An Analysis of Current Continual Learning Algorithms in an Image Classification Context

Artificial Intelligence aims to mimic natural intelligent learning by using lifelong machine learning, which allows an AI to train and learn over its lifetime. Various algorithms have been suggested and developed to allow lifelong learning, these algorithms require deeper analysis, to evaluate and highlight performance benefits. In this research, we will study three state-of-the-art algorithms for lifelong learning: Rehearsal, elastic weight consolidation and synaptic intelligence. We do an analysis and evaluation of their performance in a multiple-task experiment, using different amounts of data, and measuring several performance metrics. We found that these algorithms are similar in performance, but some algorithms perform better than others with less data, or show good performance in task one, but not subsequent tasks. These algorithms could be built upon and improved in future research. The evaluation demonstrated in this research is in the image classification context.