Data61 PhD Scholarships project list

Granular matter is an example of a complex system in which the fine details of the interactions between particles at small scales profoundly affects the observed macroscopic behaviour of the material. This has numerous real world consequences, from the dramatic in for example the sudden onset of landslides, to the mundane (but highly impactful) problems of processing powders in pharmaceuticals and ensuring proper mixing of ingredients in food products. Numerical simulations lie at the forefront of granular matter research. This project will explore new ways of combining AI and physics based computational modelling within a consistent uncertainty quantification framework.

Supervisors

Richard Scalzo and Gary Delaney (CSIRO's Data61)

Location

Clayton, Victoria

In many real-world applications, decisions need to be taken at multiple levels where a current decision may be based on several previous decisions. These decisions may be made by several Machine Learning (ML) models along the pipeline, each producing an intermediate decision. Uncertainty from errors made by models in earlier steps of decision making, propagate through the process and the final decision will be impacted from errors made in earlier levels. We are currently conducting a project on prawn measurement under Digiscape funding. We have collected data for the prawn measurement across multiple Digiscape projects and have a use case. The PhD student will have access to real world dataset to validate the developed algorithms as part of PhD project.

Supervisors

Ashfaqur Rahman and Joel Dabrowski (CSIRO's Data61)

Location

Sandy Bay, Tasmania

Strategic foresight is a popular tool that policy makers use to anticipate potential opportunities and challenges that could arise when a policy is implemented. Traditionally, researchers often consider qualitative approaches in strategic foresight due to the complexity of data. This project aims to provide a probabilistic data-driven framework for strategic foresight and scenario planning. This project will provide policy makers a probabilistic data-driven tool for scenario planning.

Supervisors

Stefan Hajkowicz and Kelly Trinh (CSIRO's Data61)

Location

Sydney, Melbourne, Brisbane, Canberra or remote where requirements can be met

Government and industry are increasingly interested in AI-assistance in document preparation. Recent methods in artificial intelligence (AI) and natural language processing (NLP) have shown the tremendous potential of neural network language models to capture language use and even some knowledge/societal norms. The PhD student will work with open domain data sets to explore key NLP problems in the context of this NLP application space, such as: (1) neural language modelling, (2) information extraction for scientific documents, (3) discourse modelling, (4) improving language models with domain knowledge, and (5) data augmentation and data set preparation for NLP in low-resource scenarios. Applicants should have a research background in one or more of the following areas: linguistics, computational linguistics, natural language processing, artificial intelligence, machine learning.

Supervisors

Stephen Wan (CSIRO's Data61)

Location

Eveleigh, NSW

The project involves developing novel methods for state inference in complex dynamical systems with later focus on the estimation of joint model and state from observations of the state over time. The student will look at novel methods for inference in these settings that consider these factors with an emphasis on rigorous algorithm development and performance characterisations. Expected outcomes will be methods as just detailed with software and published articles to be produced. The student will likely work closely with practitioners in application fields (like climate scientists) as well as with data scientists and mathematical statisticians and machine learners with the aim of producing practical methods that may be tested in real systems.

Supervisors

Edwin Bonilla (CSIRO's Data61) and Adrian Bishop (University of Technology Sydney and CSIRO’s Data61)

Location

Eveleigh and University of Technology Sydney, NSW

Machine learning (ML) has revolutionized many aspects of our daily lives with incredible breakthroughs in computer vision, speech analysis and natural language processing. However, modern ML techniques such as deep learning are notorious for being data hungry (requiring large amounts of labelled data to train) and difficult to interpret. In this project we will investigate different approaches to combining ML with mechanistic models in order to make ML methods more data efficient and provide a better understanding of the world through their mechanistic counterpart.

