Continual learning is the task of teaching machines on streamed data. In this context neural networks have exhibited the unsolicited capability of strongly forgetting past knowledge focusing on more recent targets. This is mainly due to the gradient descent training techniques and this phenomenon is exacerbated by the absence of a proper and ad-hoc regularization for streamed incremental training.
In this talk we will introduce the problem of continual learning and show how a simple yet competitive baseline for the problem can be set up by mixing replaying techniques with regularization and knowledge transfer.
The seminar purpose is to set up a starting point for CL practitioners introducing the problems and simple tools to start research in the field.
About the Speaker
Simone Calderara is Associate Professor of Machine and Deep Learning at the Engineering Department Enzo Ferrari of University of
Modena and Reggio Emilia (UNIMORE). He is representative for APRE-EU Cluster Digital, Industry and Space, UNIMORE representative in the CINI National Artificial Intelligent and Intelligent System AIIS laboratory, senior member of the Artificial Intelligence Research and Innovation center, senior member of AI Academy of UNIMORE, senior member of the AImageLab laboratory (aimagelab.unimore.it). He is also an ELLIS member and core member of the UNIMORE ELLIS Unit. His main research interests are deep learning techniques for online, incremental and continual representation learning with application to human behavior understanding in industrial and urban scenarios. In this context he has participated in numerous national and international projects and co-authored more than 90 publications with an h-index of 35 (Google Scholar).
You are welcome to participate both in presence and online on Zoom.
If you are going to participate on Zoom, please send an e-mail to andrea.passerini [at] unitn.it in order to get the link.