• artificial intelligence
• continual learning
• deep learning
• machine learning
• online continual learning
• representation learning
• self supervised learning

The objective of Continual Learning is to build Artificial Intelligence agents able to efficiently learn from non-stationary streams of data and to prevent catastrophic forgetting of past data.
This mimics real world scenarios where distributions of future data is not known.
Further, we explore an Online Continual Learning scenario, where the agent observes only a few examples at a time.
Self Supervised Learning is a machine learning paradigm where a model learns meaningful representations from unlabeled data, by creating its own supervisory signals.
Self Supervised Learning is of relevance in this matter, as the presence of data labels is questionable in an uncertain context, such as Online Continual Learning.

The aim of this thesis is to present a framework for Online Continual Self Supervised Learning; we introduce 3 novel Continual Self Supervised Strategies (AEP, ARP, APRE), based on aligning extracted representations with past representations.
We conducted a thorough analysis of experimental results, proving the superiority of the novel strategies in representation quality, in the context of our Online Continual Learning benchmarks.