Date of Award

12-2018

Degree Name

Doctor of Philosophy

Department

Computer Science

First Advisor

Dr. Ajay Gupta

Second Advisor

Dr. Ala Al-Fuqaha

Third Advisor

Dr. Alvis Fong

Fourth Advisor

Dr. Driss Benhaddou

Abstract

The massive amount of streaming data generated and captured by smart service appliances, sensors and devices needs to be analyzed by algorithms, transformed into information, and minted to extract knowledge to facilitate timely actions and better decision making. This can lead to new products and services that can dramatically transform our lives. Machine learning and data analytics will undoubtedly play a critical role in enabling the delivery of smart services. Within the machine-learning domain, Deep Learning (DL) is emerging as a superior new approach that is much more effective than any rule or formula used by traditional machine learning. Furthermore, it can even alleviate engineers from the task of defining features. This research explores the efficacy of DL in support of smart services. Despite recent advancements in DL for Internet of Things (IoT) smart services, there are still significant challenges that need to be addressed for this technology to mature. Hyper parameter tuning for DL models, besides the expensive computational cost and heavy memory overhead make DL unaffordable. In order to address these challenges, we propose a two-pronged solution.

Successful application of DL requires careful and sometimes very expensive hyper parameter searches, tuning, and testing. Manual parameter setting, and grid search are common approaches that can ease the users tasks for setting these important parameters values of the DL models. Nonetheless, these two approaches can be very time-consuming. Optimization methods therefore need to be used to help find optimal parameter settings. In this research, we show that the Particle Swarm Optimization (PSO) technique holds a great potential to optimize parameter settings; thus, saving valuable computational resources during the tuning process of DL models.

Another goal that we focus on during this research, is to investigate techniques to optimize execution of DL models on resource-constrained wearable IoT devices. Because of their overhead (e.g., memory, computation and energy), DL models are yet to become mainstream on mobile/IoT embedded platforms. Therefore, one of the main objectives of this research is to utilize cloud hosted Machine Learning as a Service (MLaaS) providers to collect big data from the resource-constrained devices and build prediction models on the cloud. Those prediction models are then sent to the underlying resource-constrained devices to be run locally without necessarily being always connected to the cloud. However, models’ trust represents a potentially serious threat in this paradigm. In this research, we propose a heuristic that maximizes the level of trust of DL models by selecting a subset of models from a superset of models. During each period, our proposed heuristic switches between the subset of selected models in order to maximize the trustworthiness while respecting given reconfiguration budget and rate. Due to the difficulty of the problem in real-world scenarios, we propose an intelligent real-time heuristic that can be used in large-scale deployments of IoT resource-constrained devices. The heuristic algorithm strives to make learning techniques more trustworthy since it avoids frequent reconfigurations. This also minimizes the communications overhead between the cloud and the resource-constrained devices. We prove that the competitive ratio of our proposed heuristic is O(1) when the reconfiguration rate is proportional to the total time and the reconfiguration budget is constant. Therefore, our proposed heuristic achieves an optimal competitive ratio in a polynomial time approximation scheme for the problem.

Access Setting

Dissertation-Campus Only

Restricted to Campus until

6-2019

Share

COinS