior.

My favorite example of this last point is that, since Python objects can be assigned any attribute, and since many Python objects are easily serializable with the pickle module, adding a field to the struct-model in Python code consists of a simple assignment like = 5 at some point, and that's it. Using SVMstruct in C, one would add a field to the relevant struct, add an assignment, add code to write it to a model file, add code to parse it from a model file, and then test it to make sure all these little changes work well with each other.

The primary disadvantage to using SVMpython is that it is slower than equivalent C code. For example, considering the time outside of SVM optimization, the Python implementation of multiclass classification takes 9 times the time of SVMmulticlass. However, on this task the SVM optimization takes about 99.5% of the time anyway, so the nine-fold increase in that 0.5% is negligible. Of course, other problems may have different performance characteristics.

Winter 2005-Summer 2005

Ornithology Data Mining


Performing data mining on data from the Project FeederWatch ornithology survey. This project is a collaboration between the Lab of Ornithology and the CS department. In this survey, many thousands of biologists and citizen volunteers collect bird abundance data from all around the US and Canada. The high-level goal of the project is to take this raw data and to transform it into something readily accessible and understandable by everyone to explore changes in bird abundance. This project is for the birds!

I am working on this project with Mohamed Elhawary and Daria Sorokina, under the supervision of Mirek Riedewald and Rich Caruana. The birds are counting on us!

Winter 2005-Spring 2005


Fob Window

Developed Fob under the name Leaky Puppy Software. Fob is freeware software for Mac OS X that lets one set countdown timers. The purpose was to teach myself Objective-C and the Cocoa API. I was extremely gratified when Fob received three-and-a-half mice 3 1/2 mice (July 2004) and a later version received four mice 4 mice (November 2004) from Macworld!

The presence of presets is Fob's main gimmick for decreasing the effort required to start a timer, and seems to be what people like most about it. To draw an analogy to web browsers, they are a kind of time bookmark for commonly used timers. For example, I may define a six minute preset to let tea steep. Thence, I need only choose "Tea" from my preset list, and I never need deal with a neglected lukewarm cup in my microwave again. Starting a preset timer requires only a single double click rather than many clicks and typing.

Winter-Summer 2003, minor dev. in 2004


JFLAP Automata Editor Window

Developed JFLAP 4.0, a new version of a visual interactive software tool for teaching automata theory. JFLAP lets users create structures related to complexity theory (automata, regular expressions, grammars, L-systems) and perform guided procedures on structures that correct mistakes students make. JFLAP is another project of Susan Rodger's.

Currently at well over 50,000 lines of beautiful code, JFLAP has the ability to create automata (FA, NPDA, TM), grammars (context-free and unrestricted), regular expressions, and L-systems (normal and contextual). Once these structures are created, one can simulate input on these structures; this includes simulating input on automata, building LL(1) and SLR(1) parse trees for grammars and parsing input with these trees, brute force parsing of input on even an unrestricted grammar, and rendering L-systems. One can also determine various properties of structures, like if two FA are equivalent. There is also a suite of conversion operations like converting an NFA to a DFA, minimizing a DFA, converting an FA or PDA to a grammar, convering an FA to an RE, converting an RE to an NFA, converting context-free grammars to NPDA (either LR and LL), converting a right linear grammar to an FA, and converting grammars to CNF.

One of my favorite abilities of JFLAP is the ability to "print" structures, not so much because I like to hang them on my wall, but because by saving it as a PDF, you can have a very pretty vector graphic of either an automata or an L-system rendering.

Summer 2002-Summer 2003, minor dev. to present

JFLAP User Manual

Coauthored a user manual for JFLAP 4.0 with Susan Rodger. I wrote 6 of the 11 chapters, including the introductory chapters on finite automata, regular grammars, regular expressions, Turing machines, unrestricted and context-sensitive grammars, and L-systems.

See the publisher's page for the user manual!

Spring-Fall 2003, revision through Fall 2005

JAWAA 2.0 Editor

JAWAA Editor Window

Developed a JAWAA editor for easily defining JAWAA animations in a GUI. JAWAA (Java And Web based Algorithm Animation) is an animation language invented by Susan Rodger, and can be used to visualize algorithms or just make fun animations.

For the life of me, I cannot figure out how to get the JAWAA applet to play any .anim file once it has been defined.

Summer 2001

Thomas Finley : Home About Me Research Projects Notes