December 2009
- Wednesday, 16 December 2009, 1-2pm*
Host: Taras Kowaliw
Title: An Empirical Exploration of a Definition of Creative Novelty for Generative Art
Abstract: We explore a new definition of creativity — one which emphasizes the statistical capacity of a system to generate previously unseen patterns — and discuss motivations for this perspective in the context of machine learning. We show the definition to be computationally tractable, and apply it to the domain of generative art, utilizing a collection of features drawn from image processing. We next utilize our model of creativity in an interactive evolutionary art task, that of generating biomorphs. An individual biomorph is considered a potentially creative system by considering its capacity to generate novel children. We consider the creativity of biomorphs discovered via interactive evolution, via our creativity measure, and as a control, via totally random generation. It is shown that both the former methods find individuals deemed creative by our measure; Further, we argue that several of the discovered "creative" individuals are novel in a human-understandable way. We conclude that our creativity measure has the capacity to aid in user-guided evolutionary tasks.
- Thursday, 3 December 2009, 1-2pm*
Host: Marsha Minchenko
Title: Counting closed walks in regular graphs
Abstract: Using a series of equations, we can relate the eigenvalues of a graph to the number of vertices, edges, and 3-cycles by counting the number of closed walks in two different ways. This information has been used to help in the search for members of interesting families of graphs. One such graph family is the integral graphs. These have only integers as eigenvalues when the adjacency matrix of the graph is considered. This talk will develop these endings, and report on work in progress to extend them by considering counting closed walks about subgraphs other than 3-cycles.
November 2009
- Tuesday, 24 November 2009, 1-2pm [Seminar room 135/26] *
Host: Tom Mitchell (Carnegie Mellon University)
CRIS Week Postgrad Q&A Session (in conjunction with CRIS Week)
Tom is Fredkin Professor of AI and Machine Learning & Chair of the Machine Learning Department, School of Computer Science, Carnegie Mellon University.
His research includes machine learning to analyze human brain activity (fMRI) and never-ending language learning research.
Everyone is invited to a question and answer session with Tom to learn more about his research, the state of the art in machine and language learning, and what life as a leading researcher is like.
- Wednesday, 18 November 2009, 1-2pm*
Hosts: Julie Bernal and Ben Porter
Title: ACAL Talks
Abstracts:
Julie: Despite the existence of vaccines, the Hepatitis B virus (HBV) is still a serious global health concern. HBV targets liver cells. It has an unusual replication process involving an RNA pre-genome that the reverse transcriptase domain of the viral polymerase protein transcribes into viral DNA. The reverse transcription process is error prone and together with the high replication rates of the virus, allows the virus to exist as a heterogeneous population of mutants, known as a quasispecies, that can adapt and become resistant to antiviral therapy. This study presents an individual-based model of HBV inside an artificial liver, and associated blood serum, undergoing antiviral therapy. This model aims to provide insights into the evolution of the HBV quasispecies and the individual contribution of HBV mutations in the outcome of therapy.
Ben: Modelling the geometry of organic forms using traditional CAD or animation tools is often difficult and tedious. Different models of morphogenesis have been successfully applied to this problem; however many kinds of organic shape still pose difficulty. This paper introduces a novel system, the Simplicial Developmental System (SDS), which simulates morphogenetic and physical processes in order to generate specific organic forms. SDS models a system of cells as a dynamic simplicial complex in two or three dimensions that is governed by physical rules. Through growth, division, and movement, the cells transform the geometric and physical representations of the form. The actions of the cells are governed by conditional rules and communication between cells is supported with a continuous morphogen model. Results are presented in which simple organic forms are grown using a model inspired by limb bud development in chick embryos. These results are discussed in the context of using SDS as a creative system.
