Dr Rohitash Chandra

Methodology Research

1. Bayesian deep learning: Markov Chain Monte Carlo (MCMC) methods provide a probabilistic approach for the estimation of the free parameters in a wide range of models. Parallel tempering is an MCMC method that features parallelism with enhanced exploration capabilities. We have developed a Bayesian neural network framework that features parallel tempering MCMC with parallel computing [1]. We have addressed the challenge of applying MCMC methods for deep learning network architectures that features millions of parameters with Bayesian Autoencoders [2] and Bayesian Graph Convolutional Neural Networks [3]. We have also presented a framework that provides a synergy of multi-source transfer learning with Bayesian neural networks powered by MCMC [4]. Our current focus is on their application to other deep learning methods, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-term Memory (LSTM) networks. 
2. Generative adversarial networks: The major challenge is to develop machine learning models given a low number of training examples. In this area, we use generative adversarial networks (GANs) with machine learning models to generate data in scenarios with space and limited data. Our current focus is on pattern classification problems [5],  but the method can be used for spatiotemporal problems, and also augmented with Bayesian inference for robust uncertainty quantification. 
3. Surrogate-assisted and Bayesian optimisation: Surrogate-assisted optimization considers the estimation of an objective function for models given computational inefficiency or difficulty to obtain clear results. Surrogate-assistance inference addresses the inefficiency of parallel tempering MCMC for large-scale problems by combining parallel computing features with surrogate assisted estimation of likelihood function that describes the plausibility of a model parameter value, given specific observed data [6][7].  The challenge is to have a good estimation by the surrogates when the actual model features hundreds of free parameters.  
4. Neuroevolution and modular learning algorithms: Neuroevolution features evolutionary algorithms that provide a gradient-free and black-box approach to learning in neural networks. Hence, the learning algorithm is not constrained to the architecture of the network and does not face the limitations of gradient descent such as local minima and vanishing gradients. We have developed novel neural network learning algorithms using neuro-evolution with motivations from transfer learning, multi-task learning and reinforcement learning [8] [9] [10] [11].  The challenge is in problems that have missing information, noise and inconsistencies in the organisation of data.   

Applications

1. COVID-19 drug repurpose and modelling: We focus on long-COVID19 and vaccine testing using machine learning methods such as Bayesian optimisation and deep learning.  The goal is to repurpose drugs for COVID-19 using human tissue and organoid models in an ongoing project funded by the NHMRC. We use artificial intelligence and machine learning (AIML) techniques to characterise the systems biology responses better in a follow-on project with the US FDA. This brings together deep learning and tissue/organoid models along with experimental design using Bayesian optimisation. The approach can be extended to other diseases and enable AIML-based management of future pandemics.
2. Cyclone modelling and prediction: The drastic effect of climate change is visible with extreme weather conditions such as tropical storms and cyclones. In this research, we use machine learning methods for forecasting cyclone formation for decades to come given drastic changes in the climate. We use global circulation models with machine learning methods to estimate cyclone categories decades ahead in the future. 
3. Language models: In this research, we use novel deep learning models to develop language models via social media to understand public behaviours in events such as COVID-19.  We have applied deep learning-enabled language models for COVID-19 sentiment analysis with a case study of the Indian first wave of the pandemic. Currently, we are extending the methodology with topic modelling for comparing the three major waves of COVID-19 along with emerging topics in India. Furthermore, we are also reviewing anti-vaccine tweets during COVID-19 and sentiments related to them as the peak was reached in the first and the second wave around the world. Deep learning-based language models with sentiment analysis have also been used to model US 2020 Presidential elections. 
4. Artificial intelligence for philosophy of religion: It is well known that artificial intelligence methods have immense success in their applications in areas of science and technology. It is important to uncover the potential of these methods in areas of arts and humanities. The Bhagavad Gita is a Hindu sacred and philosophical text which has been one of the most translated texts over the course of history. We use artificial intelligence methods to analyse the sentiments uncovered with philosophical issues presented in the Bhagavad Gita. We show that artificial intelligence methods powered by deep learning can be used to guide the study of religious and philosophical texts. We show that artificial intelligence can be used for understanding sentiments expressed in ancient philosophical texts. We use novel language models to analyse selected translations of the Bhagavad Gita (from Sanskrit to English) using semantic and sentiment analyses which help in the evaluation of translation quality.
5. Mineral exploration and remote sensing: The extraction of geological lineaments from digital satellite data is a fundamental application in remote sensing. We use computer vision techniques for extracting geological lineaments using optical remote sensing data. Furthermore, in another research direction, we provide a synergy of deep learning methods with remote sensing for lithological mapping which is a critical aspect of geological mapping that can be useful in studying the mineralization potential of a region. Currently, we are using variational autoencoders and remote sensing for the identification of lineaments along with novel clustering methods. We would like to extend these methods for space exploration projects with the study of the Moon and Mars using satellite sate.
6. Reef modelling and remote sensing: Geological reef models such as Py-Reef-Core provide insights into the flux of carbon by analysing carbonate platform growth and demise through time, and modelling their evolution using landscape dynamics and reef modelling. We estimate and provide uncertainty quantification of free model parameters using Bayesian inference with Py-Reef-Core. This can help us understand reef evolution on a geological timescale that can help in predicting the future evolution of coral reefs. The challenge here is in the estimation of the parameters which involves highly non-separable and constrained optimisation.  Currently, we are utilising remote sensing and machine learning method to study reef areas using satellite and drone datasets.
7. Solid Earth evolution: Bayesian inference has been a popular methodology for the estimation and uncertainty quantification of parameters in geological and geophysical forward models. Badlands is a basin and landscape evolution forward model for simulating topography evolution at a large range of spatial and time scales. Our solid Earth evolution projects consider Bayesian inference for parameter estimation and uncertainty quantification for a landscape evolution model (Bayeslands). The challenge is in parameter estimation for computationally expensive models which are being addressed by high-performance computing and surrogate-assisted Bayesian inversion.  
8. Paleoclimate reconstruction: The reconstruction of paleoclimate precipitation can provide light to Earth’s climate history of millions of years in the past. Although global circulation models have been used with success for the reconstruction of precipitation in the Miocene period, their application to an era back in time is a major challenge due to limited data. We use an alternate approach that features machine learning methods to predict precipitation that defines paleoclimate that spans up to 400 million years in the past. The data features a range of geological indicators including sedimentary deposits (coal, evaporates, glacial deposits). The challenge has been in addressing missing values in the dataset and providing rigorous uncertainty quantification in order to develop paleo-maps of forests and vegetation.   