Supervisors

Edwin Bonilla (CSIRO's Data61)

Location

Eveleigh, NSW

Musculoskeletal injuries and disorders (MSDs) comprise more than 70% of workplace injury claims and affect most people who enjoy sports and physical pastimes. Whilst very common and highly costly to the individual, country and businesses, MSD risk cannot yet be predicted. CSIRO have developed a markerless motion capture (MMC) system that can non-invasively measure human movement and predict external body loads in any environment. The La Trobe University team perform studies with healthy and injured participants during which motion capture, health and strength data are measured in the laboratory. They have hundreds of datasets of joint surgery patients and suburban athletes (with and without MSDs) from which new models can be developed. The PhD students will develop novel hybrid biomechanical physics-based and machine learning (PB-ML) models that will predict workplace or sports injury risk and clinical outcomes from data collected in the gait lab. They will test these models in the workplace, clinic and sports environments for improving health and performance outcomes.

Supervisors

Simon Harrison and Raymond Cohen (CSIRO's Data61) in collaboration with Kay Crossley, Jodie McClelland and Kane Middleton (La Trobe University)

Location

Clayton, Victoria

Algorithms for inferring the structure of Graphical models (e.g. Bayesian networks) from data have become increasingly popular methods for uncovering the direct and indirect influences among random variables and facilitate causal discovery. The main challenge is that in complex systems, there is a vast number of possible combinations of the network structures which strongly limits the inference scalability.
The aim of this project is to tackle the scalability problem via transforming the original discrete/continuous hybrid parameter space into a completely continuous space where scalable approximate inference via Markov chain Monte Carlo (MCMC) and variational inference (VI) techniques are feasible.
Students participating in this project will obtain a deep insight into the theory and applications of the structure learning problem as well as the state-of-the-art approximate inference tools.

Supervisors

Hadi Afshar and Edwin Bonilla (CSIRO's Data61) in collaboration with Robert Kohn (UNSW)

Location

Eveleigh, NSW

Causal modelling provides decision-makers with a tool to assess the efficacy of a proposed intervention such as resource allocation. However, in many scenarios, computing the likelihood is computationally intractable. This project will tackle this problem by developing approximate inference methodologies for models with intractable likelihood, and leveraging the power of quantum computation.

Supervisors

Roman Marchant and Gilad Francis (CSIRO's Data61) in collaboration with Minh-Ngoc Tran (The University of Sydney).

Location

Eveleigh NSW or Pullenvale, QLD

Structure learning in causal inference is a complex challenge which needs large quantities of observational data to infer causal relationships. Generally, Randomised Control Trials are used to determine causal effects in a frequentist setting, affecting large numbers of the population. However, these trials are both costly and inefficient. Bayesian Adaptive Trials (BATs) are a novel technique for sequentially acquiring relevant information, deciding simultaneously the characteristics of individuals recruited in the trial and the associated intervention. This project will further develop BATs for specific use in causal discovery.

Supervisors

Roman Marchant and Gilad Francis (CSIRO's Data61) in collaboration with Robert Kohn (UNSW).

Location

Eveleigh, NSW or Pullenvale, QLD

Performing inference on edge devices via deep-learning models for data-driven tasks such as classification, semantic segmentation, and object detection finds numerous applications, e.g., in environmental sensing, digital agriculture, supply chain integrity, and advanced manufacturing. This PhD project is around developing new deep neural-network architectures with compact models that can be used to perform end-to-end inference on embedded systems/edge devices with limited memory, energy, and computational resources.

Supervisors

Reza Arablouei, Volkan Dedeoglu and Jiajun Liu (CSIRO's Data61) in collaboration with Yifan Liu (University of Adelaide).

Location

Pullenvale, QLD

Digital twins (DTs) are virtual representations of physical objects or systems. They often rely on sensor data to facilitate forecasting, analysis, and decision support. In this project, we will develop trust mechanisms for DTs to evaluate, quantify, and monitor the trust in the DT sensor data and the related models, and consequently the confidence in the DT-induced insights and predictions.

Supervisors

Volkan Dedeoglu and Reza Arablouei (CSIRO's Data61) in collaboration with Raja Jurdak (QUT)

Location

Pullenvale, QLD

In recent years, the development of deep learning has brought significant success to the scene of understanding tasks on benchmark datasets, but often with a high computational cost. This task will be even more expensive when extended to the video sequence. To apply to real-world, especially edge applications, such as detecting animals in the wild, tracking the products on the conveyor belt, and monitoring the working environment of miners to prevent hazards, it is important to build an efficient video scene understanding system. This project aims to design efficient scene understanding frameworks based on compressed video streams.