October 2009
- Tuesday, 27 October 2009, 1-2pm*
Host: Marc Cheong
Title: Integrating Web-Based Intelligence Retrieval and Decision-Making from the Twitter Trends Knowledge Base
Abstract: Twitter as a microblogging platform has vast potential to become a collective source of intelligence that can be used to obtain opinions, ideas, facts, and sentiments. This paper addresses the issue on collective intelligence retrieval with activated knowledge-base decision making. Our methodology differs from the existing literature in the sense that we are doing analysis on Twitter microblog messages as opposed to traditional blog analysis in the literature which deals with the conventional ‘blogosphere’. Another key difference in our methodology is that we apply visualization techniques in conjunction with artificial intelligence-based data mining methods to classify messages dealing with the trend topic. Our methodology also analyzes demographics of the authors of such Twitter messages and attempt to map a Twitter trend into what's going on in the real world. Our findings reveal a pattern behind trends on Twitter, enabling us to see how it ‘ticks’ and evolves though visualization methods. Our findings also enable us to understand the underlying characteristics behind the ‘trend setters’, providing us a new perspective on the contributors of a trend.
- Wednesday, 21 October 2009, 1-2pm*
Host: Alex Healing (Senior Researcher at BT)
Title: Structuring Folksonomic Systems using Adaptive Visual Clustering
Abstract: With the explosion of unstructured information in the last decade where everyone is a publisher but no one has time to create metadata, tags as a means of marking up data has proven to be a success, in terms of their ubiquity at least, but how far do so-called folksonomic systems based on collaborative tagging scale? When mining a large corpus for as much relevant information, how can inconsistencies between tags be accounted for without assuming too much about the dataset in advance, or mandating a large amount of a priori engineering effort? How can the world of the unstructured be exploited in highly structured environments within organisations, and how can this be done in an adaptive and tailored manner? This talk will touch the surface of these questions through the discussion of a prototype visual data mining system which can be used to create taxonomies from the data-up whilst keeping the user in the loop, and (machine) learning along the way.
- Thursday, 15 October 2009, 1-2pm*
Host: Sarah Boyd
Title: Computational Modelling and Prediction of Protease Action
Abstract: Proteases are the class of enzymes that act on proteins. Consequently, in biological terms, they are essential for life, and not surprisingly they are found in all organisms. If they malfunction, they cause disease, pathogens use them to invade and subvert the host's normal biological processes, and they are important in many biotechnological and industrial applications, such as the production of detergents, leather and dairy products. I have been working on the problem of computational modelling and prediction of protease action for the last 10 years. In this talk, I will explain the whys and wherefores of this journey, where it taken myself and my collaborators, and where to next.
September 2009
- Wednesday, 30 September 2009, 1-2pm*
Host: Michael Niemann
Title: Expert Identification within a Verbose Online Community
Abstract: Expertise finding recommender systems endeavour to help organisations and individuals locate people who are experts in particular fields. They form expertise profiles from documents that are related to the experts. However, no existing system examines an expert's contributions to a community, nor the community's opinion of the contribution. This talk will discuss how my research aims to remedy this. My research will investigate how an expert's postings to an online discussion forum can be used to identify their areas of expertise, through a linguistic analysis of the text and the identification of semantic relationships between the terms used in the postings. The community's replies within the forum will also be examined to determine the community's opinion of the experts. This will be done through further linguistic analysis and by identifying the intent behind each reply, e.g., support and rebuttal. Both of these models will contribute to an expertise profile for each expert.
- Thursday, 24 September 2009, 1-2pm***
Host: Asanka Fonseka
Title: A Cognitive Approach for Multimodal Information Fusion for Bioinformatics Applications
Abstract: In this talk I will give an introduction to my research topic and then demonstrate a novel integrated hierarchical information fusion algorithm for clustering diverse genomic data sources. Developments of high-throughput technologies and genome sequencing projects have resulted in a shift of focus from single component analysis to large-scale global level analysis. There is a great demand for integrative analysis, where knowledge about a biological system is derived by information fusion, using heterogeneous data sources. Several technologies and data sources provide different types of information about the biological system, and fusion of heterogeneous data sources is therefore of great importance in the efforts to understand a particular biological process. For example, combining gene expression analysis with gene ontology, localization information, and promoter sequences may improve the final outcome compared to the solely using expression data. The new integrated hierarchical approach for clustering multi-source genomic data provides a sophisticated framework to reveal underlying biological patterns by combining heterogeneous data sources. In order to evaluate the performance of the proposed algorithm gene expression data and known transcription factor binding sites frequencies were clustered and experiments consistently showed that better clusters were obtained when applying collaborative learning.
