Development and usage of self-organising maps in high energy physics analysis with high performance computing / Mohd Adli Md Ali

The Self-Organizing Map (SOM) was put forward by Teuvo Kohonen in 1982 as a computational technique to produce a set of globally ordered quantized vectors. At the present time, it is regarded as one of the primary machine learning techniques to perform unsupervised clustering analysis on a large variety of huge data. Implementation wise, the algorithm is also parallelizable to a large extent thus allowing it to scale up/down vertically and horizontally and its adaptable to the high-performance computing environment. Thus, development of an SOM algorithm for high energy physics datasets was performed. In this research, the effects of several SOM hyperparameters such as the similarity functions, learning rate functions and map size on the clustering outcome was also performed. Moreover, a test case on how the Kullback-Leibler divergence and Multivariate Bhattacharyya Distance equation can be used as a validation parameter for SOM is performed. Additionally, it is demonstrated that a classification model can be created by staking the SOM model with a Linear Discrimination Analysis model, and the performance of this model is compared with other classification models. A demonstration of unsupervised clustering of particle physics datasets with SOM and SOM+Dirichelet Gaussian Mixture Modelling was also carried out in this research