Uncertainty, Risk and Related Concepts in Machine Learning. Machine learning is the science of making sense of data. It does not and cannot remove all risk and uncertainty. This project proposes to study the foundations of how machine learning uses, represents and communicates risk and uncertainty. It aims to do so by finding new theoretical connections between diverse notions that have arisen in allied disciplines. These include risk, uncertainty, scoring rules and loss functions, divergences, ....Uncertainty, Risk and Related Concepts in Machine Learning. Machine learning is the science of making sense of data. It does not and cannot remove all risk and uncertainty. This project proposes to study the foundations of how machine learning uses, represents and communicates risk and uncertainty. It aims to do so by finding new theoretical connections between diverse notions that have arisen in allied disciplines. These include risk, uncertainty, scoring rules and loss functions, divergences, statistics and different ways of aggregating information. By building a more complete theoretical map it is expected that new machine learning methods will be developed, but more importantly that machine learning will be able to be better integrated into larger socio-technical systems.Read moreRead less
Principled statistical methods for high-dimensional correlation networks. This project aims to develop a novel and principled approach for building correlation networks. Correlation networks aim to identify the most significant associations present in modern massive datasets, and have numerous applications, ranging from the biomedical and environmental sciences to the social sciences. Nodes of such networks represent features, and edges represent associations, or the lack thereof. Current method ....Principled statistical methods for high-dimensional correlation networks. This project aims to develop a novel and principled approach for building correlation networks. Correlation networks aim to identify the most significant associations present in modern massive datasets, and have numerous applications, ranging from the biomedical and environmental sciences to the social sciences. Nodes of such networks represent features, and edges represent associations, or the lack thereof. Current methods are not readily scalable to modern ultra-high dimensional settings, and do not account for uncertainty in the estimated associations. This project will develop a principled, highly scalable methodology for building such networks, which incorporates uncertainty quantification. Emphasis is placed on modern ultra-high dimensional settings in which differentiating a true correlation from a spurious one is a notoriously difficult task.Read moreRead less
Discovery Early Career Researcher Award - Grant ID: DE180101252
Funder
Australian Research Council
Funding Amount
$343,450.00
Summary
Statistical theory and algorithms for joint inference of complex networks. This project aims to address the challenges in jointly modelling complex networks by applying an integrated approach encompassing statistical theory, computation, and applications. The project expects to contribute to core statistical methodology development for complex inference and generate new knowledge in the fields of genomics, neuroscience, and social science through in-depth analyses of large-scale multilayered net ....Statistical theory and algorithms for joint inference of complex networks. This project aims to address the challenges in jointly modelling complex networks by applying an integrated approach encompassing statistical theory, computation, and applications. The project expects to contribute to core statistical methodology development for complex inference and generate new knowledge in the fields of genomics, neuroscience, and social science through in-depth analyses of large-scale multilayered network data. Expected outcomes include enhanced theoretical and computational frameworks for probabilistic network models to better utilise the power of multiple observations. This should foster international and interdisciplinary collaborations and add significant value to the rapidly progressing field of networks research.Read moreRead less
Inference for Hawkes processes with challenging data. The Hawkes processes are statistical models for the analysis of high-impact event sequences, such as bushfires, earthquakes, infectious diseases, and cyber attacks. When the times and/or marks are missing for some events or when the data is otherwise incomplete, it is challenging to fit these models and perform diagnostic checks on the fitted models. This project aims to develop novel statistical methods to fit these models in the presence of ....Inference for Hawkes processes with challenging data. The Hawkes processes are statistical models for the analysis of high-impact event sequences, such as bushfires, earthquakes, infectious diseases, and cyber attacks. When the times and/or marks are missing for some events or when the data is otherwise incomplete, it is challenging to fit these models and perform diagnostic checks on the fitted models. This project aims to develop novel statistical methods to fit these models in the presence of incomplete data and to check the goodness-of-fit of the fitted models. The expected outcomes include publications documenting these methods and software packages implementing them. The primary benefits include the advancement of statistical methodology and the training of junior research personnel. Read moreRead less
Fast flexible feature selection for high dimensional challenging data. The project aims to provide new frameworks for fast flexible feature selection and appropriate modelling of heterogeneous data through structural varying-coefficient regression models. The outcomes will be a series of new statistical methods and concepts enabling more powerful modelling of complex bioscience data. The project will create the science for building reliable statistical models taking model uncertainty into accoun ....Fast flexible feature selection for high dimensional challenging data. The project aims to provide new frameworks for fast flexible feature selection and appropriate modelling of heterogeneous data through structural varying-coefficient regression models. The outcomes will be a series of new statistical methods and concepts enabling more powerful modelling of complex bioscience data. The project will create the science for building reliable statistical models taking model uncertainty into account, impacting how results will be interpreted, and with accompanying software. This will be a significant improvement in the assessment of model confidence in the food and health research priority areas including areas such as meat science, Huntington’s disease, and kidney transplantation.Read moreRead less
