• dataset condensation
• machine learning
• online continual learning

Continual learning is a Machine Learning paradigm that studies the problem of learning from a potentially infinite sequence of tasks. The main challenge of Continual Learning is to avoid the catastrophic forgetting phenomena, namely the problem of forgetting the previously acquired knowledge while learning new tasks. We developed a replay-based continual Learning strategy based on the use of Knowledge Condensation techniques to mitigate the catastrophic forgetting problem. These techniques are used to save and compress, in an external memory, the knowledge extracted from the examples provided by the input data stream. The developed strategy is applied to Online Continual Learning, a scenario in which, unlike Continual Learning, the input data stream has no task boundaries, thus it is considered more challenging. The proposed approach is compared with state of the art Online Continual Learning strategies.