Skip to main content
Gloved hand holding a floating bunch of nodes
Research_

Scientific discovery

Furthering important scientific research
There is still so much to discover within the field of science. The research undertaken in this theme will help us to truly understand everything from evolution, to agriculture and human behaviour.

Understanding Earth’s evolution

Vision: Mineral deposits often form at weak points in the Earth’s crust lying on boundaries between geological rock layers with a common formation history. These boundaries are currently mapped subjectively by geologists in the field. We will update these boundaries using geophysical data and provide estimates of uncertainty in boundary locations.

Work: We will use Obsidian, a high-performance computing framework for simulating sensor data given the underlying geology (previously developed in part by centre staff). Our contributions may include new sensor models and/or more efficient Monte Carlo methods for searching the space of geologies.

Research impact: Besides making it easier to locate and exploit mineral deposits, we will develop a framework to quantify uncertainty in qualitative geological observations, and to fuse them with geophysical sensor data.  Our collaborative effort will form a key part of our upcoming Centre of Excellence bid.

Collaboration team:

Research focus:

Vision: Many problems in natural sciences are too complex to write down a closed-form solution, but obey differential equations that can be solved numerically. Fusing, and/or making decisions with, data generated by such processes requires efficient schemes to sample from infinite-dimensional spaces of functions.

Work: We will build on the work of Professor Mark Girolami, who has developed Monte Carlo schemes for inverse problems on linear partial differential equations. One interesting extension would be to sample solutions of non-linear equations or learned kernels, greatly expanding the space of functions that can be sampled.

Research impact: The new sampling schemes will be of great theoretical interest, but should also apply to many areas, including geology or geophysics problems that involve sampling over numerically simulated histories of geological areas. Solving such problems are crucial to reconstructing the Earth’s evolution over the last two billion years.

Collaboration team:

  • Sally Cripps
  • Mark Girolami, Turing Insititute
  • Richard Scalzo (Project lead)

Research focus:

Collaboration team:

  • M Bertolucci
  • Edward Cripps
  • Sally Cripps (Project lead)
  • Ori Rosen

Research focus:


Human behaviour

Vision: Risk prediction instruments are widely used in the criminal justice system. We propose improving the current system by building data-driven risk prediction instruments that can use existing data to learn about the population and predict risk in an individualised manner, but also finding patterns at the group level.

Work: Building a hierarchical model predicting and inferring the factors that explain risk under two different scenarios: domestic violence reoffending and juvenile justice incident risk estimation. The project involves developing and deriving a hierarchical random effects model for assessing risks and also implementing the framework on real data.

Research impact: These risk prediction instrument models can be used for parole decision-making, risk assessment in juvenile justice detention centres and correctional services across the adult population.

Collaboration team:

Research focus:

Vision: There is a global need for building predictive policing models that can use historical data to predict where and when crime will happen. The vision for this project is to improve on existing models for capturing the dynamics of the population demographics over time and couple that with the changing criminal activity over the last 15 years.

Work: Incorporate time into current spacial-demographic crime models (see previous publication by Marchant et al – (2017) Applying Machine Learning to Criminology: Semi-Parametric Spatial-Demographic Bayesian Regression. Security Informatics). By factoring the time dependence structure directly into the model we can capture more complex patterns and extrapolate into the future with more confidence. The work involves implementation of space-time and demographics Gaussian Process regression.

Research impact: These predictive policing tools can be used by the NSW Police Force and other law-enforcing agencies across the country. The impact of these models is of high relevance, considering that predictive policing can inform short- and long-term decision-making in the justice system. For example, patrol optimisation and long term resource allocation strategies can be informed by using this model.

Collaboration team:

Research focus:

 

Vision: Estimating academic capability and potential collaboration by conducting researcher network analysis and visualisation is a key insight for various areas of a complex organisation. Many institutes have implemented such collaborative networks on the researchers’ profile pages based on co-authorship. However, such similarity metrics do not show much value to the researchers themselves because they already know who they have collaborated with in the past. This project will use state-of-the-art statistical analysis and visualisation techniques to build a network of collaboration between researchers based on natural language processing of their existing publications and grants.

Work: The goal of this project is to further develop the front and back ends of the existing prototype web app and develop the following features:

  1. New distance metrics between researchers. How can potential successful collaborations be described and predicted? Not only based on document similarity.
  2. Implement other visualisation options.
  3. Including search capabilities on front-end for adding new academic comparison and project proposal.

Research impact: The models and visualisations informed in this model can improve the efficiency at which multidisciplinary research is conducted within a university or large company.

Collaboration team:

Research focus:

Vision: Improve police response time to criminal events by optimising patrolling routes based on spatial-temporal patterns of criminal activity.

How to navigate within a complex street network in order to maximise a reward function is a complex problem. There are existing algorithms that allow approximate decision-making under uncertainty in a Euclidian space. However, this project aims to provide better informed decision- making by planning directly over street networks.

