Files
Abstract
Deep learning technologies have demonstrated immense capabilities in numerous domains. In the current landscape, deep learning technologies play a key role in the success of modern businesses. However, adoption of deep learning technologies has a substantial amount of untouched potential. The cost of developing custom deep learning solutions for unique business problems is a major inhibitor to far-reaching adoption of deep learning technologies. We recognize that the monolithic nature prevalent in today's deep learning applications stands in the way of efficient and cost effective customized deep learning solution development. Taking a modular approach for deep learning solution development can yield a number of advantages that ultimately helps to make deep learning solutions more accessible and widespread.
This dissertation explores the benefits and trade-offs of developing deep learning solutions taking a modular approach as opposed to taking widespread monolithic approaches while addressing key challenges of modular deep learning solution development. Towards this end, we conduct experiments by implementing modular and monolithic solutions for representative deep learning problems and evaluate them in terms of accuracy, latency, reusability and maintainability. Our experiments show that modular solutions can be comparable to monolithic solutions in terms of accuracy while offering much desired benefits of modular solutions. However, the modular solutions often show higher latency compared to monolithic solutions. In our experiments, we show that this challenge can be mitigated by utilizing a black-box knowledge distillation approach. Furthermore, we propose a knowledge graph-based machine learning service description framework that further augment the benefits of modular machine learning solutions primarily by enhancing reusability and composability of existing machine learning modules. Finally, we propose a novel approach to develop supervised deep learning models in a modular manner enabling a number of solution engineering advantages that includes module reusability, module updates and replacement, transparency, debuggability and domain expert intervention for performance optimization while maintaining comparable accuracy and efficiency characteristics to the baseline models.
This dissertation explores the benefits and trade-offs of developing deep learning solutions taking a modular approach as opposed to taking widespread monolithic approaches while addressing key challenges of modular deep learning solution development. Towards this end, we conduct experiments by implementing modular and monolithic solutions for representative deep learning problems and evaluate them in terms of accuracy, latency, reusability and maintainability. Our experiments show that modular solutions can be comparable to monolithic solutions in terms of accuracy while offering much desired benefits of modular solutions. However, the modular solutions often show higher latency compared to monolithic solutions. In our experiments, we show that this challenge can be mitigated by utilizing a black-box knowledge distillation approach. Furthermore, we propose a knowledge graph-based machine learning service description framework that further augment the benefits of modular machine learning solutions primarily by enhancing reusability and composability of existing machine learning modules. Finally, we propose a novel approach to develop supervised deep learning models in a modular manner enabling a number of solution engineering advantages that includes module reusability, module updates and replacement, transparency, debuggability and domain expert intervention for performance optimization while maintaining comparable accuracy and efficiency characteristics to the baseline models.