Abstract
The thesis aims to present Graphics Processing Units (GPUs) as a tool that can significantly reduce compute-time in ecological statistics applications. As serial computation speed approaches theoretical limits, parallelism offers an opportunity for performing computationally expensive statistical analyses at reduced cost, energy consumption, and time---attributes of increasing concern in environmental science and further afield. Accelerated inference is necessary to leverage datasets of increasing size in the modern and digital world, to make complex machine learning architectures viable, and to allow models to be extended to better reflect real-world environments.However, highly parallel computing architectures require different styles of programming and are therefore less explored, despite their known potential. GPU-specific programming principles are discussed in detail alongside examples of best practices. Code optimisations and algorithm design concepts that allow for efficient use of GPUs are introduced.
This thesis demonstrates the benefits of many-core parallelism with two GPU-accelerated implementations of algorithms in statistical ecology.
The first case study focuses on parameter inference for a Bayesian grey seal (Halichoerus grypus) population dynamics state space model, via particle Markov chain Monte Carlo. A speedup factor of over two orders of magnitude is achieved, providing an efficient alternative to current state-of-the-art fitting algorithms.
The second case study involves spatial capture-recapture, an animal abundance estimation framework. A simulation study demonstrates that a speedup factor of two orders of magnitude is possible when the number of detectors and integration mesh points is high. Following this, common bottlenose dolphin (Tursiops truncatus) photo-ID data are analysed to illustrate the efficacy of GPU-accelerated SCR using less real-world observations. A speedup factor of 20 was achieved compared to using multiple CPU cores and open-source software for the same task.
Date of Award | 3 Dec 2024 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Len Thomas (Supervisor) |
Keywords
- GPU
- pMCMC
- Spatial capture-recapture
- CUDA
- Particle filter
- Many-core parallelism
Access Status
- Full text open