Henry Stern’s research while affiliated with Dalhousie University and other places

Latent Semantic Indexing (LSI) is an information retrieval (IR) method that connects IR with numerical linear algebra by representing a dataset as a term-document matrix. Because of the tremendous size of modern databases, such matrices can be very large. The partial singular value decomposition (PSVD) is a matrix factorization that captures the salient features of a matrix, while using much less storage. We look at two challenges posed by this PSVD data compression process in LSI. Traditional methods of computing the PSVD are very expensive; most of the processing time in LSI is spent in calculating the PSVD of the term-document matrix. Thus, the first challenge is calculating the PSVD efficiently, in terms of computational and memory requirements. The second challenge is efficiently updating the PSVD when the matrix is altered slightly. In a rapidly expanding environment, such as the Internet, the term-document matrix is altered often as new documents and terms are added. Updating the PSVD of this matrix is much more efficient than recalculating it after each change. We investigate the use of the PSVD updating

In this paper we describe an information architecture that supports the dynamic composition of Web based lessons based on a database of fine grained components that can be tailored for communities of users. Our goal is to support Web accessible resources of pedagogical material that can be used and reused in a variety of contexts from teacher remediation and lesson preparation to the use of lessons that reflects the diversity of individual learning styles. The Web based lessons are generated on the fly from a database of fine grained components that include explanations, interactive activities, worksheets, images, videos, audio, FAQ's, online questions and challenges. This architecture had been used in the development of learning support systems for two different communities; users of middle school mathematics in one case and users of health informatics learning modules in the other case.

Difficult non-linear problems can be mapped to a set of lo-calised linear problems. A self-organising map (SOM) is used as a gating function to a localised mixture of experts classifier and is shown to find solutions equivalent to those learned by a multi-layer perceptron while retaining the simplicity and resilience of a single-layer perceptron. Mod-ifications to the traditional softmax gate function and expert training function are shown to produce more accurate classifiers that tend not to over-fit to training data.

We present a method of exhaustively solving the popular combinatorial game Amazons in a parallel setting. The game has a high branching factor with high overlap, making it difficult to search in parallel. We exploit some structural properties of game positions to minimise game tree expansion, and use a parallel breadth-first search algorithm on the game tree. Throughput is maximised by allowing some minimal work duplication, avoiding a high communication overhead. We intend to use this method to solve the full 10x10 game with 4 amazons per player.

We present a potential vulnerability of personalised anti-spam l- ters where an attacker sends carefully constructed e-mail messages with the goal of negatively a ecting classi er accuracy. Words from the core of the English language are randomly \injected" into spam e-mails for the express purpose of manipulating the probability tables of a nave Bayesian classi er. This attack method is shown to be successful in reducing classi er accuracy within a laboratory environment. Barri- ers to a real-world implementation and potential countermeasures are discussed.