• continual learing
• replay
• continualai
• machine learning
• neural network

Nowadays, Artificial Neural Networks (ANNs) are widely adopted to solve complex classification and regression problems, performing even better than humans in numerous applications. However, some biological functions of the human brain are not fully implementable in the artificial one. One of these is the capability to continuously learn and adapt. Continual learning tackles this problem, for which several approaches have been developed in the literature. In our thesis, we focus on the Replay-based approaches, that save some samples in memory and replay them in subsequent tasks.
Several replay-based approaches have been proposed, being Replay the most famous. However, the literature lacks of an extensive comparison of them. In addition, despite the advantages of the Replay strategy with respect to other strategies, few works deepen modification of this strategy.
In our thesis, we propose some recommendations on the memory size value comparing strategies. Moreover, we experiment with some weighing policies, to understand the contribution of each task to the learning process, discovering that low and middle-distance tasks are more important in order to achieve good performances. Moreover, we deepen data augmentation applied to Replay strategy, which allows reaching better performance with lower memory sizes.