Research

       My primary research interest is in unsupervised learning techniques for unstructured and noisy environments, with a specific focus on language aquisition. I am currendly developing algorithms for automatic speech segmentation with an eye towards unsupervised linguistic concept learning. I draw inspiration from developmental psychology, and believe that, while it may not be neccessary to replicate human physiology or development, doing so is the most obvious path towards creating artificial intelligence. Consequently, I am interested in both the neural basis of cognition, and the stages of human development.
       Along these same lines, I am interested in grounded concept learning, and have been involved with the creation of the Developmental Robotics Lab at Iowa State University. We have constructed a torso robot with two Barrett WAMs for arms. When the head, face and waist actuation is complete, it may be the most advanced humanoid torso in the world. It will be used to study affordances, specifically of tools and containers, as well as to provide a platform for other embodied learning research. I have worked closely with Alexander Stoytchev and Jivko Sinapov on this project, and am mainly involved in writing software to control the arms.
       I am also working on an Infomax inspired version of the Efficient Sparse Coding algorithm. Sparse Coding is an extremely powerful feature extraction algorithm, which has recently been solved using an efficient itterative quadratic optimization technique. The algorithm learns a set of basis vectors that are then used to sparsely represent a set of datapoints in some high dimensional space (i.e. a very few number of basis vectors are used to represent each point). Unfortunately, this algorithm has no constraint on how much each basis vector must be used. In many cases, the algorithm learns a set of basis vectors such that a small number of them are used to reconstruct almost every point in the dataset, and the rest are hardly used at all. According to the infomax model, in order to maximize the efficiency of the representation, the entropy of the activations of the basis vectors should be maximized. I am currently working to design a heuristic that can be added to the objective function of the quadtratic program that will enforce this constraint. Currently all straightforward alterations make the problem immediately intractable.
       I am also currently involved in a research project in the Human-Computer Interaction department at Iowa State. We are developing software to assist humans during manufacturing tasks by detecting and alerting them if a mistake is being made. We track different tools used during the manufacturing process, and infer whether the manufacturing task is being executed correctly. I developed a graphical model based on an inverse HMM to solve this problem, and extended it to account for dynamically added states. This research is ongoing, and we are currently developing algorithms for automatic tool path creation as well as rotation invariance. We have filed for a patent on this process, and it is currently pending.