research

2024

1. 

   Intelligent optimization of borehole geophysics planning for critical mineral exploration

   2024

   Abs PDF Code

   The transition to a carbon-free energy grid and transportation network hinges on accelerating the discovery of the battery metals: copper, nickel, cobalt, and lithium. Mineral explorers infer the presence of subsurface mineral deposits by relying on geophysical data collected above the surface. Airborne geophysical surveys offer expansive coverage and rapid access to vast, remote areas. Aircraft conducting geophysical surveys traditionally fly in straight lines at fixed intervals ranging from 50 m to 2 km. This comprehensive approach permits detailed geophysical maps but accumulates significant flying distances, which come at a substantial financial and time cost. In environmental monitoring, sequential data acquisition techniques that optimize locations for data collection while minimizing resource expenditure have become prevalent. These techniques allow mobile sensors to use observations to form a belief of the world, and in turn, use their current belief of the world to inform the path taken to collect observations in the future. The partially observable Markov decision process (POMDP) framework has demonstrated considerable promise in guiding path planning decisions in robotics and has recently been applied to subsurface applications where noisy observations are used to infer the underlying state. Building on this foundation, we propose a POMDP tailored to optimize the flight paths of fixed-wing aircraft in airborne geophysical surveys. We evaluate our approach using simulated geophysical maps, comparing its performance to traditional grid-based methods in terms of distance flown and resulting profitability. The results demonstrate that our approach can accurately and confidently estimate the true state of the world in significantly less survey distance. However, its performance is inconsistent across all scenarios, and therefore we suggest an alternative model formulation for future work.

2023

1. 

   Is Efficient-CapsNet the solution to medical imaging challenges in developing countries?

   2023

   Abs PDF Code

   State-of-the-art convolutional neural networks (CNNs) for medical imaging require computationally intensive training and large datasets, making them challenging to adopt in developing countries due to limited expert annotations and computing resources. Capsule networks, introduced by Hinton et al. (2011) and Sabour et al. (2017), aim to address CNN limitations, with Efficient-CapsNet leveraging self-attention to enhance routing while using fewer parameters. We explore its application in resource-constrained healthcare settings to bridge global healthcare inequalities. This study investigates Efficient-CapsNet’s generalization under typical medical dataset challenges, such as limited samples and no data augmentation during training. It significantly outperforms low-parameter CNNs and ResNet18 in classification accuracy and distributional generalization.

2020

1. 

   Shrinkage-driven unsupervised learning with clustering applications

   2020

   Abs PDF Code

   Model-based clustering has largely focused on mixtures where the component distributions correspond to the multivariate normal distribution. A detailed review of such mixtures, their estimation through expectation-maximization algorithm and an analysis of their comprehensive implementation mclust are all given here. Later, a new flexible family of regularized copula-based mixture models are introduced for cluster analysis, which are considered an extension of mixtures of copulas offered in recent literature. The technique sees a shrinkage-driven approach to selecting dependence structure permitted by unconstrained parameterization of the copula parameter using angles. An appropriate procedure for estimation through expectation-conditional-maximization (ECM), framework for model selection that maximizes cluster separation and in-depth discussion regarding computational details are all included. The new clustering technique is illustrated by the study of simulated and real data and is shown to outperform previous proposals. In addition, the report encompasses a description and documentation of the R package rcbmm, which facilitates the new methods and contains the first general ECM implementation available in R. Finally, limitations of the regularized copula-based mixture model and its implementation are outlined alongside a suggestion for future research directions.