Our School runs a research seminar series to which all interested parties are invited to attend.

The usual seminar time is 11am on Friday mornings in Room 1.24.

1. Visual tracking and SLAM at sea
2. A computational investigation of bone biology
3. Fluctuation-induced forces in and out of equilibrium
4. Heuristic RNA pseudoknot detection in long sequences based on stem-loop correlated energy modelling.
5. Bringing engineering to life
6. Tracking mobile targets using energy-constrained sensor networks
7. New insights into therapeutic drug interventions for catabolic bone diseases using an in-silico modelling approach
8. Identifying conceptual similarities using distributed symbols and backpropagation-to-representation
9. Information fusion techniques in multimodal biometric systems
10. Toxicity evidence integration

Visual tracking and SLAM at sea

• 14 August - 2:30pm
• Ian Reid

Abstract

I will discuss work with my DPhil student Charles Bibby, which considers the problem of enhancing the navigational capabilities, the collision avoidance and security systems, and the search and rescue capabilities of ships and other sea-going vessels.

Time-permitting, I will describe three novel algorithms which we have developed to this end: (i) SLAMIDE is a SLAM algorithm based on a recursive sliding window filter which allows reversible data-association decisions, and motion-model selection, allowing the incorporation of dynamic objects into the environment map. We show that in simulation this algorithm achieves excellent robustness to poor initial data-association (comparable to JCBB) but is also very robust in a cluttered environment, as well as being able to handle a high proportion of dynamic objects in the map. We also show results from challenging real radar data in a busy harbour. (ii) We extend the idea of SLAMIDE by incorporating an occupancy grid to model landmasses, and by introducing the idea of representing the trajectories of both the ego-motion and of dynamic objects in the map using splines. This has the benefit of greatly reducing the size of the map. Further, by replacing a discrete representation of the trajectory with a continuous one, we can now allow for completely asynchronous measurements. (iii) I will describe a new visual tracking and segmentation algorithm, based on the evolution of a level-set contour. The key idea that differentiates this tracker from other similar ones is that we marginalise out the uncertain foreground/background membership probabilities of all pixels in the image and solve only for the two variables of interest, namely pose and shape of the target. The algorithm is fast, requiring less than 50ms per frame on even modest hardware, robust to illumination and shape changes, and can track agile motion. We have deployed this algorithm to control a high-performance custom-built PTZ platform and recently begun trials at sea.

The work is sponsored by Servowatch Systems Ltd.

About the speaker

Ian Reid is a Reader in Engineering Science and Fellow of Exeter College, at the University of Oxford where he jointly heads the Active Vision Group. He obtained BSc from the University of Western Australia in 1987, and came to Oxford University on a Rhodes Scholarship in 1988 where he completed a D.Phil. in 1991. His research has touched on many aspects of computer vision, concentrating on algorithms for visual tracking, control of active head/eye robotic platforms (for surveillance and navigation), SLAM, visual geometry, novel view synthesis and human motion capture.  He has published over 100 papers on these and related topics. He serves on the editorial boards of Image and Vision Computing Journal and the recently formed IPSJ Transactions on Computer Vision Applications.

Back to top

---

A computational investigation of bone biology

• 7 August - 11:00am
• Devin Sullivan

Abstract

The bone cycle is a continuous and dynamic system in which old or damaged bone is constantly removed and replaced by new bone. This system consists of two basic cell types, osteoclasts and osteoblasts. The osteoclasts are generally speaking responsible for the catabolic effect of bone resorption while the osteoblasts cause bone formation, an anabolic process. The coordination between these two cell types is crucial in maintaining appropriate strength in one’s bones. This system works in highly coordinated groups called basic multicellular units or BMUs.

This work is a summary of my 3 month research scholarship visiting UWA. During that time I had the opportunity to work on a variety of problems related to bone regulation ranging from in-vitro to in-vivo applications. Despite the obvious and extensive self-regulation of bone tissue, biologists have been able to do little to understand the complexities of these cell-cell interactions. Several computational models attempt to shed light on these interactions, covering a large scope of biological queries related to the bone system. This presentation aims to highlight a few of these models and their potential applications. The first model will highlight the balance in differentiation of mesenchymal stem cells into osteoblasts versus adipocytes. The next model investigates the effects of parathyroid hormone on overall bone activity and resorption. Lastly, a spatial model of a single BMU demonstrates the spatial movement of these cells as they interact. The discussion will focus on the potential therapeutic applications of each model. Of particular interest will be the predictive nature of the models beyond that of traditional biological approaches.