Supervisors

Jiajun Liu, Olivier Salvado, Mark Hedley, Reza Arablouei (CSIRO's Data61) in collaboration with Yifan Liu (University of Adelaide)

Location

Pullenvale, QLD

The large number of emerging short videos in social media apps and online platforms poses growing challenges to current techniques in video understanding and privacy-preserving techniques. The existing techniques for detecting and defending privacy leakage issues in social app videos still face the following major challenges: lack of an in-depth understanding of behaviour leakage, and lack of effective defensive techniques for behaviour leakage. In this project we propose to design and implement a foundational adversarial behaviour learning framework to understand, discover and preserve sensitive video information.

Supervisors

Xun Li and David Ahmedt (CSIRO's Data61) in collaboration with Yulei Sui and Xin Yu (University of Technology Sydney)

Location

Marsfield, NSW or University of Technology Sydney, NSW

In this project, the student will investigate and develop novel weakly supervised machine learning-based methods to detect, classify and recognize actions from large-scale videos. In this context, both vision-based data and audio-visual data will be analysed and learned in a discriminant manner. Aspirational aims include efficient video analytics for action indexing and to benchmark its performance in a resource-constrained environment.

Supervisors

Zeeshan Hayder (CSIRO's Data61) in collaboration with Dr Ajmal Saeed Mian (University of Western Australia), Dr Jing Zhang (ANU) and Ali Zia (CSIRO’s Data61 and ANU).

Location

Canberra, ACT (CSIRO and ANU) or University of Western Australia, WA

The sandstone country of Cape York hosts one of the richest bodies of rock art in the world where spectacular galleries document the life-ways of generations of Aboriginal peoples. Classification of motifs in images of rock art provides a unique and challenging machine learning problem. In this project the student will investigate attention Graph Neural Networks to interpret the visual and contextual relationship for image captioning in a hierarchical manner. The system will be capable of categorising rock art motifs from this incredible rock art province using a taxonomy developed in collaboration with the expert Indigenous community.

Supervisors

Mohamad Ali Armin and David Ahmedt (CSIRO's Data61) in collaboration with Gervase Tuxworth, Lynley Wallis and Paulo de Souza (Griffith University)

Location

Pullenvale and Griffith University, QLD

This project intends to develop novel state-of-the-art dexterous dual-arm manipulation capabilities for soft materials (like vegetation) in field robots. This PhD project will look at developing dual-arm manipulation capabilities in robot-vegetation interaction and will showcase the ability to pluck fruits, sense under dense canopy, and collect leaves and other biological samples for plant health monitoring.

Supervisors

Tirthankar Bandyopadhyay and Brendan Tidd (CSIRO's Data61)

Location

Pullenvale, QLD or Clayton, Victoria

Legged locomotion for robots has become very popular in the recent years, especially with the availability of multiple commercially quadruped platforms. While they are impressive, their operation is largely limited to flat or moderately uneven terrain. This is in stark contrast to multi-legged animals in nature who can effortlessly traverse extremely challenging terrain. The focus of this work would be to address the problem of traversing and climbing on discontinuous terrain where multi-limbed multi-contact locomotion is required for effective traversal. This project will bring together the best of traditional control dynamics and mechanism design with state of the are machine learning techniques to design and implement a truly acrobatic multi-legged robot for traversing unstructured terrain.

Supervisors

Navinda Kottege, Tirthankar Bandyopadhyay (CSIRO's Data61) and Ian Manchester (University of Sydney)

Location

Pullenvale, QLD

Sharing knowledge about a certain environment is something humans do on a regular basis. It is common for someone to call a friend providing information about the status of the road they have just driven by: “watch for snow” or “careful with potholes”. That information is usually processed by the recipient of the call, and they adapt their driving behaviour accordingly, so there are no “surprises”. Similarly, when a driver sees a “Kangaroo” sign on the road, they know they should adapt their speed and heighten awareness, even if no kangaroo is currently present or “sensed” at all in the surroundings. This PhD project aims to apply this analogy of high-level knowledge sharing to robots.

