Dr Rohitash Chandra

TECHNOLOGY DEVELOPMENT

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. It features a number of replicas with slight variations in the acceptance criteria. We have developed Bayesian deep learning methods that feature parallel tempering and parallel computing. The challenge is in the inference for deep learning network architectures that features millions of parameters, such as convolutional neural networks and LSTM neural networks. Collaboration: Prof. Scott Sisson and Dr Pavel Krivitsky (UNSW)

Meta-learning and 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 generative data in scenarios with space and limited data. The current focus is on pattern classification problems but the method can be used for spatiotemporal problems, and also augmented with Bayesian inference for robust uncertainty quantification. 

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 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.  The challenge is to have a good estimation by the surrogates when the actual model features hundreds of free parameters. Collaboration: Prof. Dietmar Muller,  University of Sydney; Prof. Yew Soon Ong, Nanyang Technological University, Singapore.

Neuro-evolution and learning algorithms: Neural networks are loosely modelled after biological neural systems and have a wide range of data-driven applications that include time series prediction and pattern recognition. Opposed to gradient-based methods, neuro-evolution features evolutionary algorithms that provide a 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.  The challenge is in problems that have missing information, noise and inconsistencies in the organisation of data. Collaboration: Prof. Yew Soon Ong, Nanyang Technological University; Prof. Christian Omlin, University of Agder, Norway.



APPLICATIONS

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 landscape dynamics 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. Collaboration: Prof. Dietmar Muller and Dr Tristan Salles, University of Sydney.

Reef modelling: Geological reef models such as Py-Reef-Core provides 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. Collaboration: A/Prof. Jody Webster and Dr Tristan Salles,  University of Sydney.

Mineral exploration: The extraction of geological lineaments from digital satellite data is a fundamental application in remote sensing. The location of geological lineaments such as faults and dykes are of interest in terms of mineralization. Although a wide range of applications utilized computer vision techniques, a standard workflow for the application of these techniques to mineral exploration is lacking. We use computer vision techniques for extracting geological lineaments using optical remote sensing data. Furthermore, in another research direction, we provide a synergy of geophysical forward models and Bayesian inference for 3D joint inversion for mineral prospecting and exploration. Collaboration: Prof. Dietmar Muller, Dr Ehsan Farahbakhsh 

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. Collaboration: Prof. Dietmar Muller, Dr Nathaniel Butterworth, and Prof. Sally Cripps,  University of Sydney.

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. 

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 also develop language models to analyse translations of ancient Hindu texts that include the Bhagawad Gita and the Upanishads.