Deep transfer learning on continual learning

• Main Author: Sousa Leite de Carvalho, Marcus Vinicius
• Other Authors: Zhang Jie
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/171082
• Institution: Nanyang Technological University

Artificial intelligent agents acting in the real world interact with a multitude of data streams. As a result, they must attain, accumulate and record various tasks from non-stationary data distributions. In addition, self-governing computational agents must acquire an understanding of new experiences and transfer knowledge from prior learning over a long period. Continual learning or lifelong learning refers to the capacity to continuously acquire new knowledge, adapt to changing circumstances, and integrate new experiences with previously learned ones over an extended span of time. Catastrophic forgetting is the primary obstacle in computational models that use continual learning. It refers to the neural networks' tendency to disrupt the existing learned knowledge while being trained on new information. This leads to a sudden decline in performance when the new information overwrites the previous knowledge entirely or partially. The learning agents should assimilate the new information seamlessly to enhance their existing knowledge and prevent catastrophic forgetfulness. The model must be capable of preserving most or all of the acquired knowledge, ensuring that the new information does not hinder the previously acquired knowledge. Computational models that engage in continuous learning are often designed to mimic the learning abilities of humans and other mammals. These animals can acquire, distill, and communicate knowledge over long periods of time. They benefit from various neurophysiological processes that facilitate the development of perception and motor skills through experience. Unlike machines, humans and other mammals can effortlessly acquire new skills and transfer knowledge between different domains and tasks. Moreover, the human brain has a remarkable ability to integrate multisensory information, allowing it to respond effectively in situations where there is sensory ambiguity and to draw on knowledge from different domains to achieve a common objective. Consequently, agents that engage in continual learning in the real world need to be able to deal with uncertainty, process a continuous stream of multisensory data, and learn multiple tasks without disrupting their previously acquired knowledge. Achieving these goals has proven to be a persistent challenge for machine learning, neural network research, and the development of general intelligent systems.