Back to top

---

Fluctuation-induced forces in and out of equilibrium

• 24 July - 11:00am
• Pascal Buenzli

Abstract

Fluctuation-induced forces are forces arising between objects due to the fluctuations of their surrounding media. Their presence reveals the incessant jiggling motion that otherwise averages into what seems still. Probably the best-known example of such a force is the London--van der Waals force between neutral atoms in a fluid. The atoms' fluctuating dipole moments average zero individually, nevertheless produce a nonvanishing attraction between them. In 1948, H. B. G. Casimir predicted the existence of a similar attractive force between two neutral metallic plates due to the quantum fluctuations of the electromagnetic field. This so-called "Casimir force", aside from thrilling SF, has been the subject of extensive research since the late '90s when experimental setups were able to probe its strength. Many other fluctuation-induced phenomena have since entered the field.

This talk will overview two distinct studies performed in this field. The first concerns a microscopic (bottom-up) approach to understand the Casimir force between metallic plates. This calculation resolves a long-lasting and heavily-debated controversy about the value of the force at large separations, that dates back in works by E. M. Lifshitz and J. Schwinger. The second topic studies various interesting properties that fluctuation-induced forces acquire when they arise out of nonequilibrium fluctuations. In particular, forces and torques can be induced on single (asymmetric) objects and they can be tuned both in strength and sign. These properties are illustrated between various objects immersed in a basic reaction-diffusion fluid and calculations compared with a Boundary Element Method (BEM) scheme.

Back to top

---

Heuristic RNA pseudoknot detection in long sequences based on stem-loop correlated energy modelling.

• 17 July - 2:00pm
• Jana Sperschneider

Abstract

There are two types of nucleic acids in the living cell: deoxyribonucleic acid (DNA) and ribonucleic acid (RNA). RNA is a versatile macromolecule which is no longer seen as the passive intermediate between DNA and proteins. Numerous functional RNA molecules with an astonishing variety have been uncovered in the last decade. Macromolecule function is closely connected to its three-dimensional folding and therefore, structure prediction from the base sequence is of great importance.

When only a single sequence is given, the most popular approach for RNA structure prediction is free energy minimization using dynamic programming. However, the inability to predict crossing structure elements, so-called pseudoknots, is a major drawback. Pseudoknots are functional structure elements which occur in most classes of RNA and in many viruses. From a theoretical point of view, general pseudoknot prediction is not  an easy task and has been shown to constitute an NP-complete problem. In general, most practical methods as reported in the literature suffer  from low accuracy for longer sequences and high running times.

This talk will give an introduction to a different algorithmic framework which aims to detect pseudoknots in a sequence with high confidence. The importance of the underlying folding model for pseudoknots will be discussed and recent progress in pseudoknot structure prediction will be presented.

Back to top

---

Bringing engineering to life

• 19 June - 11:00am
• Bruce Gardiner

Abstract

A new research group, calling themselves engineering computational biology, has recently joined the School of Computer Science and Software engineering. The two main aims of this seminar are to give an introductory overview of the research conducted by this group and to invite interactions with other members of CSSE. In the past 5 or so years the engineering computational groups has been developing mechanistic mathematical/computational models of biological systems at the tissue, cellular and sub-cellular levels.

Current projects include: Intercellular communication in bone regulation, developmental biology and prostate cancer progression. Intracellular communication in colorectal cancer, cell adhesion and programmed cell death. Cell-tissue interactions regulating cartilage health and kidney oxygenation.

Finally clinical applications such as the design of glaucoma drainage implants and strategies to predict preterm birth. Some of these projects will be discussed to illustrate the different aspects of this research and what this engineering approach may offer to the biomedical community.

Back to top

---

Tracking mobile targets using energy-constrained sensor networks

• 25 May - 11:00am
• Prof Bijendra Jain

Abstract

In this talk we first give a general overview of sensor networks and their applications. But for most part of the talk we consider the problem of tracking mobile targets using energy-constrained sensor networks.

