Muhammad Asiful Islam

Bagging and Boosting Negatively Correlated Neural Networks

In this paper, we propose two cooperative ensemble learning algorithms, i.e., NegBagg and NegBoost, for designing neural network (NN) ensembles. The proposed algorithms incrementally train different individual NNs in an ensemble using the negative correlation learning algorithm. Bagging and boosting algorithms are used in NegBagg and NegBoost, respectively, to create different training sets for different NNs in the ensemble. The idea behind using negative correlation learning in conjunction with the bagging/boosting algorithm is to facilitate interaction and cooperation among NNs during their training. Both NegBagg and NegBoost use a constructive approach to automatically determine the number of hidden neurons for NNs. NegBoost also uses the constructive approach to automatically determine the number of NNs for the ensemble. The two algorithms have been tested on a number of benchmark problems in machine learning and NNs, including Australian credit card assessment, breast cancer, diabetes, glass, heart disease, letter recognition, satellite, soybean, and waveform problems. The experimental results show that NegBagg and NegBoost require a small number of training epochs to produce compact NN ensembles with good generalization.

Hearsay: a new generation context-driven multi-modal assistive web browser

This demo will present HearSay, a multi-modal non-visual web browser, which aims to bridge the growing Web Accessibility divide between individuals with visual impairments and their sighted counterparts, and to facilitate full participation of blind individuals in the growing Web-based society.

Inference in probabilistic logic programs with continuous random variables

Probabilistic Logic Programming (PLP), exemplified by Sato and Kameyas PRISM, Pooles ICL, Raedt et al.s ProbLog and Vennekens et al.s LPAD, is aimed at combining statistical and logical knowledge representation and inference. However, the inference techniques used in these works rely on enumerating sets of explanations for a query answer. Consequently, these languages permit very limited use of random variables with continuous distributions. In this paper, we present a symbolic inference procedure that uses constraints and represents sets of explanations without enumeration. This permits us to reason over PLPs with Gaussian or Gamma-distributed random variables (in addition to discrete-valued random variables) and linear equality constraints over reals. We develop the inference procedure in the context of PRISM; however the procedures core ideas can be easily applied to other PLP languages as well. An interesting aspect of our inference procedure is that PRISMs query evaluation process becomes a special case in the absence of any continuous random variables in the program. The symbolic inference procedure enables us to reason over complex probabilistic models such as Kalman filters and a large subclass of Hybrid Bayesian networks that were hitherto not possible in PLP frameworks.

Bridging the Web Accessibility Divide

The Web has become the primary medium for accessing information and for conducting many types of online transactions, including shopping, paying bills, making travel plans, etc. The primary mode of interaction over the Web is via graphical browsers designed for visual navigation. Sighted users can visually segment web pages and quickly identify relevant information. On the contrary, screen readers - the dominant assistive technology used by visually impaired individuals - function by speaking out the screens content serially. Consequently, users with visual impairments are forced to listen to the information in web pages sequentially, thereby experiencing considerable information overload. This problem becomes even more prominent when conducting online transactions that often involve a number of steps spanning several pages. Thus, there is a large gap in Web accessibility between individuals with visual impairments and their sighted counterparts. This paper we describe our ongoing work on this problem. We have developed several techniques that synergistically couple web content analysis, users browsing context, process modeling and machine learning to bridge this divide. These techniques include: 1) context-directed browsing that uses link context to find relevant information as users move from page to page; 2) change detection that separates the interface from the implementation of web pages and helps users find relevant information in changing web content; and 3) process modeling that helps users find concepts relevant in web transactions. We describe these three techniques within the context of our Hearsay non-visual web browser.

Mixture model based label association techniques for web accessibility

An important aspect of making the Web accessible to blind users is ensuring that all important web page elements such as links, clickable buttons, and form fields have explicitly assigned labels. Properly labeled content is then correctly read out by screen readers, a dominant assistive technology used by blind users. In particular, improperly labeled form fields can critically impede online transactions such as shopping, paying bills, etc. with screen readers. Very often labels are not associated with form fields or are missing altogether, making form filling a challenge for blind users. Algorithms for associating a form element with one of several candidate labels in its vicinity must cope with the variability of the elements features including labels location relative to the element, distance to the element, etc. Probabilistic models provide a natural machinery to reason with such uncertainties. In this paper we present a Finite Mixture Model (FMM) formulation of the label association problem. The variability of feature values are captured in the FMM by a mixture of random variables that are drawn from parameterized distributions. Then, the most likely label to be paired with a form element is computed by maximizing the log-likelihood of the feature data using the Expectation-Maximization algorithm. We also adapt the FMM approach for two related problems: assigning labels (from an external Knowledge Base) to form elements that have no candidate labels in their vicinity and for quickly identifying clickable elements such as add-to-cart, checkout, etc., used in online transactions even when these elements do not have textual captions (e.g., image buttons w/o alternative text). We provide a quantitative evaluation of our techniques, as well as a user study with two blind subjects who used an aural web browser implementing our approach.

Tightly coupling visual and linguistic features for enriching audio-based web browsing experience

People who are blind use screen readers for browsing web pages. Since screen readers read out content serially, a naive readout tends to mix irrelevant and relevant content thereby disrupting the coherency of the material being read out and confusing the listener. To address this problem we can partition web pages into coherent segments and narrate each such piece separately. Extant methods to do segmentation use visual and structural cues without taking the semantics into account and consequently create segments containing irrelevant material. In this paper, we describe a new technique for creating coherent segments by tightly coupling visual, structural, and linguistic features present in the content. A notable aspect of the technique is that it produces segments with little irrelevant content. Preliminary experiments indicate that the technique is effective in creating highly coherent segments and the experiences of an early adopter who is blind suggest that it enriches the overall browsing experience.