Optimizing deep learning networks using multi-armed bandits

Deep learning has gained significant attention recently following their successful use for applications such as computer vision, speech recognition, and natural language processing. These deep learning models are based on very large neural networks, which can require a significant amount of memory and hence limit the range of applications. Hence, this study explores methods for pruning deep learning models as a way of reducing their size, and computational time, but without sacrificing their accuracy. A literature review was carried out, revealing existing approaches for pruning, their strengths, and weaknesses. A key issue emerging from this review is that there is a trade-off between removing a weight or neuron and the potential reduction in accuracy. Thus, this study develops new algorithms for pruning that utilize a framework, known as a multi-armed bandit, which has been successfully applied in applications where there is a need to learn which option to select given the outcome of trials. There are several different multi-arm bandit methods, and these have been used to develop new algorithms including those based on the following types of multi-arm bandits: (i) Epsilon-Greedy (ii) Upper Confidence Bounds (UCB) (iii) Thompson Sampling and (iv) Exponential Weight Algorithm for Exploration and Exploitation (EXP3). The algorithms were implemented in Python and a comprehensive empirical evaluation of their performance was carried out in comparison to both the original neural network models and existing algorithms for pruning. The existing methods that are compared include: Random Pruning, Greedy Pruning, Optimal Brain Damage (OBD) and Optimal Brain Surgeon (OBS). The thesis also includes an empirical comparison with a number of other learning methods such as KNN, decision trees, SVM, Naïve Bayes, LDA, QDA, logistic regression, Gaussian process classifier, kernel ridge regression, LASSO regression, linear regression, Bayesian Ridge regression, boosting, bagging and random forests. The results on the data sets show that some of the new methods (i) generalize better than the original model and most of the other methods such as KNN and decision trees (ii) outperform OBS and OBD in terms of reduction in size, generalization, and computational time (iii) outperform the greedy algorithm in terms of accuracy