Research Directors: Prof. Christian Omlin and Dr John Hawkins
We are interested in developing the next generation of machine learning models that address the grand challenges in artificial intelligence, including the development of models for sparse and missing data, uncertainty quantification in predictions, class imbalance and extreme value forecasting. We are currently extending extreme-value forecasting deep learning model with Bayesian inference for enabling uncertainty quantification in predictions. Furthermore, we are developing compact Bayesian neural networks through post-pruning via a Monte Carlo approach. Particularly, we are keen in data imputation, particularly uncertainty quantification in imputation models and also extension of such models in spatiotemporal data imputation. We plan to enhance the Bayesian deep learning models with data augmentation methods for multi-modal data fusion utilising a wide range of data streams. Furthermore, we plan to use a combination of variational inference and Monte Carlo sampling methods to provide uncertainty quantification in the data and the model space. We also envision incorporating explainable artificial intelligence and knowledge distillation techniques in our models.
Theme Leaders: Dr Rohitash Chandra and Prof Christian Omlin
Modern natural language processing deep learning models have accelerated research into the major fields of computational linguistics. The Transformer deep learning architecture has improved performance for tasks like sentiment analysis, document classification and named entity recognition. In addition, its generative abilities have improved translation and summarizations performance in a range of domains. Nevertheless, the improvements are not uniformly distributed across all languages and domains. In this research stream, we develop machine learning models for specific language applications and evaluate the performance of large language models for machine translation using sentiment and semantic analysis. We provide longitudinal study of social media and news media when it comes to political issues such as Sinophobia, Hinduphobia, Antisemitism, and anti-vaccination trends. Furthermore, we review political speeches in the United States Presidential elections in terms of sentiments, abusive language and personal attacks. We are keen in using large language models at specialised domains such as low-resource languages and area of humanities and social science that struggle with research funding. In addition, we explore the newer emergent themes in natural language processing, such as reasoning, deception, and world models implicit in text generation.
Theme Leaders: Dr Rohitash Chandra and Dr John Hawkins
We promote sustainable exploration practices, improving our understanding of Earth’s resources, and assessing the impacts of human activities on natural landscapes. We aspire to contribute significantly to responsible resource management and environmental stewardship through the integration of artificial intelligence and geosciences. Under this theme, we aim to harness advanced machine learning and deep learning models, to transform mineral exploration and environmental monitoring. We plan to develop innovative data driven methods that enhance the accuracy and efficiency of mineral discovery. We build on our previous expertise (data and models) in creating reliable multidimensional prospective models. Furthermore, we develop novel methodologies to innovate the processing of remote sensing data through machine learning and uncover deeper insights into geological formations and environmental change. We are keen to develop novel remote sensing tools driven by machine learning and deep learning methods that are not only limited to our planet, but also applicable to study remote sensing data for Mars and the moon. We are also interested in using machine learning and Bayesian inference methods for the application of landscape evolution models in studying the paleoclimate and geological history of the planet. Finally, we envision the usage of plate tectonic models such as GPlates for the acquisition of data for region-specific mineral productivity mapping.
Theme Leaders: Dr Ehsan Farahbukhsh and Dr Rohitash Chandra
The drastic effect of climate change is visible given extreme weather conditions such as tropical storms and cyclones. Currently, we are using deep learning for forecasting cyclone genesis in the coming decades given drastic changes in the climate via sea surface temperature data from the general circulation model (GCM). Furthermore, we developed novel conditional ensemble learning framework for modelling streamflow and flood prediction in selected Australian catchments using precipitation and environmental data. We are interested in using machine learning and Bayesian inference methods to integrate landscape evolution models with reef evolution models to have a better understanding of the development of the Great Barrier Reef and other reef systems around the planet. Finally, we envision robust remote sensing and machine learning models for environmental management and post-disaster management including landslides, cyclones, and floods.
Theme Leaders: Ratneel Deo and Dr Rohitash Chandra
We are interested in various biomedical data problems that can be solved using machine learning models. These problems range from pure molecular biology problems where we process data about specific biological molecules to predict properties such as function, movement, structure or interactions with other molecules. These problems contribute to fundamental research into drug discovery and the causal processes involved in diseases. At the other end of the spectrum, we look at biomedical data science problems that involve improving the processes by which human beings receive treatment. We are keen in building machine learning models that determine whether people are susceptible to complications during surgery, or short-term readmission following discharge from the hospital. These models help improve the efficiency of medical care and thereby help lift the quality of care. We envision the development of novel deep learning models that can incorporate uncertainty projections in predictions for medical data imputation and clinical decision making process. Furthermore, our focus is in the development of novel deep learning models with uncertainty projection in guiding medical practitioners with digital twin models.
