Invited Talk at Fondazione Bruno Kessler: Focused Learning Over Time

Matteo Tiezzi was invited to give a talk summarizing the group research activities on Learning over Time at the Fondazione Bruno Kessler (FBK), Trento, Italy. The talk spanned several works revolving around the concept of human-like focus of attention.

Abstract

Outstanding results have been obtained by Machine Learning solutions applied to well-defined Computer Vision tasks, usually leveraging huge collections of data, processed in an offline manner using stochastic updates of the model parameters. The recent debate on the risks and shortcomings of centralization of data collections calls for the need of a different perspective. For instance, smaller effort has been devoted by the research community to the direction of learning from a single, potentially life-long data stream, with limited computational and memory budget. We delve into a novel scenario, where neural models continuously process a stream of frames without the privilege of storing exemplars, that is a very natural setting resembling the flow of information that hits the eyes of each human. Indeed, an important feature of the human visual system that is frequently neglected is the “focus of attention” mechanism, that drives the gaze over different spatial regions of the input stimulus. In this talk, we will explore the advantages brought by injecting such a mechanism in neural networks, both from the representational and computational point of view, in the quest for novel learning mechanisms and architectural solutions to devise lifelong learning agents.

Slides