• Continual Learning
• Model Merging

Catastrophic forgetting is a phenomenon that afflicts a model during learning on a stream of sequential tasks. This phenomenon destroys knowledge of already learned tasks when the model encounters new ones. The hypothesis proposed in the thesis is that it is possible to mitigate catastrophic forgetting by integrating past task knowledge into the current network with a model fusion technique, preserving old knowledge while learning new tasks. The methodology we propose performs merging at each epoch of the train and handles the stability-plasticity dilemma by modifying the merging coefficient according to the number of experiences. Moreover, our method can be combined with a buffer replay to increase performance. The experiments included scenarios such as Domain-Incremental and Class-Incremental learning on different datasets. The results demonstrated that our model merging approach outperformed existing state-of-the-art techniques. In conclusion, this research contributes to the Continual Learning field by providing a novel approach that integrates past task knowledge into the current network effectively.