Marc Parizeau

Training hidden Markov models with multiple observations-a combinatorial method

Hidden Markov models (HMM) are stochastic models capable of statistical learning and classification. They have been applied in speech recognition and handwriting recognition because of their great adaptability and versatility in handling sequential signals. On the other hand, as these models have a complex structure and also because the involved data sets usually contain uncertainty, it is difficult to analyze the multiple observation training problem without certain assumptions. For many years researchers have used the training equations of Levinson (1983) in speech and handwriting applications, simply assuming that all observations are independent of each other. This paper presents a formal treatment of HMM multiple observation training without imposing the above assumption. In this treatment, the multiple observation probability is expressed as a combination of individual observation probabilities without losing generality. This combinatorial method gives one more freedom in making different dependence-independence assumptions. By generalizing Baums auxiliary function into this framework and building up an associated objective function using the Lagrange multiplier method, it is proven that the derived training equations guarantee the maximization of the objective function. Furthermore, we show that Levinsons training equations can be easily derived as a special case in this treatment.

GENERICITY IN EVOLUTIONARY COMPUTATION SOFTWARE TOOLS: PRINCIPLES AND CASE-STUDY

This paper deals with the need for generic software development tools in evolutionary computations (EC). These tools will be essential for the next generation of evolutionary algorithms where application designers and researchers will need to mix different combinations of traditional EC (e.g. genetic algorithms, genetic programming, evolutionary strategies, etc.), or to create new variations of these EC, in order to solve complex real world problems. Six basic principles are proposed to guide the development of such tools. These principles are then used to evaluate six freely available, widely used EC software tools. Finally, the design of Open BEAGLE, the framework developed by the authors, is presented in more detail.

A fuzzy-syntactic approach to allograph modeling for cursive script recognition

This paper presents an original method for creating allograph models and recognizing them within cursive handwriting. This method concentrates on the morphological aspect of cursive script recognition. It uses fuzzy-shape grammars to define the morphological characteristics of conventional allographs which can be viewed as basic knowledge for developing a writer independent recognition system. The system uses no linguistic knowledge to output character sequences that possibly correspond to an unknown cursive word input. The recognition method is tested using multi-writer cursive random letter sequences. For a test dataset containing a handwritten cursive text 600 characters in length written by ten different writers, average character recognition rates of 84.4% to 91.6% are obtained, depending on whether only the best character sequence output of the system is considered or if the best of the top 10 is accepted. These results are achieved without any writer-dependent tuning. The same dataset is used to evaluate the performance of human readers. An average recognition rate of 96.0% was reached, using ten different readers, presented with randomized samples of each writer. The worst reader-writer performance was 78.3%. Moreover, results show that system performances are highly correlated with human performances. >

Genetic programming, validation sets, and parsimony pressure

Fitness functions based on test cases are very common in Genetic Programming (GP). This process can be assimilated to a learning task, with the inference of models from a limited number of samples. This paper is an investigation on two methods to improve generalization in GP-based learning: 1) the selection of the best-of-run individuals using a three data sets methodology, and 2) the application of parsimony pressure in order to reduce the complexity of the solutions. Results using GP in a binary classification setup show that while the accuracy on the test sets is preserved, with less variances compared to baseline results, the mean tree size obtained with the tested methods is significantly reduced.

Segmentation and reconstruction of on-line handwritten scripts

Abstract On-line handwritten scripts consist of sequences of components that are pen tip traces from pen-down to pen-up positions. This paper presents a segmentation and reconstruction procedure which segments components of a script into sequences of static strokes, and then reconstructs the script from these sequences. The segmentation is based on the extrema of curvature and inflection points in individual components. The static strokes are derived from the delta log-normal model of handwriting generation and are used in component representation and reconstruction. The performance of the procedure is measured in terms of root-mean-square reconstruction error and data compression rate.

Signature verification from position, velocity and acceleration signals: a comparative study

The performance of position, velocity, and acceleration signals for automatic signature verification is compared. Three types of signal comparison algorithms (dynamic time warping, regional correlation and tree matching) are used on a single database. Variance analysis shows that significant differences exist among the three signal representation spaces. Among other things, it is shown that the most discriminant are signals which reflect the vertical activity of the signing process and that the best representation space for a 2-D signature verification system is the velocity domain.<<ETX>>

Probabilistic Sensing Model for Sensor Placement Optimization Based on Line-of-Sight Coverage

This paper proposes a probabilistic sensor model for the optimization of sensor placement. Traditional schemes rely on simple sensor behaviour and environmental factors. The consequences of these oversimplifications are unrealistic simulation of sensor performance and, thus, suboptimal sensor placement. In this paper, we develop a novel probabilistic sensing model for sensors with line-of-sight-based coverage (e.g., cameras) to tackle the sensor placement problem for these sensors. The probabilistic sensing model consists of membership functions for sensing range and sensing angle, which takes into consideration sensing capacity probability as well as critical environmental factors such as terrain topography. We then implement several optimization schemes for sensor placement optimization, including simulated annealing, limited-memory Broyden-Fletcher-Goldfarb-Shanno method, and covariance matrix adaptation evolution strategy.

Experiments on eigenfaces robustness

This paper is an experimental study on the robustness of the eigenfaces method for face recognition. To build a face recognition system, especially in an unconstrained surveillance system where a clear, direct, and normalized view of the face cannot be assumed, one needs to implement several image preprocessing steps like segmentation, deskewing, zooming, rotation, warping, etc., before processing the face image per se. Our aim is to determine how efficient these preprocessing steps must be in order to apply the eigenfaces method with success. The experiments are conducted on a subset of the AR-face color image database. Real images are used and altered synthetically to study the effects of 7 parameters that can be translated into corresponding preprocessing artifacts: horizontal and vertical translations, downsampling, zooming, rotation, morphing and lighting.

The Agile Stereo Pair for active vision

This paper presents a new stereo sensor for active vision. Its cameras are mounted on two independent 2-DOF manipulators, which are themselves mounted on two translation stages. The system is designed for fast and accurate dynamical adjustments of gaze, vergence, and baseline. A complete description of its software and hardware components is given, including a detailed discussion of its calibration procedure. The performance of the sensor with respect to dynamical properties and measurement accuracy is also demonstrated through both simulations and experiments.

DEAP: a python framework for evolutionary algorithms

DEAP (Distributed Evolutionary Algorithms in Python) is a novel volutionary computation framework for rapid prototyping and testing of ideas. Its design departs from most other existing frameworks in that it seeks to make algorithms explicit and data structures transparent, as opposed to the more common black box type of frameworks. It also incorporates easy parallelism where users need not concern themselves with gory implementation details like synchronization and load balancing, only functional decomposition. Several examples illustrate the multiple properties of DEAP.