Theme Leaders: Dr John Hawkins and Dr Rohitash Chandra
The challenges in economics intersect with machine learning in areas including forecasting, causal inference, and social choice theory with ensemble learning methods such as bagging. This also includes model learning paradigms in general, e.g. reinforcement learning, submodular function learning, and preference learning. Our multidisciplinary focus has led to a project that used recurrent neural networks for decadal economic forecasting, focusing on country-wise GDP growth. Currently, we are developing an ease of living index using data imputation and dimensionality reduction methods which will be later used for developing novel deep learning models for the decadal cost of living outlook. We are also developing a variational deep learning model for uncertainty quantification in stock price forecasting and multimodal deep learning that incorporates text and numeral data streams with large language models for credit rating forecasting. In future, there is scope for multimodal data fusion for forecasting decadal economic trends taking into account, energy utilisation, climate change, migration and estimates of poverty elimination. Applications include environmental and resource economics, climate change, distributive justice, and fair allocation.
Theme Leaders: Dr Rodney Beard and Dr Rohitash Chandra
The cinema and music industry has been used as an instrument to promote social change which has also been widely abused. Therefore, longitudinal study of the impact of cinema and music on society can be helpful in reforms and education policies. This is also a concern as social media and news media have been mediums influencing and disseminating information that had an impact on culture, social dynamics and politics. One of our projects focuses on evaluating the right and left-wing biases of developed and emerging economies and studying how they impact political and economic development. We envision the use of large language models and computer vision methods for the analysis of movies, scripts, and songs from the viewpoint of social sciences, economics and psychology. The focus is not just Hollywood, but world cinema, including the evolution of Indian cinema.
Theme Lead: Dr Rohitash Chandra
Artificial Intelligence is a field that forms one of the core spokes of cognitive science. It asks us to try and understand the mind by replicating some of its behaviour. As it progresses these questions encroach on the world of pure philosophy which has always engaged humanity in the question of understanding the nature of knowledge, experience, rationality and sentience. We explore the questions that are emerging from cutting-edge research, as well as the cross-cultural consideration of these technologies and their impact on society. Furthermore, we envision a multidisciplinary approach that looks into philosophical aspects of artificial intelligence from the viewpoint of machine consciousness, psychology, philosophy of mind, and comparative religion.
Theme Leaders: Dr John Hawkins, Elvin Prasad, and Dr Rohitash Chandra
Team: Rodney Beard, John Hawkins, Karan Sutariya
Determine whether Sobol sensitivity analysis and global sensitivity analysis or conformal prediction is more robust as an approach to uncertainty quantification for a variety of machine learning methods.
Team: Rodney Beard, John Hawkins, Yogendra Meena, Vishal Gupta
Develop a retrieval augmented generation solution for an expert system based on LLMs and the WCPFC conservation management measures for tuna fishing
Team: Rodney Beard
Conduct a large scale comparison of different diversity measures for feature selection when building machine learning systems. Include evaluation of explainable Artificial Intelligence (xAI) applications of diversity measures.
Team: John Hawkins
Extending work on building literature filtering systems to improve efficiency in systematic reviews. Selection of papers for deeper review on the basis of titles and abstracts using semantically sensitive machine learning models like pre-trained BERT systems.
Team: John Hawkins
Extending our work on GDP forecasting to use time lagged information about investment in energy infrastructure, energy mix and consumption trends to improve GDP forecasting.
Team: Aditya Pramar, John Hawkins
There is an emerging literature on methods used to circumvent the safety methods of LLMs and solicit undesirable content. These methods are called ‘jailbreaks’ and involve using certain phrases and structures inside a prompt to elicit the desired response from the LLM. This project involves experimenting with machine learning approaches to identify potential jailbreak attempts and analyse what drives them.
Team: Rodney Beard
Explore ways to implement submodular function learning. The topic relates to questions of fairness, distributive justice, diversity but from a machine learning perspective. Possible other applications relate to substitutability and complementarity of goods in markets.
Team: Rodney Beard
The Incas used mysterious stringy objects called ‘khipus’ to record data. This project is a data science/machine learning for scientific progress project with elements of mathematical language analysis.
Please apply here for a Research Internship.