Supervisors

Paulo Borges, Jason Williams and Tirthankar Bandyopadhyay (CSIRO's Data61)

Location

Pullenvale, QLD

Globally localising an agent within a map is key to autonomous robotics. Also, reliably finding the robot's current location can be used in Simultaneous Localisation and Mapping (SLAM) to close loops in revisited areas to deal with drift. Place Recognition (PR) approaches, using deep learning, show impressive results in finding the robot's location in a map database without requiring any prior knowledge. However, these PR approaches are based on heuristic supervised learning which restricts the use of the approach in different environments.
This PhD project will jointly address the weaknesses of the existing PR methods to produce a self-supervised, adaptive and reliable PR model that can learn during deployment.

Supervisors

Milad Ramezani, Dimity Miller and Peyman Moghadam (CSIRO's Data61)

Location

Pullenvale, QLD

Soft and flexible robots are ideally suited to agricultural applications like fruit picking, pest management and animal husbandry. Their soft materials allow them to conform to the shape of target objects and robustly grasp them without damage. This project looks to provide these capabilities for the design of soft robotic components that safely and efficiently perform on-farm tasks.

Supervisors

Josh Pinskier and David Howard (CSIRO's Data61) in collaboration with Michael Wang and Chao Chen (Monash University)

Location

Pullenvale, QLD

3D scene understanding and 3D mapping are core topics in the field of computer vision. Specifically, current approaches to creating large scale maps typically rely on a single, or at most a few, high-end, expensive sensors that scan an area in a relatively controlled and restricted way. What this project attempts to explore is to radically reduce the cost of each individual sensor, and the platforms that carries them, and instead significantly increase the number of sensors/platforms used.

Supervisors

Lars Petersson, Olivier Salvado, Kasra Khosoussi (CSIRO's Data61) and Hongdong Li (ANU)

Location

Canberra, ACT or Pullenvale, QLD

This PhD project will focus on developing fundamental collaborative machine perception capabilities such as Collaborative Simultaneous Localization and Mapping (CSLAM) for a fleet of heterogeneous robots operating in large-scale unknown GPS-denied environments. As part of this project, new decentralized and distributed optimization methods will be designed for collaborative perception tasks in challenging time-varying communication regimes. Our work will enable robots to efficiently share and learn from self and peer experiences, while seamlessly adapting to and leveraging heterogeneity in resources and sensing modalities. The algorithms developed in this project will be implemented and tested on CSIRO Data61's fleet of ground and aerial robots.

Supervisors

Kasra Khosoussi, Navinda Kottege and Paulo Borges (CSIRO's Data61)

Location

Pullenvale, QLD

On-device machine learning (ML) is rapidly gaining popularity on mobile devices. Mobile developers can use on-device ML to enable ML features on users’ mobile devices, such as face recognition, augmented virtual reality, voice assistance, and medical diagnosis. Compared to cloud-based machine learning services, on-device ML is privacy-friendly, of low latency, and can work offline.
On-device ML requires models to be deployed at the local mobile devices, thereby inevitably creating a requirement of model compression and security to prevent new attack surfaces. This project aims to combine the adversarial robustness, model compression and our novel compression encryption into one task.

Supervisors

Shuo Wang and Jason Xue (CSIRO's Data61)

Location

Marsfield, NSW

Quantum computing (QC) can be a game-changer in fields such as, cryptography, chemistry, material science, agriculture, and pharmaceuticals once the technology is more mature. Briefly, the quantum supremacy of a QC is demonstrated by its ability to quickly solve a problem that never has been solved by a classical computer, e.g., breaking TLS encryptions. Quantum supremacy requires a reliable set of measurements but in quantum mechanics, measurement is, famously, an inherently destructive process.

Supervisors

Meisam Mohammady, Dongxi Liu and Muhammad Usman (CSIRO's Data61)

Location

Marsfield, NSW

This project aims to develop a novel synthesis of physical and computational approaches to ML powered cyber-physical system security. Synthesis of computation, communication and control models with machine learning for practical relevance. Software plays a fundamental role in modern control (e.g., cruise control, self-driving cars), communication (e.g., 6G), and power systems (e.g., DER). However, complex software architectures used in real-world critical infrastructure are usually not captured by abstract system-theoretic models. This project will combine data from (simulation of) critical cyber-physical systems with principles and strategies derived from control and coding theories.

