91���˰涶��

Research Themes

We developed a transdisciplinary program of research encircling methodologies and applications of data science. The methodologies include�� Bayesian deep learning, neuroevolution,�� ensemble machine learning, and data augmentation. The applications include climate extremes, geoscientific models, mineral exploration, medical diagnosis and bioinformatics, and COVID-19 drug re-purposing, infection forecasting, and social media with language models. Our key strength is in the development of novel deep-learning software frameworks with a focus on uncertainty quantification in decision-making. We pioneered the area of language models for studying ancient religious-philosophical texts.

Machine Learning

• Machine learning and deep learning: We developed novel neural network learning algorithms using neuroevolution with motivations from transfer learning and multi-task learning to a wide range of problems that include multi-step ahead and dynamic time series prediction (Chandra et al., 2017; Chandra et al., 2018) and modular pattern recognition for dynamic environments (Chandra and Cripps, 2018) with the goal of modular deep learning.��The challenge has been problems that have missing information, noise and inconsistencies in the organisation of data and these are major research directions for future studies.��In order to address class imbalance problems and limited training data, we used generative adversarial networks (GANs) with machine learning models for data augmentation (Sharma et al., 2022). Our current focus is on extending the method to extreme-value forecasting, and also combining it with Bayesian inference for uncertainty quantification in predictions via the posterior distribution.
• Bayesian deep learning:��We developed novel algorithms for Bayesian neural networks that feature parallel tempering MCMC and parallel computing in order to address computational challenges (Chandra et al., 2019}. We later extended this for deep learning models that feature millions of parameters that were “believed to be unachievable” in the MCMC sampling community. We developed Bayesian autoencoders using MCMC (Chandra et al., 2022) and Bayesian Graph convolutional neural networks (Chandra et al., 2022). Taking into account the benefits of evolutionary algorithms, we developed a framework that provides a synergy of multi-source transfer learning with Bayesian neural networks using MCMC (Kapoor et al., 2022). 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 MCMC sampling methods to provide uncertainty quantification in the data and the model space.

Earth and Climate Sciences

• Remote sensing and machine learning: There is a huge potential for remote sensing when combined with emerging machine learning and deep learning methods. We demonstrated that the extraction of geological lineaments from satellite data via remote sensing and machine learning can be used for mineral exploration (Farahbakhsh et al.,2020). In collaboration with the EarthByte Group, we developed deep learning models for lithological mapping via remote sensing (Shirmard et al.,2022), and applied the same technology for the detection of alteration zones for mineral exploration. Currently, we are using variational autoencoders and remote sensing for the identification of lineaments with novel clustering methods. We are planning to use geochemical datasets with spatial machine learning and data augmentation methods for prospectively of critical metals in Australia. Furthermore, we are currently developing an open-source framework for land-cover mapping in Fiji where we plan not only to publish a paper but to release online digital maps that can help in environment restoration initiatives. We plan to extend these methods for space exploration projects with the study of the surfaces of the Moon and Mars using satellite data. We also plan to apply these methods to environmental protection and conservation problems, such as quantifying damages to reefs after high-category cyclones and monitoring invasive species (eg. African tulips in Fiji).
• Machine learning for the GBR: The Great Barrier Reef (GBR) is the world’s largest coral reef system which is vital for a healthy ecosystem in Australia and the Pacific Ocean. In collaboration with the Geocoastal Research Group (University of Sydney), we used remote sensing and clustering methods for reef community mapping from the One-Tree Island of the Great Barrier Reef (Barve et al., 2023). Our interest is in the study of the geological development of the GBR going back thousands of years in time. We developed an open-source software framework for reef modelling (Bayesreef) that provides insights into coral platform growth and demise through time (Pall et al, 2020) using the Py-Reef-Core model. We developed a machine learning framework for processing coral reef drill core data (Deo et al. 2023). Currently, we are using novel deep learning models for the study of organisms on the deep sea floor in collaboration with the University of Sydney and Imperial College London. In future, we are interested to use machine learning and Bayesian inference methods in connecting landscape evolution models with reef evolution models to have a better understanding of the development of the GBR.
• Climate extremes and environmental problems: The drastic effect of climate change is visible given extreme weather conditions such as tropical storms and cyclones. We used deep learning models for forecasting cyclone wind intensity and trajectory for South Pacific and South Indian Oceans (Chandra and Dayal, 2015; Deo and Chandra, 2016; Chandra et al.,2016 ). We also focused on the rapid intensification of cyclones (Chandra, 2017) and uncertainty quantification in predictions using Bayesian neural networks (Deo and Chandra, 2019). We developed variational deep learning models for cyclone trajectory prediction with uncertainty quantification (Kapoor et al., 2023). 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) in collaboration with the Natural Hazards and Climate Risk group, Data 61. Furthermore, we used novel deep learning models for modelling Australian floods by taking into account precipitation and stream flow which led to an Honours thesis with potential publication. We are also combining hydrological models with deep learning models for modelling stream flow. We collaborated with UNSW Water Research Laboratory where we use deep learning models for groundwater modelling. Apart from climate extremes, we developed a framework that featured machine learning methods to predict precipitation that defines paleoclimate that spans up to millions of years in the past (Chandra et al., 2021). The data features a range of geological indicators including sedimentary deposits (coal, evaporates, glacial deposits). We addressed the challenges of missing values in the dataset and uncertainty quantification with Bayesian machine learning and developed paleo-maps of forests and vegetation that span 250 million years.�� In future work, we would like to use these maps and connect them with other paleoclimate studies.
• Bayesian geoscientific models: Bayesian inference has been a popular methodology for parameter estimation in geological and geophysical forward models, also known as geoscientific models. In collaboration with the EarthByte Group (University of Sydney), we developed Bayesian inference via MCMC framework for parameter estimation and uncertainty quantification for landscape evolution models (Bayeslands) to demonstrate landscape evolution in synthetic models that span thousands to million years demonstrating surface evolution based on different climate and environmental conditions (Chandra et al., 2019, Chandra et al., 2020). We envision that Bayeslands and Bayesreef will create a significant impact on the research community. We addressed the computational inefficiency of MCMC for large-scale problems by combining parallel computing features with a surrogate-assisted estimation of likelihood function that describes the plausibility of a model parameter value, given observed data (Chandra et al., 2020). We plan to use our MCMC sampling framework for estimating parameters in hydrological models combined with deep learning models.