Fast approximate inference methods: new algorithms, applications and theory. This project aims to develop new algorithms and theory for fast approximate inference and lay down infrastructure to aid future extensions. Fast approximate inference methods are a principled and extensible means of fitting large and complex statistical models to big data sets. They come into their own in applications where speed is paramount and traditional approaches are not feasible. The project aims to lead to prac ....Fast approximate inference methods: new algorithms, applications and theory. This project aims to develop new algorithms and theory for fast approximate inference and lay down infrastructure to aid future extensions. Fast approximate inference methods are a principled and extensible means of fitting large and complex statistical models to big data sets. They come into their own in applications where speed is paramount and traditional approaches are not feasible. The project aims to lead to practical outcomes from better business decision-making for insurance data warehouses, to improved medical imaging technology.Read moreRead less
Learning the meso-scale organization of complex networks. This project aims to model and learn the organization of online social networks. We will combine mathematical models, inference, and domain knowledge from computational social sciences to obtain interpretable descriptions of the role groups of users play in the network. The expected outcomes are new mathematical models and computational methods that learn from data how to best decompose a complex network into building blocks and their int ....Learning the meso-scale organization of complex networks. This project aims to model and learn the organization of online social networks. We will combine mathematical models, inference, and domain knowledge from computational social sciences to obtain interpretable descriptions of the role groups of users play in the network. The expected outcomes are new mathematical models and computational methods that learn from data how to best decompose a complex network into building blocks and their interactions, linking connectivity to function. This should provide benefits to industries and policy makers interested in how information spreads in social media, including the critical questions of understanding the mechanisms contributing to political polarization and fragmentation.Read moreRead less
Large dynamic time-varying models for structural macroeconomic inference. This project aims to broaden the range of macroeconomic models that have an integrated capacity for both greater realism and efficiency in analysis. This approach will be applied to two contexts at the forefront of current macroeconomic research, the effects of noisy productivity signals on business cycles and the effects of fiscal policy shocks. Flexible macro-econometric models underpin accurate inference by economists ....Large dynamic time-varying models for structural macroeconomic inference. This project aims to broaden the range of macroeconomic models that have an integrated capacity for both greater realism and efficiency in analysis. This approach will be applied to two contexts at the forefront of current macroeconomic research, the effects of noisy productivity signals on business cycles and the effects of fiscal policy shocks. Flexible macro-econometric models underpin accurate inference by economists and policymakers and the project outputs should provide widespread and significant benefits by improving policy and boosting Australia’s comparative advantage.Read moreRead less
New methods for modelling real-world extremes. This project aims to develop new theory and methods for analysing and predicting extreme values observed in real-world processes. Many existing techniques are limited by convenient mathematical assumptions that commonly do not hold in practice: dependence at asymptotic levels, process stationarity, and that the observed data are direct measurements of the process of interest. As a result, using these techniques may produce undesirable results. Expec ....New methods for modelling real-world extremes. This project aims to develop new theory and methods for analysing and predicting extreme values observed in real-world processes. Many existing techniques are limited by convenient mathematical assumptions that commonly do not hold in practice: dependence at asymptotic levels, process stationarity, and that the observed data are direct measurements of the process of interest. As a result, using these techniques may produce undesirable results. Expected outcomes of this project include theoretically justified data analysis techniques that can accurately model extreme values seen in the real world. Project benefits include more realistic analyses of nationally important applications in climate, bushfire insurance risk, and anomaly detection.Read moreRead less
Discovery Early Career Researcher Award - Grant ID: DE210100256
Funder
Australian Research Council
Funding Amount
$415,283.00
Summary
Extracting the hidden structure of glass from particle vibrations. Predicting the rigid behaviour of glass from its disordered, amorphous atomic structure remains a challenge in materials science. This project aims to define an innovative measure of structure based on how constrained each particle is, which can be quantified by measuring the particles’ vibrations. Using this new measure of structure, this project expects to link the microscopic structure of glass to its macroscopic properties v ....Extracting the hidden structure of glass from particle vibrations. Predicting the rigid behaviour of glass from its disordered, amorphous atomic structure remains a challenge in materials science. This project aims to define an innovative measure of structure based on how constrained each particle is, which can be quantified by measuring the particles’ vibrations. Using this new measure of structure, this project expects to link the microscopic structure of glass to its macroscopic properties via computer simulations. Expected outcomes of this project include a new methodology for characterising amorphous materials and an improved understanding of the nature of glass. This should provide significant benefits, such as an increased ability to rationally design amorphous materials with desired properties.Read moreRead less