Honours Projects
This is a collection of honours projects being offered in 2009. These research projects are closely tied with work being undertaken on the Thinking Systems Project. For more information about these projects or to apply for a project please contact Dr Daniel Angus (d (dot) angus (at) uq (dot) edu (dot) au)
Visualisation Techniques for High Dimensional Data
Background
The recent growth of digital media, in large part due to the expansion of the internet, has allowed increased access to information. However the gap between knowledge and information still remains. To close this gap many systems have been proposed to assist in gaining understanding of these many varied information sources. These concept mapping tools have allowed users to gain insights into data in ways that were previously impractical or impossible to obtain.
Computational techniques currently used to visualise these multi-dimensional datasets (in particular text documents) have been developed using input from many diverse academic backgrounds, including but not limited to: linguistics, psychology, physics and mathematics. Many of the existing techniques are good at achieving the goal of reducing large dimensional data to few dimensions, however most use particular assumptions about the data organisation and content or require human supervision in the development of their representations.
ConceptSLAM is a newly proposed unsupervised knowledge representation and visualisation system. ConceptSLAM extends the physical navigation system, RatSLAM, which uses visual input to perform physical navigation tasks including self localisation, terrain mapping and route planning. RatSLAM takes inspiration from the mammalian hippocampus which is important in performing physical navigation tasks. The ConceptSLAM system represents text documents as pathways in a conceptual space, similar to how RatSLAM maps pathways (roads, office corridors) from the physical space.
Project
Both of the RatSLAM and ConceptSLAM systems represent input information (visual scenes and text sentences respectively) as individual experiences that are connected to other similar experiences according to their physical or conceptual organisation. In the physical domain expriences are organised on a 2D plane, thus making visualisation relatively simple. In the concept domain the intra-concept relationships exist on a multi-dimensional plane, making it difficult to preserve these relationships with a visualisation system.
This project will require the student to develop a visualisation system that can preserve interesting conceptual linkages, while allowing for easy inspection by an end user. These are conflicting goals so any solution will likely have to be a trade-off thus testing the student's judgement and creativity. There are many existing mapping techniques in the literature which may be useful as starting points for this investigation.
Skills
This project would suit a student with a background in Computer Science, Software Engineering, Multimedia, or similar degree program. Strong programming skills are essential, as well as a basic understanding of information visualisation techniques (Principle Componenet Analysis, Correspondance Analysis). Students should be self-motivated and be able to work with minimal supervision. Good teamwork skills are also desirable.
Investigating the use of Bayes Theorem for Text Classification Tasks
Background
The recent growth of digital media, in large part due to the expansion of the internet, has allowed increased access to information. However the gap between knowledge and information still remains. To close this gap many systems have been proposed to assist in gaining understanding of these many varied information sources. These concept mapping tools have allowed users to gain insights into data in ways that were previously impractical or impossible to obtain.
ConceptSLAM is a newly proposed unsupervised knowledge representation and visualisation system. ConceptSLAM extends the physical navigation system, RatSLAM, which uses visual input to perform physical navigation tasks including self localisation, terrain mapping and route planning. RatSLAM takes inspiration from the mammalian hippocampus which is important in performing physical navigation tasks. The ConceptSLAM system represents text documents as pathways in a conceptual space, similar to how RatSLAM maps pathways (roads, office corridors) from the physical space.
Central to the design of these concept mapping tools are lexical statistics which obtain information about the content of input data. These lexical statistic techniques are generally used to obtain rich information about input data such as what words are good descriptors to describe a large document, and how thematic content is arranged within the document. One particularly simple, yet powerful, statistic is word co-occurrence data. This data counts the relative frequency of particular words occurring with other words.
Project
Although co-occurrence data is easy to obtain, its use in determining sentence conceptual content is less than clear. How to combine term co-occurrence data to obtain good estimates of conceptual content is an open research problem. Some ideas exist in the literature, and these use techniques such as Bayes theorem to achieve this concept estimation. However such estimations must strike a balance between the computational complexity and accuracy of the the end result.
Where such a system fits into the wider research of ConceptSLAM is in translating input text data into generalised conceptual information. In the physical domain visual scenes contain information which is condensed into pixel intensity data for processing by RatSLAM. This pixel data is useful for obtaining information about the speed of the robotic agent and its rotation. In the conceptual domain, text documents are somewhat different as there are many ways to consider how the passage of text through the ConceptSLAM system could translate to movement. By translating text data into a conceptual `code' we hope to obtain information about the passage of a document through the conceptual space.
Skills
This project would suit a student with a background in Computer Science, Software Engineering, or a similar degree program. Strong programming skills are essential, as well as a basic understanding of probability theory. Students should be self-motivated and be able to work with minimal supervision. Good teamwork skills are also desirable.