networks & systems laboratory> research> current projects> research interests & current projects

Research Interests & Current Projects
Computer Human Adapted Interaction Research Group

Research Interests

- Machine learning, neural networks, data mining, pattern recognition.

- Methods for classification, prediction and clustering; hybrid symbolic/neural techniques for intelligent processing.

- Applications, especially in bio- and medical informatics, multimedia and video processing, content-based video retrieval, recommendation systems and text categorizations.

Content-Based Access to Digital Video Libraries

- Applications such as video-on-demand, digital TV broadcast, digital libraries, distance learning generate and use large collections of video data. Content-based systems are currently being developed to automatically organize video in order to provide fast and meaningful nonlinear access to the relevant material in video. They allow to browse the content of the video and search quickly for subsequences of interest without the need to watch the entire video.

- The goal of this project is to investigate a neural network based framework for more efficient temporal video segmentation and video summarization of MPEG-2 compressed video. It involves the application of incremental self-organizing neural networks such as Growing Neural Gas and Growing Cell Structures.

- Supported by SESQUI grant

Fingerprint Recognition

- Although many systems for fingerprint recognition are already available, there is still a need for further research to improve their reliability and performance. Most commercial systems involve complicated data processing to locate minutiae and then combinatorics algorithms to match them. In contrast, computational intelligence methods (e.g. neural networks, machine learning techniques, genetic algorithms) learn by examples and are powerful mechanism for problems that are hard to solve analytically.

- Anna Cequerra (Honors student in 2001), in collaboration with Piero Calucci and Andrea Aizza, from Tender S.p.A., and Fabio Vitali and Sergio Carrato, Image Processing Laboratory, University of Trieste, Italy, has developed an approach which combines local and global fingerprint image features. In her specific implementation minutiae and shape signatures were integrated and the final recognition was done by a neural network.

- The aim of this project is to further study the potential of computational intelligence methods for fingerprint recognition and also to investigate other combinations of local and global fingerprint recognition schemes.

Mining of Microarray Gene Expression Data

- DNA microarrays have revolutionized DNA analysis by allowing biologists to detect and monitor simultaneously the activities of thousands of genes. The raw microarray data are images which can be transformed into gene expression matrices - the rows correspond to genes, columns represent different experimental conditions, and the numbers in each cell are the expression level of the genes in the respective conditions. These matrices need to be analyzed further and the challenge is how to interpret such massive and complex data. This is where data mining can help! Data mining refers to the process of automatically extracting implicit, previously unknown and potentially useful information from large databases.

- The aim of this project is to develop data mining methods and tools for analysis, interpretation and visualization of microarray gene expression data. For example, unsupervised methods (like clustering) can be applied to determine how a set of genes cluster into groups, and hence to understand more about the complex mapping from genes to functions (similar genes yield similar expression patterns) and also to study the effects of various treatments or diseases.

- Supervised algorithms (neural networks, decision trees, support vector machines, etc.) can be used to construct a classifier based on labeled gene expression data that is able to distinguish between different conditions (e.g. healthy or diseased) or tissues (tumor or tumor-free) and be used for diagnostics.

- Honors student 2002: Ehremin Avila.

- Honors student 2001: Kim Jackson; third year advanced project: Keith Player; collaboration with Dr. Janette Burgess, Dept. of Pharmacology.

 

Recommendation Systems

- Recommender systems make suggestions to a user applying statistical and machine learning techniques. For example, they may predict whether a user would be interested in seeing a particular movie. Consider the following scenario: you tell the system what movies you have already seen and how you like them; the system uses this information to make suggestion in 2 ways: 1) collaborative filtering - it matches your tastes to those of the other users who share your likes and dislikes and suggests movies that they have seen and liked but you haven’t seen; 2) content-based approach - it recommends movies similar to those you prefer based on a comparison of movie content, i.e. according to the subject matter.

- A movie recommender system which makes such predictions has been developed by the group eliteAI (team leaders: Damien McMonigal and James Clark).

- Harry Mak, Honors student 2002 (co-supervision with Dr. Josiah Poon) is exploring the potential of learning from text (e.g. movie description, review, etc.) to movie recommendation.
One of the project goals is to compare the three approaches (content-based prediction based on limited set of feature, content-based prediction based on text and collaborative filtering).

E-mail Classification

- Electronic e-mail is rapidly becoming one of the most popular, fastest and cheapest means of communication. As a result, the volume of e-mail that we get is constantly growing. This is one aspect of a more general information overload problem faced by many people. People are spending more and more time and effort filtering e-mails and organizing them into folders.

- Most modern e-mail software packages provide some form of programmable automatic filtering, typically in the form of sets of rules that organize mail into folders or dispose of junk mail based on keywords detected in the headers or message body. Unfortunately manually constructing a set of robust rules is an arcane task. Moreover, users are constantly creating, deleting and reorganizing their folders. A system that can automatically learn how to classify e-mail messages into a set of folders is highly desirable.

- The aim of this project is to investigate methods for automatic classification of e-mail messages. In addition to some standard machine learning approaches for classification, the project aims to explore the potential of neural networks and genetic algorithms.

- Honors student: James Clark, co-supervision with Dr. Josiah Poon

Comparison of Text Categorization Techniques

- More and more documents are available on the web. Effective retrieval requires good indexing and summarization of the content of the documents. Text categorization is one solution to this problem. It is defined as automatically assigning predefined categories to text documents. Many statistical and machine learning techniques have been applied to text categorization. Research also shows that combining various classifiers typically improves the performance.

- The goal of this project is to compare techniques for text categorization, to study their advantages and limitations for the different situations, to develop methods which integrate multiple classifiers and to evaluate their performance.

- Honors student 2002: Elisabeth Crawford, co-supervision with Prof. Jon Patrick

Meta-learning

- Meta-learning refers to learning to learn. Given a set of tasks, training experience for each of them and performance measures, an algorithm is said to learn to learn if its performance on each task improves with experience and with the number of tasks.

- PhD Student: Daren Ler, co-supervision with A/Prof. Judy Kay and Dr. Eric McCreath (ANU)

Machine Learning for the Smart Personal Assistant

- Smart personal assistants should be able to learn from experience and adapt. This project aims at investigating a number of machine learning techniques for design of adaptive personal assistant architectures.

- Supported by the CRC Smart Internet

- Collaborators: Prof. John Lloyd (ANU, project leader), Dr. Eric McCreath (ANU), Dr. Abdul Sattar (Griffith) and Dr. Josiah Poon.

Contacts

Dr Irena Koprinska
 
University of SydneyDesigned by eliu