Work: The work involves using a decision-making algorithm under uncertainty that can allow patrolling units to maximise the chances to catch criminals.

The applications of this novel framework allow generalisation of existing techniques for decision- making in street networks and a new family of kernels for non-parametric regression models that can improve accuracy over Euclidian space kernels.

The goals of this project are the following.

  1. Conduct decision-making over a street network using Open Street Maps.
  2. Apply a decision-making algorithm to police patrolling scenarios and car sharing/pooling optimisation.

Research impact: The research impact of this project is highly relevant because it has implications across both the data-science domain and criminology. As part of this project we will develop Bayesian optimisation for generalised linear models, and apply it to patrolling.

Collaboration team:

Research focus:

Vision: Improve existing models of criminal activity by adding features from the environment. The new models will include green space quantification and density of premises (different types). In the long term, it could be possible to use convolutional neural networks to extract characteristics of the environment that contribute to features of the environment.

Work: The work involves building data handling modules that can extract different information from the environment. After the information is extracted it is fused with other sources of information by using the semi-parametric spatial model of crime.

Research impact: The impact of this project is high for criminology, mainly because it allows to uncover patterns for the relationships between characteristics of the environment and criminal activity.

Collaboration team:

Research focus:

Vision: Recently there has been increased interest in building quantitative models of crime. Authors have particularly used linear regression and other non-parametric statistical models to infer the factor creating crime and predict crime occurrence, which is currently understood as predictive policing. However, even though the techniques used for building these models allow for uncertainty quantification and inference, they are not the most powerful in terms of predictive performance.

Compare existing methodology for predictive policing with deep learning networks that can potentially improve predictive accuracy at the cost of reduced interpretability.

Work: The work involves using deep learning models over spatial-demographic data and training these deep networks using existing criminal records.

Research impact: The impact of this project is high for criminology, mainly because it can help inform predictive policing models in a more accurate manner and in turn improve patrolling response to dynamics of crime.

Collaboration team:

Research focus:

Vision: There is an increasing number of tools to visualise spatial/temporal data. Lately, the use of java-script and web visualisation applets is becoming the norm due to their flexibility. Data61 has developed an open source platform called TerriaJS. However, it is not easy to interface in order to represent continuous density function.

Work: The goals of this project are the following.

  1. Implement the visualisation of spatial/temporal density surfaces over a geographical map. In order to achieve this, the project will make use of the latest release of Deck.GL, a large-scale WebGL-powered data visualisation platform.
  2. Visualise over Deck.GL the paths of autonomous cars, including the decision tree and reward function over street levels, to understand decision-making.

Research impact:

State-of-the-art visualisation techniques allow for human comprehension complex results exposed by machine learning algorithms. The developed visualisation platform will be used by researchers and could also be helpful for police, to help them visualise spatial temporal density of crime and allocate police resources.

Collaboration team:

Research focus:

Vision: This project seeks to map out putative relationships between the number, expenditure and distribution of gaming machines across geographical regions, considering relevant socio-economic variables, density of criminal offences and the prevalence of gambling disorders and gambling-related harms.

Work: Generating hierarchical probabilistic models for longitudinal data that can capture dependencies between demographic characteristics, spatial correlations, gambling revenue and the dynamics of gambling disorders and crime.

Research impact: The advantages presented by the integrated analyses of these databases are that they would be based on objective data and will highlight the prevalence of gambling problems within specific geographical regions, and relevant socio-demographic variables that could be used to identify vulnerable sectors of the community.

The analyses would be unique because they would allow for the inclusion of additional variables such as distribution of alcohol venues, poverty and social disadvantage that could potentially contribute to, or exacerbate, the incidence of gambling problems. These variables would inform policy decisions on the optimal location of treatment/counselling services and responsible gambling strategies targeting at-risk subpopulations.

Collaboration team:

Research focus:

Collaboration team:

Research focus:

Collaboration team:

Research focus:

Vision: Risk prediction instruments are widely used in the criminal justice system. We propose improving the current system by building data-driven instruments that can use existing data to learn about the population and predict risk in an individualised manner, but also find patterns at the group level.

Work: Building a hierarchical model predicting and inferring the factors that explain risk under two different scenarios: domestic violence reoffending and juvenile justice incident risk estimation. The project involves developing and deriving a hierarchical random effects model for assessing risks and also implementing the framework on real data.

Research impact: These risk prediction instruments can be used for parole decision-making, risk assessment in juvenile justice detention centres and correctional services across the adult population.

Collaboration team:

Research focus:

Vision: There is a global need for building predictive policing models that can use historical data to predict where and when crime will happen. The vision for this project is to improve on existing models for capturing the dynamics of the population demographics over time, and couple that with the changing criminal activity over the last 15 years.

