Continual deep learning via progressive learning

Fayek, H 2018, Continual deep learning via progressive learning, Doctor of Philosophy (PhD), Engineering, RMIT University.


Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Fayek.pdf Thesis application/pdf 5.22MB
Title Continual deep learning via progressive learning
Author(s) Fayek, H
Year 2018
Abstract Machine learning is one of several approaches to artificial intelligence. It allows us to build machines that can learn from experience as opposed to being explicitly programmed. Current machine learning formulations are mostly designed for learning and performing a particular task from a tabula rasa using data available for that task. For machine learning to converge to artificial intelligence, in addition to other desiderata, it must be in a state of continual learning, i.e., have the ability to be in a continuous learning process, such that when a new task is presented, the system can leverage prior knowledge from prior tasks, in learning and performing this new task, and augment the prior knowledge with the newly acquired knowledge without having a significant adverse effect on the prior knowledge. Continual learning is key to advancing machine learning and artificial intelligence.

Deep learning is a powerful general-purpose approach to machine learning that is able to solve numerous and various tasks with minimal modification. Deep learning extends machine learning, and specially neural networks, to learn multiple levels of distributed representations together with the required mapping function into a single composite function. The emergence of deep learning and neural networks as a generic approach to machine learning, coupled with their ability to learn versatile hierarchical representations, has paved the way for continual learning. The main aim of this thesis is the study and development of a structured approach to continual learning, leveraging the success of deep learning and neural networks.

This thesis studies the application of deep learning to a number of supervised learning tasks, and in particular, classification tasks in machine perception, e.g., image recognition, automatic speech recognition, and speech emotion recognition. The relation between the systems developed for these tasks is investigated to illuminate the layer-wise relevance of features in deep networks trained for these tasks via transfer learning, and these independent systems are unified into continual learning systems.

The main contribution of this thesis is the construction and formulation of a deep learning framework, denoted progressive learning, that allows a holistic and systematic approach to continual learning. Progressive learning comprises a number of procedures that address the continual learning desiderata. It is shown that, when tasks are related, progressive learning leads to faster learning that converges to better generalization performance using less amounts of data and a smaller number of dedicated parameters, for the tasks studied in this thesis, by accumulating and leveraging knowledge learned across tasks in a continuous manner. It is envisioned that progressive learning is a step towards a fully general continual learning framework.
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Engineering
Subjects Signal Processing
Keyword(s) Artificial Intelligence
Machine Learning
Deep Learning
Neural Networks
Continual Learning
Machine Perception
Computer Vision
Speech Recognition
Versions
Version Filter Type
Access Statistics: 76 Abstract Views, 105 File Downloads  -  Detailed Statistics
Created: Mon, 04 Mar 2019, 09:42:52 EST by Adam Rivett
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us