Professor Stephan Chalup

Because of the number of different architectures, numerous settings of their hyper-parameters and disparity among their sizes, it is difficult to equitably compare various deep neural network (DNN) architectures for sentence classification. Evolutionary algorithms are emerging as a popular method for the automatic selection of architectures and hyperparameters for DNNs whose generalisation performance is heavily impacted by such settings. Most of the work in this area is done in the image domain, leaving text analysis, another prominent application domain of deep learning, largely absent. Besides, literature presents conflicting claims regarding the superiority of one DNN architecture over others in the context of sentence classification. To address this issue, we propose a genetic algorithm (GA) for optimising the architectural and hyperparameter settings in different DNN types for sentence classification. To enable the representation of the wide variety of architectures and hyperparameters utilised in DNNs, we employed a generalised and flexible encoding scheme in our GA. Our study involves optimising two convolutional and three recurrent architectures to ensure a fair and unbiased evaluation of their performance. Furthermore, we explore the effects of using F1 score versus accuracy as a performance metric during evolutionary optimisation of those architectures. Our results, using ten datasets, show that, in general, the architectures and hyperparameters evolved using the F1 score tended to outperform those evolved using accuracy and in the case of CNN and BiLSTM the results were significant in statistical measures. Of the five architectures considered, the GA-evolved gated recurrent unit (GRU) performed the strongest overall, achieving good generalisation performance while using relatively few trainable parameters, establishing GRU as the preferred architecture for the sentence classification task. The optimised architectures exhibited comparable performance with the state-of-the-art, given the large difference in trainable parameters.

Conservation managers cannot manage what they don't know about, yet our existing biodiversity monitoring is idiosyncratic and small in scale. One of Australia's commitments to the Convention for Biological Diversity in 2015 was the creation of a national biodiversity monitoring programme. This has not yet occurred despite the urgent need to monitor common and threatened species, as highlighted by the challenges of determining the biodiversity impacts of the Black Summer fires of 2019/20. In light of improvements to automation, miniaturisation and powering devices, the world urgently needs to scale-up biodiversity monitoring to become coordinated, comprehensive and continuous across large scales. We propose the BIOMON project that could achieve this where individual sensor nodes use machine learning models to identify biodiversity via sound or photos onboard. This could be coupled with abiotic data on temperature and humidity, plus factors such as bushfire smoke. Nodes would be set within networks that transmit the results back to a central cloud repository where robust analyses are conducted and provided free to the public (along with the raw data). Network arrays could be set up across entire continents to measure the change in biodiversity. No one has achieved this yet, and significant challenges remain associated with training the algorithms, low power cellular network coverage, sensor power versus memory trade-offs, and sensor network placement. Much work is still needed to achieve these goals; however we are living in the 21st Century and such lofty goals cannot be achieved unless we start working towards them.

This study presents a new computer vision approach to perform affective analysis of a scene or object. The approach utilises a simulation of the phenomenon of face pareidolia that can be described as the perception of non-existent faces, for example, in random textures, clouds or rock formations. The emergence of face pareidolia in product designs and natural scenes can modulate affective perception of our everyday experiences without our conscious awareness. We propose a new deep learning method to simulate the face pareidolia ability of humans and predict associated emotional responses in two-dimensional valence and arousal space. Starting from a face detector that was trained on images of human faces, we propose a novel cross-domain weakly supervised three-step progressive domain adaptation approach to simulate face pareidolia by fine-tuning the human face detector on three types of synthetically generated sample images. Our approach fuses two deep network models, one model for predicting valence and arousal from abstract and minimal face-like patterns producing face pareidolia, and a second for recognising overall moods and context associated with the scene. To evaluate our approach, we constructed a new dataset containing instance-level annotations of face pareidolia occurrences as well as valence and arousal emotional values associated with the overall scene. The quantitative and qualitative experimental results of the present study demonstrate that our approach can outperform other state-of-the-art methods in face pareidolia detection as well as in predicting associated emotions.