Supervisors

Mohammed Bahutair and Jason Xue (CSIRO's Data61)

Location

Marsfield, NSW

Recent advancements in federated learning (FL) allow multiple parties to collaboratively train to learn an artificial intelligence (AI) model without sharing raw data. However, FL is challenging in real-world applications because datasets may differ in both the sample and feature spaces. Transfer learning is an effective way to solve the difficulty of data annotation by transferring knowledge from a related source domain to the target domain. In this project, we aim to address the challenges in dealing with heterogeneous data from multiple sources to perform model training safely and efficiently without violating data privacy and confidentiality.

Supervisors

Ming Ding, Thilina Ranbaduge and Thierry Rakotoarivelo (CSIRO's Data61)

Location

Eveleigh, NSW

This project will advance academic knowledge and industry practice guidelines in the development and deployment of AI-based Learning Analytics (LA), which adheres to the principles of Responsible and Trustworthy AI. This project will focus on this technical side and provide designs and approaches to orchestrate effective, ethical, and trustworthy incorporation of AI Learning Analytics. More specifically, this project will research methods that i preserve the utility of AI-based LA methods in concrete industry scenarios, while providing provable assurance on principles such as Privacy, Fairness, Explainability and or Robustness

Supervisors

Thierry Rakotoarivelo (CSIRO's Data61)

Location

Eveleigh, NSW

Artificial intelligence (AI) has gained significant attention because of the achievements of machine learning (ML) and deep learning algorithms that rapidly accelerate research and transform practices in multiple fields, including health, agriculture, cybersecurity, and advanced manufacturing. Given the constraints of data sharing, distributed learning has emerged as a strategy for effective collaboration between data owners while enhancing privacy, ensuring governance, and complying with regulatory aspects. Distributed ML techniques enable machine learning without directly accessing raw data, which can often be personal and sensitive, held by clients such as hospitals or end devices such as the Internet of Things (IoT). Nevertheless, privacy remains to be a significant issue for distributed ML since the shared information in distributed learning can still reveal certain knowledge of the underlying data.

Supervisors

Thilina Ranbaduge, Ming Ding and Youyang Qu (CSIRO's Data61)

Location

Eveleigh, NSW

In food supply chains – capturing, interpreting, and disseminating data analytics from farmer to food-plate, is a vital digital solution for the maintenance and growth of Australian agriculture. For trusted supply chains – to enable participation, provision, and dissemination of analytics – a key driver is preserving privacy/confidentiality of participants (such as farmers, processors, and distributors) while providing improved utility for all stakeholders. Building on existing pilot work this PhD project will design for, and investigate, such trusted supply chains for a range of primary produce important to the Australian agriculture sector, such as, grains, legumes, honey, and beef.

Supervisors

David Smith (CSIRO's Data61) in collaboration with Robert Barlow (CSIRO Agriculture and Food)

Location

Eveleigh, NSW

In this project, we aim to systematically analyse the privacy leakage of the training data and/or the model itself from two popular generative models, i.e., generative adversarial network (GAN) and variational autoencoder (VAE). Various assumptions will be considered, such as whether it is explainable (black-box or white-box), which granularity of auxiliary information the attacker has access to, etc. Based on the analytical results, we then work to develop a privacy-preserved deep generative model, which mitigates the privacy inference attacks. Upon completion of this project, we aim to deliver a trustworthy and privacy-preserved GAN/VAE network and may extend to other application domains such as digital manufacturing. This will fundamentally solve the critical issue of lack of data in the mentioned application domains.

Supervisors

Youyang Qu and Ming Ding (CSIRO's Data61)

Location

Eveleigh, NSW

The use of Automated Decision Systems (ADS) is increasing at a rapid pace in both Government and non-Government contexts. The importance of incorporating responsibility into ADS is broadly recognized to guarantee fairness, accountability, interpretability and transparency in the socio-legal-technical systems. In the EU and Australia, the General Data Protection Regulation (GDPR) and the Australian Human Rights Commission are introduced in recently years. In research communities, significant efforts have been dedicated to the responsible development of ADS. Nevertheless, existing efforts often focus on the last mile of data analytics, namely, on ensuring responsibility in model designing and deployment. The key question is how to embody trust and responsibility into the holistic lifecycle of data management and analytics in ADS? In this project, we aim to answer this question and identify critical opportunities of improving data representativeness and controlling bias.