Bioinformatics and Medicine

• Medical diagnosis and bioinformatics: In this area, the focus is on applying novel machine learning and deep learning models for medical diagnosis and bio-informatics that includes COVID-19. In NHMRC funded project, we developed an unsupervised machine learning framework for COVID-19 drug re-purposing (Bansal et al., 2023) where the goal was to down-select a small subset of drugs for COVID-19 clinical trials. We also used deep learning for COVID-19 infection forecasting in India (Chandra et al., 2022). Furthermore, we used deep learning models for respiratory rate prediction using bio-signals (Kumar et al., 2022). Currently, I am leading projects where we use deep learning models for skin cancer detection given imbalanced and limited training data. In future, we plan to use Bayesian deep learning for a wide range of medical diagnoses and health-related datasets where uncertainty quantification in model predictions plays a vital role in decision-making.

Media and Religion

• Language models for social media analyses: In this research direction, we used novel deep learning models to develop language models via social media to understand public behaviours in events such as COVID-19 (Chandra and Krishna, 2021).��Currently, I am leading a project that reviews anti-vaccine tweets during COVID-19 with sentiment analysis. I led a related project that used sentiment analysis to model US 2020 Presidential elections (Chandra and Saini, 2021). Currently, we are working towards the analysis of political leaning (left vs right wing) in popular news media reporting using pre-trained deep learning-based language models. In the future, we plan to evaluate the right and left-wing biases of developed and emerging economies and study how it impacts political and economic development. We are currently developing models to study the emergence of Hinduphobia in India and around the world during COVID-19 via Twitter using sentiment analysis.
• Artificial intelligence for religion:��There is an immense potential for data science methods and language models in areas of arts and humanities, particularly philosophy and religion. The Bhagavad Gita is a sacred Hindu philosophical text that is one of the most translated texts over the course of history. We used language models to analyse the sentiments uncovered with philosophical issues presented in the Bhagavad Gita and study the difference when comparing prominent translations (from Sanskrit to English) (Chandra and Kulkarni, 2022). We demonstrated that deep learning-based language models can be used to compare related texts, i.e. the Bhagavad Gita with Upanishads via topic modelling (Chandra and Ranjan, 2022). Currently, we are developing model for metaphor detection in the Bhagavad Gita and Sermon of the Mount of the Holy Bible. We are also using sentiment analysis for the comparison of translations of the Sermon of the Mount. In collaboration with Midam Charitable Trust in India, we evaluated Google Translate Sanskrit using Bhagavad Gita translations for comparison (Shukla et al., 2023). Currently, our focus is on large language models such as Chat-GPT for developing Vedanta-GPT. This model will be trained with Bhagavad Gita and Upanishads that will be released as an online discussion chat-bot for Hindu philosophy. We are also developing language models for humour detection in Hindu texts, and plan to use related language models for Buddhist texts. In future, we plan to use computer vision methods for the automatic archival of ancient sacred scripts in India in collaboration with the Oxford Centre for Hindu Studies.

��

Economics and Society

• Deep learning for forecasting future economic trends:��We developed Bayesian neural networks that feature uncertainty quantification in forecasting future trends in the stock price of selected markets (Chandra and He, 2021). This led to a project that used recurrent neural networks for decadal economic forecasting focusing on country-wise GDP growth (Wang et al., 2023). Currently, we are developing novel deep learning models for the decadal cost of living outlook and variational deep learning for uncertainty quantification in stock price forecasting. We also used 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.