- Wednesday, 9 September 2009, 1-2pm*
Host: Kerri Morgan
Title: Galois Groups of Chromatic Polynomials
Abstract: The chromatic polynomial, P(G,x), gives the number of proper colourings of a graph in at most x colours. Our research considers possible links between the Galois groups of these polynomials and the structure of their graphs. Two graphs are Galois equivalent if they have chromatic polynomials with the same Galois group. Chordal graphs are Galois equivalent as their chromatic polynomials have the trivial Galois group. We are interested in other families of Galois equivalent graphs. First we give some operations that can be performed on a graph to produce Galois equivalent graphs. We then give an infinite family of Galois equivalent graphs that have chromatic polynomials with Galois group D(4).
- Thursday, 3 September 2009, 1-2pm***
Host: Greg Paperin
Title: Towards Formalising the Theory of Dual Phase Evolution
Abstract: This is a further talk on Dual Phase Evolution (DPE), and parts of it will be familiar to those how attended my previous seminars. I intend to use it as training for the presentations I will be giving at the European Conference for Artificial Life and the International Conference on Evolutionary Computation next month. Evidence from several fields suggests that DPE may account for distinctive features associated with complex adaptive systems. I will review empirical and theoretical evidence for DPE in natural systems and examine the relationship of DPE to self-organised criticality and adaptive cycles.
August 2009
- Wednesday, 19 August 2009, 1-2pm*
Host: Ingrid Zukerman
Title: A Spoken Language Interpretation Component for a Robot Dialogue System
Abstract: The DORIS project aims to develop a spoken dialogue module for an autonomous robotic agent. In this talk, I will examine the techniques used by Scusi?, the speech interpretation component of DORIS, to postulate and assess hypotheses regarding the meaning of a spoken utterance, including our formalism for disambiguating referring expressions. I will also present the results of our evaluation experiments, and discuss our progress towards the interpretation of multiple utterances.
- Wednesday, 12 August 2009, 1-2pm
Host: Jenny Kashmirian
Title: Modelling Self-Organising Chemical Networks
Abstract: The complexity of nature is often manifested as some type of network. A fundamental network at the molecular level is the hydrogen bond network formed by water molecules. This network drives the hydrophobic effect and is responsible for membrane formation, catalytic activity and organization within a cell. One feature of chemical networks is that once a critical point of connectivity is reached they self-organise. The question is: Can we make software networks self-organise like chemical networks, and would it be useful? The assumptions of this research are: 1. If we know the connectivity rules of chemical networks, then we can create software that self-organises like chemical networks; and 2. Self-organising software agents will be useful (maybe critical) to help us meet the challenge of large, distributed network environments. I will give a talk on the model developed to test the idea that self-organisation will emerge from modeling the connectivity rules of the network.
- Tuesday, 4 August 2009, 1-2pm***
Host: Jonathan Ito (University of Southern California)
Title: Self-Deceptive Decision Making (Normative and Descriptive Insights)
Abstract: Computational modeling of human belief maintenance and decision-making processes has become increasingly important for a wide range of applications. We present a framework for modeling the psychological phenomenon of self-deception in a decision-theoretic framework. Specifically, we model the self-deceptive behavior of wishful thinking as a psychological bias towards the belief in a particularly desirable situation or state. By leveraging the structures and axioms of Expected Utility (EU) Theory we are able to operationalize both the determination and the application of the desired belief state with respect to the decision-making process of expected utility maximization. While we categorize our framework as a descriptive model of human decision making, we show that in certain circumstances the realized expected utility of an action biased by wishful thinking can exceed that of an action motivated purely by the maximization of perceived expected utility. Finally, we show that our framework of self-deception and wishful thinking has the descriptive flexibility to account for the inconsistencies highlighted by the Common Ratio Effect and the Allais Paradox.