In particular, we consider the problem of estimating the location of a moving target ‘T’ in a 2D plane. We assume that it is possible for sensors to detect the presence of the target in its vicinity and to (possibly) measure the distance from/to the target. Given that available energy in sensors is at a premium, we have proposed protocols for target detection and route activation that require sensors to conserve energy by switching between ‘inactive’ and ‘active’ modes of operations, while waking-up frequently in inactive mode to evaluate the need to become active. Yet another method to save energy is to reduce the number of measurements and, as a result, the number of transmissions. We therefore propose that energy be conserved by (a) requiring that a sensor switch to an ‘inactive’ mode whenever feasible, and (b) selecting fewer but adequate number of  sensors that measure distance and communicate with the central tracker. Given an adequate spread of sensors, and an ability to only detect  presence or absence of the target within its vicinity in a timely manner, it is feasible to obtain an approximate trajectory of a mobile target as a function of time. Alternatively, distance measurements from several such sensors may be used to estimate the location of the target T at a point  of time. Clearly, the latter approach is expected to track the target more accurately. However, it will necessarily require that there be at least three sensors within the vicinity of the target and, therefore, this approach requires a significantly greater density of sensors. The error in  estimating location of the target using distance measurements from multiple sensors is shown to be dependent on two measures viz. proximity of  sensors to the target, and co-linearity of sensors. We also propose a new measure, ideal direction for selecting a 3rd sensor, given locations of two sensors and the location of the target. We propose algorithms to estimate the track of the target by using distance measurements from sensors selected on the basis of the above measures. Further, we evaluate all the protocols and algorithms using simulations.

About the speaker

Professor Bijendra Jain obtained B. Tech. from IIT Kanpur in 1970, and Ph. D. from SUNY, Stony Brook in 1975, both in Electrical Engg. Since 1975 he has been with IIT Delhi, where he is presently Deputy Director (Faculty) and Professor of Computer Science. In the past he has held visiting assignments with Universities of Texas and Maryland, Bell Labs, and Cisco Systems. Presently, he is a Gledden Senior Visiting Fellow at University of Western Australia.

His interest is in Computer Networks and Systems, including network models and analysis, algorithms for large sparse matrix operations, scheduling algorithms for hard real-time systems, fault-tolerant routing. His recent interest is, however, in ad hoc and sensor networks. His research is funded in part by Government of India, UNDP, US Army, Sun Microsystems, Microsoft and Media Lab Asia. As early as 1989, Prof. Jain, together with developers from other institutions in India, built and launched India’s first data network, ERNet. Today, ERNet is a thriving not-for-profit company connecting over a million users spread across over 2000 institutions in India. He is a co-inventor in seven US patents, assigned to Cisco Systems. These cover methods to speed-up access to Web pages and efficient monitoring of IP network performance. He has co-authored "OSI: Its Architecture and Protocols", a book published by McGraw Hill, New York.

He is an active industry consultant. He is also a member of several Government committees, including Naval Research Board. Lastly, Professor Jain is a co-founder and past-Chairman of Kritikal Solutions, a technology start-up incubated on IIT Delhi campus and one which is focused on computer vision, embedded systems and networks.

Back to top

---

New insights into therapeutic drug interventions for catabolic bone diseases using an in-silico modelling approach

• 1 May - 11:00am
• Peter Pivonka

Abstract

The conceptual model employed by bone scientists today is based on the dynamic balance between two main bone cell types, the osteoclasts and osteoblasts, which continuously resorb bone and form new bone. This process is referred to as bone remodelling. This conceptual approach has given many new insights into how bone structure and function is influenced by a variety of factors including hormones, cytokines, mechanical loading, and gene mutations to name only a few. But these two cell types do not work independently, but rather work in a coordinated way in so-called basic multi-cellular units (BMUs). There is cross-talk between the cell types to coordinate their functional behaviours. Given this interaction, it is very difficult to predict what might happen given some changes of the bone microenvironment. To date there have been few attempts to integrate all the key observations into a theoretical framework, which allows theoretical predictions and subsequent investigation experimentally. Mathematical modelling, provides the basis for "translation" of conceptual models into theoretical models which can then be employed to quantitatively investigate various hypotheses and study system behaviour as a whole rather than single-component behaviour. This presentation will discuss how in-silico modelling can be applied to advance our current knowledge on bone biology with particular emphasis on therapeutic interventions in catabolic bone diseases.

