LTTM Laboratory

This month we are going to present our recent paper at CVPR 2021.

In our work, we teach machines to automatically improve the learned knowledge to recognize and segment new classes. Our main intuition is that if features are representative of their class and well-separated, then it’s easier to preserve high accuracy on them whilst accommodating future classes. We achieve this by disentangling the feature space: we constrain features of the same classes to be tightly clustered, to be spaced apart from features of other classes and to be sparse. In this way, we preserve geometrical space, since few channels are active, and expressiveness, since features are divided into well-separated clusters. Then, it is easier to accommodate the latent representations of future classes.

Our technology will be helpful for agents deployed in the real world, which continuously need to adapt their discriminative knowledge.

Paper webpage: https://lttm.dei.unipd.it/paper_data/SDR/