Supervisors

Sherry Xu, Qing Liu (CSIRO's Data61) in collaboration with Wenjie Zhang (UNSW)

Location

Eveleigh, NSW

Australia is currently facing a huge challenge in addressing natural disasters such as bushfires, floods, and the ongoing pandemic. These events have high social, financial, environmental and human costs. Australian Government and organisations are increasingly relying on investing in technological solutions to better prepare for unprecedented crisis events. It is predicted that by 2025 almost 21.5 million Australians will have a smartphone. The diversity of utilities, simplicity of use, ease of access, personalisation, ubiquity and flexibility of mobile technologies and apps make them a valuable tool in current times. However, these apps must work within a socio-technical ecosystem during a crisis event where users’ diversity and their social attitude toward the technology play almost as critical a part as the technical excellence and effectiveness of the app itself.

Supervisors

Didar Zowghi (CSIRO's Data61) in collaboration with Chetan Arora (Deakin University) and Muneera Bano

Location

Eveleigh, NSW or Clayton, Victoria

With sheer amount of software vulnerabilities and high dependency on third-party libraries, traditional rule-based or human-forced vulnerability detection approaches are challenged by heterogeneous data resources, complex library dependencies and conflicts, limited vulnerability support, and heavy human workloads that cannot ensure the security of complex cyber systems. The development of machine learning has improved automation and effectiveness of vulnerability detection to a new level. However, the black-box machine learning suffers from high false positive rate with unexplainable and unreliable detection results. Security experts still need to manually find key clues of the vulnerability to validate AI outcomes.
Knowledge graph opens a new door to solve the problems. Its structure can efficiently integrate heterogeneous resources from different databases, supporting automatic knowledge induction, threat modelling, knowledge retrieval and risk analysis. Besides, knowledge graph provides direct visualization to clients, offering explainable information to AI decisions that improve trustworthiness and responsibility of AI outcomes.

Supervisors

Jiamou Sun and Zhenchang Xing (CSIRO's Data61)

Location

Eveleigh, NSW

Unmanned Aerial Vehicles (UAVs) or drones, are flying mini robots and have been widely deployed in various scenarios (e.g., surveillance, delivery) because of their features of pilotless, a small physical shape, and fast speed. When implementing the flight code, developers not only design the customized flight code, but also invoke the APIs provided in the Dronekit, a development kit containing APIs developed by third parties to support advanced use cases including computer vision, path planning, 3D modelling, etc. Although the Dronekit provides a lot of convenience for developers, it introduces the potential maintenance issue, which can cause security problems. This project aims to address the security issues of UAVs from two perspectives: Dronekit vulnerability repair and closed-source emulation

Supervisors

Zhi Zhang, Surya Nepal (CSIRO's Data61) in collaboration with Siqi Ma (UNSW)

Location

Marsfield, NSW

As data exists in different modalities in the real world, viable interactions and combinations among multimodal data feature the creation and discernment of multimodal information in deep learning research. However, pre-trained large multimodal models can often carry more information (because they have much stronger data representation ability) than single modal models and they are usually applied in sensitive scenarios such as medical report generation and disease identification. Thus, multimodal models may lead to severe data privacy problems, for example, the attacker can recover the patient information from pre-trained models. This project studies the privacy leakage of large-scale pre-trained multimodal models through the lens of membership inference attack, the process of determining whether a data record belongs to the training dataset of a model or not.

Supervisors

Jason Xue (CSIRO's Data61)

Location

Marsfield, NSW

Causality plays an integral role in all forms of decision making, particularly it is important for responsible decision making by machine learning systems. Whatever we consider the potential effects of our decisions, we are thinking about cause. Causal inference goes beyond making associations and observations about what happens in our world. The importance of causal inference for making informed decisions has long been recognised in health, medicine, social sciences and other domains. The availability of “big data” in today’s world further presents opportunities to unleash the power of causal analysis to transform decision-making systems.
However, in many cases, it is not clear what data should be used for analysis, let alone how they are suitable. Also, data may come from different sources, with different formats (e.g., texts, images and relational tables). This project aims to address these two challenges and support causal analysis in heterogeneous and dynamic settings through a knowledge-driven causal framework.