Work: Incorporate time into current spacial-demographic crime models (see the previous publication by Marchant et al – (2017) Applying Machine Learning to Criminology: Semi-Parametric Spatial-Demographic Bayesian Regression. Security Informatics). By considering the time dependence structure directly into the model we can capture more complex patterns and extrapolate into the future with more confidence. The work involves implementation of space-time and demographics Gaussian Process regression.

Research impact: These predictive policing tools can be used by the NSW Police Force and other law-enforcing agencies across the country. The impact of these models is of high relevance, because predictive policing can inform short- and long-term decision-making in the justice system. For example, patrol optimisation and long-term resource allocation strategies can be informed by using this model.

Collaboration team:

Research focus:


Life sciences

Vision: Automatic dairy farms, and the consequent data capture, offer an opportunity to drive business improvements in all aspects of dairy farm operations. The aim of this project is to find the optimal allocation of supplemental feed to cows and how this time-varying allocation varies with cow specific attributes.

Work: To determine optimal feed allocation, we develop a dynamic Bayesian model to capture the dependence between the milk yield and supplemental feed at the individual cow level, taking into account other confounding factors.

Research impact: The research will allow development of a novel Bayesian model for optimal feed allocation of dairy cows and expected to increase the milk yield with utilisation of the same resources.

Collaboration team:

Research focus:

  • Life sciences

Other projects

Vision: Integral field units (IFUs) provide detailed maps of the chemical components and dynamics of galaxies in three-dimensional “datacubes”. These cubes often contain spurious, non-astrophysical structures caused by the atmosphere or residual instrumental signatures. Our new analysis method removes these effects, maps to standard coordinates, and provides reliable uncertainties.

Work: We used IFU data from the SAMI Galaxy Survey as a test case. Our datacubes are modeled as a Gaussian process, with covariance transformed by the instrument response (which can be numerically calculated, and specified as an input for use of the method on data from different instruments).

Research impact: We believe our method has the potential to become widely used for IFU data analysis, which will dramatically improve the accuracy and precision of observations in many upcoming next-generation surveys of thousands of galaxies.

Collaboration team:

Research focus:

Vision: Type Ia supernovae are used as distance indicators to study the universe’s accelerating expansion (for which the 2011 Nobel Prize in Physics was awarded), but their explosion mechanisms are still not fully understood. The progenitor star’s mass is a key discriminant, but is not directly observable and must be inferred.

Work: We used Gaussian process regression to measure the energy emitted by each supernova as a function of time since explosion, which we used as input to a flexible probabilistic model designed to emulate common physics in three-dimensional explosion simulations. Outputs included total ejected mass and radioactive nickel mass.

Research impact: Project lead Scalzo’s previous use of this technique showed that “normal” type Ia supernovae used for cosmology eject a broad range of masses, directly refuting the traditional explosion scenario with a single mass scale. This work derives ejected masses for individual “peculiar” supernovae spanning the full diversity of explosion scenarios.

Collaboration team:

  • Chris Burns, Carnegie Observatories
  • Mark Phillips, Carnegie Observatories
  • Richard Scalzo (Project lead)
  • Max Stritzinger, Aarhus University
  • The Carnegie Supernova Project Collaboration

Research focus:

Vision: Large astronomical imaging surveys often find rare transient events, such as supernovae, stellar flares or neutron star mergers. They also have noise or outlier detections that can mimic astrophysical transient signals, and that far outnumber real transients. This project aims to improve automatic vetting of targets.

Work: The standard classifier used for this task before our work was random forest, which requires features extracted from the data. We explored convolutional neural net (CNN) architectures to automatically select relevant features and produce improved classifications, trained on data from the SkyMapper robotic survey (augmented using known data symmetries).

Research impact: Ours is the first application of deep learning to this particular task, but we doubt it will be the last. Future extensions of our approach might include online training environments for large surveys, which can learn new noise and outlier models for each telescope and update as new data arrive.

Collaboration team:

  • Stephen Bloemen, University of Nijmegen (Project lead)
  • Fabian Gieseke, University of Nijmegen
  • Richard Scalzo
  • Fang Yuan, ANU

Research focus:

Vision: The Murchison Wide-Field Array Transient Survey (MWATS) contains observations of over 300,000 time-varying radio sources, presenting a rich vein of data on high-energy phenomena in the universe. We will de-noise the data, separating true astrophysical variability from instrumental signatures and providing robust uncertainties on radio fluxes.

Work: We will build a scalable joint Bayesian model of the instrument and all relevant sources of noise, drawing on our previous experience from the SAMI datacube project (see above) and on optical-wavelength transient surveys, including surveys for supernovae and exoplanets.

Research impact: The project will result in a high-quality public catalog of variable radio sources which will be mined by astronomers worldwide. It will also showcase the potential for Bayesian techniques in radio astronomy, where they are not yet widely used, and in other large astronomical survey projects operating at scale.

Collaboration team:

Research focus: