Kernel and Margin Based Machine Learning Algorithms. Certain machine learning algorithms, such as support vector machines, utilizing
the ideas of large margins and kernels have attracted much attention lately
because of their impressive performance on real world problems such as optical
character recognition. We plan to refine and extend such algorithms to a wide
range of different machine learning problems such as gene sequence analysis,
image processing and text classification. Expected ....Kernel and Margin Based Machine Learning Algorithms. Certain machine learning algorithms, such as support vector machines, utilizing
the ideas of large margins and kernels have attracted much attention lately
because of their impressive performance on real world problems such as optical
character recognition. We plan to refine and extend such algorithms to a wide
range of different machine learning problems such as gene sequence analysis,
image processing and text classification. Expected outcomes include the
development of software that allows the solution of hitherto unsolved machine
learning problems, and the ability to solve problems larger than those solvable
by the current generation of machine learning tools.Read moreRead less
Analysing Iterative Machine Learning Algorithms with Information Geometric Methods. Online machine learning problems arise from situations where data is provided a point at a time. There are many classical algorithms for solving such problems based on the principle of stochastic gradient descent. Recent research by the CIs and others have thrown up interesting but diverse geometric connections that offer new insights. The proposed research aims to integrate the understanding of these algori ....Analysing Iterative Machine Learning Algorithms with Information Geometric Methods. Online machine learning problems arise from situations where data is provided a point at a time. There are many classical algorithms for solving such problems based on the principle of stochastic gradient descent. Recent research by the CIs and others have thrown up interesting but diverse geometric connections that offer new insights. The proposed research aims to integrate the understanding of these algorithms with the aim of designing algorithms better able to exploit prior knowledge, and to extend existing algorithms to new problem domains thus offering well principled and well understood algorithms for solving a variety of novel online problems.Read moreRead less
Discovery Early Career Researcher Award - Grant ID: DE240101129
Funder
Australian Research Council
Funding Amount
$442,000.00
Summary
Synergy between future 21-cm experiments and physical cosmology. The nature of dark matter and formation of the first galaxies are both unsolved mysteries. During the first 500 million years, our universe was filled with hydrogen atoms illuminated by the first galaxies. The 21-cm radiation from this gas encodes properties of unseen galaxies and dark matter during this so-called cosmic dawn. This project aims to build an innovative framework to leverage future 21-cm experiments using The Square K ....Synergy between future 21-cm experiments and physical cosmology. The nature of dark matter and formation of the first galaxies are both unsolved mysteries. During the first 500 million years, our universe was filled with hydrogen atoms illuminated by the first galaxies. The 21-cm radiation from this gas encodes properties of unseen galaxies and dark matter during this so-called cosmic dawn. This project aims to build an innovative framework to leverage future 21-cm experiments using The Square Kilometre Array to observe cosmic dawn, and to forecast the optimal constraints on dark matter physics. Additional outcomes include the largest cosmological simulation of the first galaxies powered by neural networks and improved knowledge of their properties using Bayes' theorem and The James Webb Space Telescope.Read moreRead less
Automatic Training Data Search and Model Evaluation by Measuring Domain Gap. We aim to investigate computer vision training data and test data, using automatically generated data sets for facial expression recognition and object re-identification. This project expects to quantify and understand the domain gap, the distribution difference between training and test data sets. Expected outcomes of this project are insights on measuring the domain gap, the ability to estimate model performance witho ....Automatic Training Data Search and Model Evaluation by Measuring Domain Gap. We aim to investigate computer vision training data and test data, using automatically generated data sets for facial expression recognition and object re-identification. This project expects to quantify and understand the domain gap, the distribution difference between training and test data sets. Expected outcomes of this project are insights on measuring the domain gap, the ability to estimate model performance without accessing expensive test labels and improvements to system generalisation. This should provide significant benefits for computer vision applications that currently require expensive labelling, and commercial and economic benefits across sectors such as transportation, security and manufacturing.Read moreRead less
Design of dynamic visual signals. Models of the design of visual signals depend heavily upon analyses of static ornaments. Nothing is known about dynamic visual signals. We will use an array of new techniques to tackle this problem for the first time. Motion analyses will define the task faced by the visual system. Sensory limitations will be measured to identify constraints on signal evolution. Digital video playback studies will assess recognition and explain aspects of signal design. Results ....Design of dynamic visual signals. Models of the design of visual signals depend heavily upon analyses of static ornaments. Nothing is known about dynamic visual signals. We will use an array of new techniques to tackle this problem for the first time. Motion analyses will define the task faced by the visual system. Sensory limitations will be measured to identify constraints on signal evolution. Digital video playback studies will assess recognition and explain aspects of signal design. Results will test the generality of principles that have been developed in studies of female mate choice and extend these ideas to address intra-sexual selection operating through opponent assessment.Read moreRead less