Seminars

1. R. Chandra, “Machine learning for paleo-geology and mineral exploration: A spatiotemporal odyssey”, ARC ITTC Data Analytics in Resources and Environments, December 2021. Youtube
2. R. Chandra, “BERT-based language models for US Elections, COVID-19, and Bhagavad Gita”, UNSW Statistics Seminar Series, December 2021. Youtube
3. R. Chandra, “Revisiting Bayesian deep learning with advancements in MCMC”, University of Auckland, Department of Statistics, April 2021. Youtube
4. R. Chandra, “Unravelling Earth’s geological history with geoscientific models powered by artificial intelligence” University of the South Pacific, Public Seminar Series, September 2019. Youtube
5. R. Chandra, ``Bayesian inference for Geoscientific models'', School of Computer Science, University of Wollongong, February 2019.
6. R. Chandra, ``Bayesian inference for computationally expensive Earth evolution models'', School of Computer Science, University of Adelaide, October 2018. Youtube
7. R. Chandra, ``Bayesian inference for modelling geo-coastal, basin and landscape evolution'', Basin Genesis Hub Workshop, The University of Sydney, February 2018.
8. R. Chandra, `` Tackling climate change problems with machine learning '', EarthByte Group, School of Geosciences, The University of Sydney, July 2017. Youtube
9. R. Chandra, ``Competitive neuroevolution with applications'', Seminar, School of Computing, Information and Mathematical Sciences, University of South Pacific, August 2015. Youtube
10. R. Chandra, ``Open source software for education in Fiji'', Seminar, South Pacific Computer Society, University of South Pacific, Suva, Fiji, April 2013.
11. R. Chandra, ``Chaotic time series prediction using recurrent networks", Seminar, School of Engineering and Computer Science, Victoria University of Wellington, Wellington, New Zealand, August, 2011.

Available Research Projects 

1. Bayesian deep learning for protein function detection (PhD), Co-supervised by Prof. Alok Sharma (RIKEN, Japan)
2. Cyclone path and intensity prediction with deep insight based deep learning (Masters/Honours)
3. Indoor path navigation for disabled persons in large buildings (Masters/Honours) 
4. Detection of electric cable hazards from Cyclones using  drones and  remote sensing and deep learning (Masters/Honours) 
5. Dynamic Earth models, landscape dynamics and basin evolution (PhD), Co-supervised by Prof. Dietmar Muller (University of Sydney)
6. Machine learning for reef modelling and Optimisation, Co-supervised by Prof. Jody Webster (University of Sydney)
7. Deep learning for the reconstruction of 3D Ore-bodies, Co-supervised by  Dr Ehsan Farahbakhsh
8. Memory in Recurrent Neural Networks and Neural Turing Machines, Honours/PhD
9. Bayesian deep learning with incomplete information, Honours/PhD
10. Variational Bayes for Spatio-temporal modelling (Honours/Masters/PhD) with Prof. Robert Kohn
11. Bayesian deep learning for language models  (Honours/Masters/PhD)
12. Sentiment analysis with deep learning during natural disasters and extreme events (Honours/Masters/PhD)
13. Deep learning for monitoring abuse in social media  (Honours/PhD)
14. Ensemble learning for class imbalanced problems (Honours/PhD)  with Dr Rodney Beard 
15. Bayesian deep learning for hydrological models (Honours/PhD) with Prof. Lucy Marshall (UNSW) and A/Prof Willem Vervoort (University of Sydney)
16. COVID-19 related vaccine research  (Honours/PhD with Prof. Seshadri Vasan (CSIRO and University of York)
17. Knowledge-based Recurrent Neural Networks (Honours/PhD) with Prof. Christian Omlin (University of Agder, Norway)
18. Deep learning for bio-diversity and ecology (Honours/PhD) with Prof. Glenda Wardle and Dr Aaron Greenville (University of Sydney)

Access research papers: https://github.com/rohitash-chandra/research