Download attachment GitHub repository
One of the main issues concerning machine learning systems nowadays is the inability to incrementally learn features from the external world without incurring unwanted effects or the inability to preserve previous knowledge.
In our work, we started by reproducing the experiments of iCaRL to have a better understanding of the incremental learning setting and of the problems of current implementations. We tried to solve some of the issues by proposing our variations of the iCaRL architecture based on deep generative approaches (GANs). In parallel, a cluster-based approach is tested for selecting class exemplars to keep in memory.
The performed experiments showed that the proposed methods are able to improve the performances on 100 classes, received in batches of 10, with respect to the original iCaRL approach.
This project was carried in 2020 as part of the Machine Learning and Deep Learning course, for the master’s degree in Data Science and Engineering at Politecnico di Torino.
Authors:
- Ivan D’Onofrio
- Gabriele Degola
- Luca Dibattista
Supervisors:
- prof. Barbara Caputo
- ing. Fabio Cermelli