Concept Drift (Time Variant) Online Learning
In cannonical machine learning it is assumed that the process trying to be learned is stationary. Under this assumption, a model is fit to some precollected sample data and then used to predict future values. In this research I considered the setting in which the process being learned changes with time. Because a precollected training set cannot reflect the currect state of the target function, online learning must be used. I examined several basic learners and developed addaptive learning methods which allow the learner to predict how quickly the target function is evolving. While our initial results are exciting, I believe there is considerable room for exploration in this field. As a graduate student I hope to continue my research in dynamic learning. (The paper and code will be posted shortly)
Typed Functional Assembly
Tomorrow's computer architectures will be vastly different from those we use today. One of the emerging concept in computer architectures is multicore programmable hardware. This will enable programs to run concurrent operations and load gate logic on the fly. Despite substantial advancement is processor architecture little has been done to redesign the basic languages in which we communicate with this hardware. To facilitate processor level parallelization I made the assembly functional. I removed the concept of registers and instead introduced a stack which can be randomly accessed but has the property that items near the top have lower latency. To enable hardware reconfiguration, I introduced function registers which can implement the gate logic of functions (think FPGAs). Finally to improve saftey, I introduced a static type inference system. All functions are required to have type annotations. Since every operation in a function is a function application, we can statically verify the types. To test this assembly we implemented a virtual machine. (The code and paper will be posted shortly)
Evaluating Facial Attractiveness
Computer dating services match people based on social characteristics. What if they could match people based on how physically attracted they would be? What if you could give the computer a sample of faces you personally found attractive and it could find other people who you would also be attractive to you? This was the focus of my research in computer vision. We applied various kernel techniques, PCA, and Fisher Faces, along with support vector machines to construct a classifier which could automatically evaluate facial attractiveness. Results thus far have only been marginally successful. However, the computer does find me to be very attractive. (The code and paper will be posted shortly.)
Query-less Search
New technologies and communication media have enabled researchers to collect data faster than it may be assimilated. To sort through vast amounts of information, powerful query-driven technologies have been developed. However, query-driven research is limited to the quality of the query generated by the user. We have developed an unsupervised query-less search technology that selects new documents based on documents the user has already read. Latent semantic analysis is used to map words to vectors in a high-dimensional semantic space. Relative changes in these vectors are used to assess the effect a new document has on the abstract concepts that are associated with each word in the reader's vocabular. We then apply various metrics to measure changes in these associations. Documents are then ranked on their relative effect on the semantic association of words. (The code and paper will be posted shortly.)
Line of Sight
The ability to rapidly and accurately assess visibility between two points on large terrain models is paramount to the success of military simulations, real time terrain visualization, and the placement of LOS- dependent communications systems. Currently, there are two primary techniques for evaluating line-of-sight (LOS) on digital elevation maps (DEMs). Both techniques may produce less accurate results than traversing a regular triangulation of a DEM, which is usually too slow for real time simulations. I have developed a new algorithm for evaluating LOS that is as accurate as traversing a regular triangulation of the DEM and is faster than the techniques currently used. Based on the trivial rejection approach commonly used in ray tracing, I have applied quadtree space partitioning to reduce the LOS evaluation problem from an O(n) search to a O(log n) search. (The code and paper will be posted shortly.)
Interval Analysis
In space exploration, computer graphics, and computational physics we often need fast and accurate ways to numerically compute the intersection of nonlinear surfaces in relatively high dimensions. Inorder to guide spacecraft along equipotential manifolds in space, time, velocity, and mass we need not only accurate numerical techniques but also a precise bound on the error. Interval analysis can be applied to this problem, to provide both an accurate and fast way to computer surface intersections and a reliable error bound on the results. Working with Professor Alan Barr I developed an efficient algorithm for solving high dimensional systems of nonlinear equations by applying interval analysis and a branch and bound search. I was able to generalize the bounding planes approach to increasingly complex surfaces and implemented a basic prototype extending the thesis work of a previous graduate student. (The code and paper will be posted shortly.)
Growing Globally Optimal Irregular Triangular Networks
The triangulation of a given set of points in a plane is common to computer graphics, computer vision, and robotics. Unfortunately, the construction of globally optimal triangulations with multiple optimality criteria is believed to be an NP complete problem. Thus, no efficient polynomial time solution exists. However, solutions to globally optimal triangulations can be computed with manageable time complexity using nondeterministic techniques. In the paper by Kolingerova and Ferko, such a method is developed. By using a genetic algorithm which favors globally optimal triangulations in an evolving "population" of solutions, Kolingerova and Ferko are applying genetic optimization to solve the globally optimal triangulation problem. We implemented their algorithm and tested in on various fitness metrics. This was done as part of a freshman class (CS20) in algorithms and numerics. <more> (The code and paper will be posted shortly.)