This article presents a computer vision approach that can detect and classify abstract face-like patterns, including subliminal faces within a scene. This can be regarded as a way of simulating the phenomenon of pareidolia, that is, the tendency of humans to 'see faces' in random structures such as clouds or rocks. The paper describes the system consisting of a component-based face detector and an expression classifier. The face detector creates a number of component images from the original image at different resolutions. A component image is a binary edge image where the edges are segmented into components using a labelling method with a border-following technique. The component images are then overlaid to produce a component height map where large and notable components across all resolutions have high values, while specular and noisy components have low values. The method retains three-shape components, representing two eyes and a mouth, that have height map values that are larger than the noise cut-off value. Support vector machines using scale-invariant feature vectors are applied for ranking these three-shape components by their geometry and size, and their shape semblance to human faces in the training data. The outcome is a facial expression analysis system that uses face components, with the potential to estimate an emotional expression value for a scene by producing an array of emotion scores corresponding to Ekman's seven Universal Facial Expressions of Emotion. An advantage of this technique, when compared to a holistic method, is that the face components are explicitly isolated. This supports a process of abstraction that can facilitate the detection of distorted and minimal face-like patterns.

Time series clustering is an important data mining topic and a challenging task due to the sequences' potentially very complex structures. In the present study we experimentally investigate the combination of support vector clustering with a triangular alignment kernel by evaluating it on an artificial time series benchmark dataset. The experiments lead to meaningful segmentations of the data, thereby providing an example that clustering time series with specific kernels is possible without pre-processing of the data. We compare our approach and the results and learn that the clustering quality is competitive when compared to other approaches. © 2014 Published by Elsevier B.V.

The presence of locations that possess distinct spatial-cognitive features (salient landmarks) is a fundamental necessity for supporting navigation. Embedding formal or structural variability sufficient to create such landmark locations is therefore an important consideration in the design of large urban and architectural spaces. Despite the availability of diverse theories that seek to identify the characteristics of a salient landmark, relatively few experimental techniques are available to empirically evaluate saliency in a given architecture plan. This study is therefore motivated by the development of an ability to measure spatial distinctiveness during the architectural design and modelling process. The information from such an analysis can prove useful for evaluating the way in which a design provides support for wayfinding and spatial appeal. Statistical summaries obtained from the three-dimensional (3D) isovists are compared using principal component analysis to differentiate monotonous regions from the more structurally distinct ones. The experiments reported in the paper demonstrate novel utilization of the isovist concept to capture spatial properties and comparison of structural saliency among two well-known architectural designs. Central contributions of the paper include the novel experimentation technique of capturing and utilizing 3D isovists, its interpretation and the quantitative methodology behind saliency computation. © 2013 Copyright Taylor and Francis Group, LLC.

This research uses deep learning to estimate the topology of manifolds represented by sparse, unordered point cloud scenes in 3D. A new labelled dataset was synthesised to train neural networks and evaluate their ability to estimate the genus of these manifolds. This data used random homeomorphic deformations to provoke the learning of visual topological features. We demonstrate that deep learning models could extract these features and discuss some advantages over existing topological data analysis tools that are based on persistent homology. Semantic segmentation was used to provide additional geometric information in conjunction with topological labels. Common point cloud multi-layer perceptron and transformer networks were both used to compare the viability of these methods. The experimental results of this pilot study support the hypothesis that, with the aid of sophisticated synthetic data generation, neural networks can perform segmentation-based topological data analysis. While our study focused on simulated data, the accuracy achieved suggests a potential for future applications using real data.

Rail traffic planning and scheduling problems have been challenging academy and industry for a few decades. Specifically, problems in the short term and real-time horizons deal with simultaneous decision-making of trains, stations and terminals. Approaches focused on decentralised decision-making have been successful in delivering real-world committed solutions. This work focuses on decentralised real-time decision-making in a closed freight rail network and applies multi-agent deep reinforcement learning (MADRL) to find efficient timetables. We apply the MADRL model to solve the traffic decisions arising in the Hunter Valley Coal Chain (HVCC) in New South Wales, Australia. The approach uses the same simulation model currently in use for capacity planning of the system, thus allowing tests with real data. The environment is modelled as a decentralised, partially observed Markov decision process (dec-POMDP), where the train, load point, and dump station agents decide upon train movements based on local observations. The observations follow a novel state encoding strategy for rail traffic management composed of nine layers. We benefit from this strategy to apply a decentralised execution with a centralised learning approach through proximal policy optimisation. The experiments revealed a significant performance improvement for the ten instances tested, which reproduce the challenges faced in the HVCC operations. The approach is suitable for varied levels of rail network complexity, generating efficient solutions without scaling issues. The MADRL outperformed the heuristic in use by HVCC's simulation model and a high-performance genetic algorithm in all instances, reaching performance improvements of up to 72.00% and 47.42%, respectively. Therefore, the framework with the MADRL and the simulation model allows its application with real world instances in an efficient and reliable way. These results show the method's consistency and draw a safe path towards a decentralised rail traffic management system.