July 2009
- Thursday, 30 July 2009, 1-2pm*
Host: Xiaojiang Ding
Title: An Intelligent Model-Based Analysis for Investigating Smokers' Quitting Motivations and Tobacco Control Policies
Abstract: Tobacco consumption has long been a recognized global health problem. Various countries have attempted to influence the behaviour of smokers using means such as warning labels on cigarette packs, increasing the price of cigarettes and TV advertisements. Conventional statistical techniques have been used to analyse the effectiveness of tobacco control policies. With the assumption of linearity, these techniques have limitations in describing the complex, non-linear relationship between tobacco control policies and smokers’ behaviours. To address this issue, we have applied an intelligent model-based analysis using International Tobacco Control survey data to investigate smokers' quitting motivations and tobacco control policies.
- Friday, 24 July 2009, 1-2pm*
Host: Joshua Ho (University of Sydney)
Title: Concepts, Methods and Engineering Issues for Systems-Level Biological Data Analysis
Abstract: Computational analysis of systems-level data (such as sequence, expression microarray and network data) are arguably the central component of computational systems biology. In my PhD thesis, I examine some concepts, methods and software engineering issues related to the analysis of systems-level data in biological and biomedical studies. The central concepts and challenges in computational systems biology are explained. Novel methods are developed to tackle specific biological problems, ranging from understanding the population gene expression dynamics in human diseases and ageing, to the evolution of gene regulatory networks. In the second half of the talk, I will touch on the software engineering issues in testing bioinformatics software.
- Thursday, 16 July 2009, 1-2pm*
Host: Gerhardus Visser
Title: Interest-Guided Bayesian Learning
Abstract: With Bayesian inductive inference one attempts to select a hypothesis which best explains the observed data. In practice one is not always interested in discovering all regularities in the observed data. This is because of limited interest in the data. Our aim is to develop a theoretical framework for Bayesian inductive inference (and more specifically for minimum message length) in which one tells the learning program what your interest is and it then works out how to use that to select the best explanation. This talk will discuss how interest and relevance can be linked to the minimum message length principle and to Occam's razor.
- Friday, 10 July 2009, 1-2pm*
Host: Jeewanee Bamunusinghe
Title: Investigating the Effect of Illumination and Viewpoint on Image Recognisability
Abstract: Artificial intelligence (AI) is an area of computer science focusing on creating machines with human like intelligence.In order to achieve intelligence, the machine should have the ability to learn and understand. The artificial systems can learn from many different sources of information. The images are one of them. Therefore having an artificial vision system which can do self learning and understanding will increase the intelligence of those systems. In my research I am investigating how human learn and understand from their visual inputs and propose a conceptual model for artificial learning and understanding based on visual inputs. In this talk I will explain a biologically inspired model for achieving understanding from images and discuss how an artificial system can separate the images based on their view point and illumination condition with out human intervention.
- Wednesday, 1 July 2009, 1-2pm*
Host: Nayyar Zaidi
Title: Feature Fusion for Generic Object Recognition
Abstract: Object recognition is a very challenging problem in computer vision. Due to huge intra-cass and inter-class variation among classes, the idea of creating a viable object recognition system using a single set of features is not very attractive. Current research is focusing more on how to combine different cues (features) to make recognition more sturdy. In this talk we will highlight the study of our framework of feature fusion for object recognition in a systematic and intuitive way. Our main contribution is the proposal of a framework for distance measurement in high dimensional feature spaces. The main insight that drove our research is: different machine learning algorithms (if analyzed from the right perspective) are distance measuring devices in high dimensional feature spaces. By formulating feature fusion as a distance measurement problem we make explicit three levels of metric learning in our framework (aimed at dimensionality reduction, introducing isotropy and optimal feature subset fusion respectively). Using our proposed framework we improve the performance of the k-nearest neighbor classifier. We show our results on a well-knowm object recognition database: Oxford flowers.
|