��

Available Research Projects��

1. Remote sensing for landscape detection on Mars (Honours/PhD) with Prof. Cedric John (Imperial College London)��
2. Detection of hazards from cyclones using��remote sensing and deep learning��(Masters/Honours)�� with Dr Ehsan Farahbakhsh��(University of Sydney)
3. Bayesian deep learning for hydrological models (Honours/PhD) with Prof. Willem Vervoort (University of Sydney)
4. Explainable recurrent neural��networks��(Honours/PhD) with Prof. Christian Omlin (University of Agder, Norway)
5. Language models����for the analysis of philosophical traditions emerging from Hindu texts (Honours/PhD)�� with Dr Bjarne Wernicke-Olesen (Oxford Centre for Hindu Studies, Oxford University)
6. Explainable artificial intelligence (Honours/PhD) with Prof. Giles Hooker (University of Pennsylvania, United States)
7. Bayesian deep learning with incomplete information (Honours/PhD) with Prof. Scott Sisson (UNSW)
8. Bayesian deep learning for crowd simulation and robot path prediction (Honours/PhD) with Dr William Zhi (Carnegie Mellon University, United States)
9. Indoor path navigation for disabled persons in large buildings��(Masters/Honours)��with Dr William Zhi (Carnegie Mellon University, United States)
10. Bayesian deep learning for robot path planning (Honours/PhD)�� with Prof. Bibhya Sharma (University of the South Pacific, Fiji)
11. Data augmentation and deep learning��(Honours/PhD)�� with Dr Anurag Sharma (University of the South Pacific, Fiji)
12. Deep learning for medical diagnosis��(Honours/PhD) with Dr Amit Krishan Kumar (Da Nang University, Vietnam)
13. Decadal economic trends with deep learning (Honours/PhD)�� with Dr Rodney Beard (FFA, Solomon Islands) and Dr John Hawkins (Sydney)
14. Sentiment analysis with deep learning during natural disasters and extreme events (Honours/Masters/PhD)��
15. Deep learning for protein function detection (Honours/PhD) with Prof. Alok Sharma (RIKEN, Japan)
16. Plate tectonics for mineral prospecting with machine learning����(Honours/PhD) with ��Prof. Dietmar Muller��(University of Sydney)
17. Machine learning for reef modelling and optimisation�� (Honours/PhD) with��Prof. Jody Webster��(University of Sydney)

��

Seminar Presentation

1. R. Chandra, ``Natural language processing for philosophy of religion: Analysis of translations, sentiments and topics in religious texts", ��and ��Centre of Data Analytics and Cognition,�� LaTrobe University (August 2022), IIT Guwahati (November 2022),�� and University of the South Pacific (February 2023).
2. R. Chandra, ``Unsupervised machine learning framework for discriminating major variants of concern during COVID-19", Lymphatic Medicine Laboratory, Monash Institute of Pharmaceutical Sciences, Monash University, August 2022.
3. R. Chandra, ``Global COVID-19 Twitter dataset and language models for sentiment analysis and topic modelling", Melbourne University (August 2022),�� University of Tasmania (September 2022), and IIT Bombay (December 2022).
4. ��R. Chandra and Ehsan Farahbakhsh, ``Machine learning for mineral exploration: a data odyssey", CSIRO Mineral Exploration, Perth (September 2022) and Fiji National University (February 2023).
5. R. Chandra, ``Bayesian deep learning via MCMC sampling with application to robot path planning", Robotics and Automation Group, Queensland University of Technology, Brisbane, September 2022. ��
6. R. Chandra, “Machine learning for paleo-geology and mineral exploration: A spatiotemporal odyssey”, ARC ITTC Data Analytics in Resources and Environments, December 2021.
7. R. Chandra, “BERT-based language models for US Elections, COVID-19, and Bhagavad Gita”, UNSW Statistics Seminar Series, December 2021.
8. R. Chandra, “Revisiting Bayesian deep learning with advancements in MCMC”, University of Auckland, Department of Statistics, April 2021.
9. R. Chandra, “Unravelling Earth’s geological history with geoscientific models powered by artificial intelligence” University of the South Pacific, Public Seminar Series, September 2019.
10. R. Chandra, ``Bayesian inference for Geoscientific models'', School of Computer Science, University of Wollongong, February 2019.
11. R. Chandra, ``Bayesian inference for computationally expensive Earth evolution models'', School of Computer Science, University of Adelaide, October 2018.
12. R. Chandra, ``Bayesian inference for modelling geo-coastal, basin and landscape evolution'', Basin Genesis Hub Workshop, The University of Sydney, February 2018.
13. R. Chandra, `` Tackling climate change problems with machine learning '', EarthByte Group, School of Geosciences, The University of Sydney, July 2017.
14. R. Chandra, ``Competitive neuroevolution with applications'', Seminar, School of Computing, Information and Mathematical Sciences, University of South Pacific, August 2015.
15. R. Chandra, ``Open source software for education in Fiji'', Seminar, South Pacific Computer Society, University of South Pacific, Suva, Fiji, April 2013.
16. R. Chandra, ``Chaotic time series prediction using recurrent networks", Seminar, ��Victoria University of Wellington, Wellington, New Zealand, August 2011.

��