Back to top

---

Identifying conceptual similarities using distributed symbols and backpropagation-to-representation

• 24 April - 11:00am
• Peter Dreisiger

Abstract

The tasks of concept formation and concept recognition underlie our ability to generalise knowledge and cope with uncertainty; these tasks, however, also assume that we are able to identify shared traits and conceptual similarities. While researchers within the cognitive sciences continue to study our ability to form concepts and estimate similarities, the growing need for data mining and analysis tools has led to the development of more heuristic forms of statistical pattern recognition and conceptual clustering; some of these techniques have even drawn upon advances in the area of dimensional reduction.

Non-linear Principal Component Analysis, self-organising feature maps and the auto-encoder neural network allow us to discover new trends in numerical data; while these techniques are able to discover many hidden patterns and relationships, they are restricted to the analysis of entities, or observations, that have a natural, metrical representation. For symbolic data, there are several forms of co-occurrence and correlation analysis, and techniques such as Latent Semantic Analysis (LSA) have been developed to discover similarities in textual documents; LSA, in particular, can discover implicit relationships based upon the terms' shared neighbourhoods. What these techniques lack, however, is the ability to capture syntax and usage.

One approach which has received very little attention from the data mining community is a variant of the auto-encoder that uses an extended form of back-propagation to capture regularities in the properties and roles of terms and other, non-numerical entities. While this technique, called FGREP, has been used to model the process of concept formation in humans, it has not, as yet, been used to discover patterns and conceptual clusters in larger sets of data; nor has it been used within the intelligent agent community to classify or group entities based upon their perceivable attributes.

The aims of this project, then, are three-fold: (1) to investigate FGREP's ability to discover meaningful concepts in larger sets of symbolic data; (2) to identify its limitations and investigate how we can improve its average performance; and (3) to compare it to other existing techniques. In this talk, we will introduce FGREP and our implementation of this technique, we will present the results of our first set of experiments, and we will see how its conceptual clusters compare to those found using LSA.

Back to top

---

Information fusion techniques in multimodal biometric systems

• 17 April - 11:00am
• Maryam Mehdizadeh

Abstract

Most biometric systems deployed in real-world applications are unimodal, i.e., they rely on the evidence of a single source of information for authentication (e.g., single fingerprint or face). These systems have to contend with a variety of problems such as: (a) Noise in sensed data: A fingerprint image with a scar, or a voice sample altered by cold are examples of noisy data. Multibiometrics is expected to overcome some of the limitations of unibiometric systems by combining the evidence presented by multiple biometric sources. This integration of information is known as information fusion and if appropriately done, can enhance the matching accuracy, increase population coverage and deter spoofing activities.

The aim of this research is to investigate how methods of semi-supervised learning and non-linear dimensionality reduction can be used in multimodal biometric systems for the purpose of feature fusion. We are intending to explore how feature selection and feature fusion can be a part of learning algorithm and be automated as much as possible. We are going to learn the feature spaces of face and hand biometric data with Locally Linear Embedding (LLE) - as a nonlinear feature descriptor - and consolidate (fuse) the two spaces in the concept of adaptive feature fusion with semi-supervised learning methods.

Back to top

---

Toxicity evidence integration

• 3 April - 11:00am
• Alison Anderson

Abstract

The environment is increasingly contaminated with thousands of chemical compounds the health effects of which are not well understood. Toxicity arises from complex interactions between environmental agents, genes and the mechanisms that control gene expression, collectively known as the epigenome. Advances in bioinformatics are driving new approaches to investigating and evaluating these complex toxicity pathways. Fundamental to these objectives is the need to integrate information, in particular, evidence of chemical toxicities from disparate biological resources. For this project information from disparate publicly available resources has been downloaded and serialised as triple graphs which comprise three components (subject, predicate, object) or (entity, attribute, value). An ontology-driven method has been developed to extract toxicity specific subgraphs of information from the aggregated triplestore. Toxicity information is mapped to the ontology via N3Logic, a W3C specification which allows rules to be expressed in a Web environment. This process provides a framework that facilitates the integration and manipulation of toxicity evidence arising from multi-disciplinary research. It is the first step in meeting the main objective of this PhD: to provide a holistic weight-of-evidence approach to assessing toxicity evidence.

Back to top