Towards in-vehicle situation awareness using visual and audio sensors. This project aims to characterise driver awareness, activity and interactions with other vehicle occupants using visual and audio cues from internally mounted sensors. Road accidents cost Australia an estimated $30 billion per year and tragic loss of thousands of lives, yet the vast majority of severe vehicle crashes are linked to driver fatigue or distraction. The expected project outcomes include advanced artificial intelli ....Towards in-vehicle situation awareness using visual and audio sensors. This project aims to characterise driver awareness, activity and interactions with other vehicle occupants using visual and audio cues from internally mounted sensors. Road accidents cost Australia an estimated $30 billion per year and tragic loss of thousands of lives, yet the vast majority of severe vehicle crashes are linked to driver fatigue or distraction. The expected project outcomes include advanced artificial intelligence to infer and predict dangerous driver and passenger behaviour. This has the potential to significantly benefit society by advancing autonomous driving capabilities and reducing driver-induced accidents and fatalities, ensuring that every driver, passenger and pedestrian arrives home safely at the end of each day.Read moreRead less
A Novel Automatic Neural Network Feature Extractor. This project aims to study feature extraction abilities of convolutional as well as traditional neural networks and develop a generic feature extractor which can be applied to wide variety of real-world image and non-image data. New concepts for automatic feature extraction, feature explanation, hybrid evolutionary algorithms and non-iterative ensemble learning will be introduced and evaluated. The expected outcomes are a generic feature extrac ....A Novel Automatic Neural Network Feature Extractor. This project aims to study feature extraction abilities of convolutional as well as traditional neural networks and develop a generic feature extractor which can be applied to wide variety of real-world image and non-image data. New concepts for automatic feature extraction, feature explanation, hybrid evolutionary algorithms and non-iterative ensemble learning will be introduced and evaluated. The expected outcomes are a generic feature extractor for automatically extracting features, an optimiser for finding optimal parameters and non-iterative ensemble learning technique for classification of features into classes. The impact of this project will be automatic feature extractors and classifiers for real-world applications.Read moreRead less
Handling unreliable, uncertain and inadequate data for Intelligence led Investigation. Intelligence led investigation has been successful recently in drug and people smuggling, preparation or instigation of acts of terrorism, and can benefit profoundly from the techniques we will develop, in the timely management and inference from many sources and kinds of uncertain information. This work will assist in making Australia a safer and more secure country.
E.g., Australian Bureau of Statistics ....Handling unreliable, uncertain and inadequate data for Intelligence led Investigation. Intelligence led investigation has been successful recently in drug and people smuggling, preparation or instigation of acts of terrorism, and can benefit profoundly from the techniques we will develop, in the timely management and inference from many sources and kinds of uncertain information. This work will assist in making Australia a safer and more secure country.
E.g., Australian Bureau of Statistics figures show that for 2004, investigations of some 35% of murders, 63% of kidnappings, and 80% of robberies are incomplete at 30 days. Terrorism investigations are harder in that usually there is no initial crime trigger for an investigation. Any assistance our tools can provide in will be of significant benefit to Australia.Read moreRead less
Democratisation of Deep Learning: Neural Architecture Search at Low Cost. The need to manually design Deep Learning-based Neural Networks (DNNs) limits their usage to AI experts and hinders the exploitation of their true potential more broadly, e.g., in farming, humanities. We aim to replace this tedious process through novel AI methods capable of generating DNNs that can perform significantly better and at a lower computational cost than manually designed DNNs. We further expand this idea to so ....Democratisation of Deep Learning: Neural Architecture Search at Low Cost. The need to manually design Deep Learning-based Neural Networks (DNNs) limits their usage to AI experts and hinders the exploitation of their true potential more broadly, e.g., in farming, humanities. We aim to replace this tedious process through novel AI methods capable of generating DNNs that can perform significantly better and at a lower computational cost than manually designed DNNs. We further expand this idea to solve complex real-world problems with both labelled and unlabelled data found in various applications including energy and climate change. The expected outcomes include the novel AI methods, highly trained AI researchers and a number of critical applications that will bring significant benefits to Australia and the world.Read moreRead less
Machine education for trusted multi-skilled evolutionary learners . Transforming data assets into organisational knowledge assets sits in the hands of a few, highly specialised, data scientists. The aim of this research is to design educational instruments to support non-experts to teach artificial intelligence (AI) systems in a similar way to educating human teachers to teach human learners. The significance of the project lies in affording the wider smart, but not necessarily AI expert, commun ....Machine education for trusted multi-skilled evolutionary learners . Transforming data assets into organisational knowledge assets sits in the hands of a few, highly specialised, data scientists. The aim of this research is to design educational instruments to support non-experts to teach artificial intelligence (AI) systems in a similar way to educating human teachers to teach human learners. The significance of the project lies in affording the wider smart, but not necessarily AI expert, community the ability to contribute to growing our knowledge-based society in a safe, transparent and trustworthy manner. Outcomes will include innovative instruments to teach machines, novel knowledge creation, trusted and transparent AI systems, and a new generation of human teachers specialised in educating AI systems.Read moreRead less