Supervisors

Yanfeng Shu and Chen Wang (CSIRO’s Data61)

Location

All sites including Sydney, Hobart, Melbourne, Canberra, Adelaide and Brisbane

Uncertainty modelling describes what a machine learning (ML) model does not (and does) know, and measures goodness of prediction. This project will study novel uncertainty estimation methods for deep learning (DL) in the Lifelong Learning system and tackle the forgetting issue with guidance from uncertainty. Students participating in this project will develop novel models with efficient estimation methods for describing and predicting the uncertainties for data and DL models, considering the scalability and efficiency in real-world applications. The students will then design the uncertainty-guided lifelong/continual learning system in various application scenarios.

Supervisors

Sally Cripps (CSIRO’s Data61) in collaboration with Dong Gong (UNSW) and Lina Yao.

Location

Eveleigh, NSW

Machine learning at the edge (ML@Edge) aims to bring the capability of running ML models locally to edge devices. It is important for many scenarios where raw data is collected from sources far from the cloud. However, Deep neural networks (DNNs) are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. This project will tackle the problems by developing efficient deep learning methods, i.e., pruning, factorization, quantization as well as compact model design. Students participating in this project will develop efficient deep learning models for devices with limited resources.

Supervisors

Sally Cripps (CSIRO’s Data61) in collaboration with Xiaojun Chang (University of Technology Sydney) and Lina Yao.

Location

Eveleigh NSW

Graphical models are an important component in explainable Machine Learning. They are widely used to understand the interactions, even causal relationships, among complex systems in many real-world scenarios. In this project, we aim to design scalable and efficient algorithms of structure learning for large-scale datasets. The problem will be tackled by using a "divide-and-conquer" strategy, which is common and effective for scalable algorithm design. The students participating in this project will learn and master state-of-the-art structure learning algorithm for graphical models, and then develop new algorithms for structure learning.

Supervisors

Andy Zhu (CSIRO’s Data61) in collaboration with Robert Kohn (UNSW) and Lina Yao.

Location

Eveleigh NSW

In many applications for Bayesian inference, the probability density is distributed on a manifold that is embedded in an ambient space of higher dimension. Such distributions are usually caused by imposing constraints on the parameters. Only very recently, effective inference algorithms have been proposed for this setting and designing more robust inference tools is an open problem.
Another important and challenging problem in Machine Learning and Statistics is Structure learning i.e., estimating the topology of directed acyclic graphs (DAGs) a.k.a. Bayesian networks.
This PhD project links these two problems via converting the structure learning setting into Probabilistic inference on a continuous space. Due to the acyclicity constraint, the target probability density is distributed on a manifold and the inference is carried out by sampling from it.

Supervisors

Hadi Afshar and Yanan Fan (CSIRO’s Data61) with Minh Ngoc Tran (University of Sydney)

Location

Eveleigh NSW

Reproducibility and replicability (R&R) in scientific research are essential to validate results and confirm new knowledge. They are even more important when these results inform policy, future scientific studies, or health-based decisions. This PhD research will build the framework and structure from which an R-index can be constructed across research in Science, Engineering and Medicine. This work will be of global importance and potentially extremely significant impact.

Supervisors

Yanan Fan (CSIRO’s Data61) with Scott Sisson (UNSW).

Location

Eveleigh NSW

This project will consider ways in which expert knowledge can be included to guide the creation of topics groups in short texts; the incorporation of meta information; and the related issues of Bayesian model choice. For large scale problems, development of new computational methods will be required, particularly in cases for when new information becomes available. This project aims to address these two challenges and support causal analysis in heterogeneous and dynamic settings through a knowledge-driven causal framework.

Supervisors

He Zhao and Yanan Fan (CSIRO’s Data61) in collaboration with Tom Stindl (UNSW)

Location

Eveleigh NSW