A new method is proposed for utilising scene information for stereo eye tracking in stereoscopic 3D virtual environments. The approach aims to improve gaze tracking accuracy and reduce the required user engagement with eye tracking calibration procedures. The approach derives absolute Point of Regard (POR) from the angular velocity of the eyes without user engaged calibration of drift. The method involves reduction of a hypothesis set for the 3D POR via a process of transformation during saccades and intersection with scene geometry during fixations. A basic implementation of this concept has been demonstrated in simulation using the depth buffer of the scene and a particle representation for the hypothesis set. Future research directions will focus on optimisation of the algorithm and improved utilisation of scene information. The technique shows promise in improving gaze tracking techniques in general, including relative paradigms such as electrooculography.

This paper proposes a simple and fast method for adapting colour lookup tables to lighting changes in real-time. The method adjusts the classified colour space regions keeping both their surface area and volume constant. Two variations of the method were compared and tested in a RoboCup soccer setting. Detection success rate was measured as a function of the speed and magnitude of hue change to the lighting environment. Compared to a static lookup table, these experimental results show improved robustness against lighting changes for detection of coloured objects.

The ability of an autonomous agent to self-localise within its environment is critically dependent on its ability to make accurate observations of static, salient features. This notion has driven considerable research into the development and improvement of feature extraction and object recognition algorithms, both within RoboCup and the robotics community at large. Instead, this paper focuses on the rarely-considered issue imposed by the limited field of view of humanoid robots; namely, determining an optimal policy for actuating a robot's head, to ensure it observes regions of the environment that will maximise the positional information provided. The complexity of this task is magnified by a number of common computational issues; specifically high dimensional state spaces and noisy environmental observations. This paper details the application of motivated reinforcement learning to partially overcome these issues, leading to an 11% improvement (relative to the null case of uniformly distributed actuation policies) in self-localisation and ball-localisation for an agent trained online for less than one hour. The method is demonstrated as a viable method for improving self-localisation in robotics, without the need for further optimisation of object recognition or tuning of probabilistic filters. © 2014 Springer-Verlag Berlin Heidelberg.

Digital Mammograms are x-ray images of the breast and one of the preferred early detection methods for breast cancer. However, mammograms are still difficult to interpret, and associated with this problem is a high percentage of unnecessary biopsies, misdiagnoses and late detections. The focus of this research is to use neuroevolutionary mechanisms for detecting breast cancer from mammographic images. The aim is to design a sophisticated classification tool that detects breast cancer at its early stages, so that treatment has a better chance of success. Wavelet neural networks have the ability to capture and extract information at various frequency levels by using different dilation and scaling values of the wavelet function. In this work, the wavelet neural network parameters are evolved using on the concept of Cartesian Genetic Programming, resulting in an evolved neural network which is trained for mass diagnosis. In the reported study the proposed algorithm achieves a classification accuracy of 89.57% on a real dataset composed of 200 images. Such a computer- based classification system has the potential to provide a second opinion to the radiologists, thus assisting them to diagnose the malignancy of breast cancer more precisely.

The aim of this work is to contribute some insights and a partial overview of how machine learning methods are used in robotics. We first discuss typical general issues in the relationship between robotics and machine learning. Then we locus on projects associated with the RoboCup competition and symposium, and review the extent to which machine learning approaches have been used in the 4-legged league at RoboCup during the years 1998-2003. Further, we summarise the machine learning methods that were used by our own RoboCup team-the NUbots-in 2002/2003.

This paper discusses the Enterprise Information System (EIS) life cycle and the phases of the EIS development life cycle. It details the stages in the EIS life cycle and the characteristics of phases in the system development life cycle and explains where it differs from traditional concepts of software engineering. In particular it defines the concept of generation change and when it is applied to a system. It also describes the nature of the rapid evolution of the EIS and how it results in version or generational change of the system, and how the EIS development life cycle involves a multitude of engineering processes, not just one. This new perspective could lead to the generation of new EIS development methodologies in business modelling, analysis, design, project management and project estimation.

This study empirically investigates variations of hill climbing algorithms for training artificial neural networks on the 5-bit parity classification task. The experiments compare the algorithms when they use different combinations of random number distributions, variations in the step size and changes of the neural networks' initial weight distribution. A hill climbing algorithm which uses inline search is proposed. In most experiments on the 5-bit parity task it performed better than simulated annealing and standard hill climbing. © 1999 IEEE.