Learning never stops: real-time machine learning in robotics
Fabio Ramos (University of Sydney)
NICTA SML SEMINARDATE: 2013-06-13
TIME: 11:15:00 - 12:15:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.
ABSTRACT:
The real-time and continuous learning requirements in long-term autonomy pose a number of algorithmic challenges for machine learning. How to keep the computational cost bounded over extensive periods of time? How to adapt the models to new environments and unseen data? How to use the interaction of the robot with the environment to self supervise as much as possible? In this talk, I will present two methods for unsupervised and self supervised learning. The first is based on extensions of affinity propagation; a message passing approach to clustering that can determine the number of clusters, can be adapted to online settings, and is scalable to large datasets. I will show applications to obstacle learning, object clustering and multi-sensor segmentation. The second method is a nonparametric classifier to segment dynamic objects in videos from a moving camera. The algorithm is totally unsupervised and improves its performance with more data, providing a library of segmented dynamic objects (such as cars, people, bikes).
BIO:
I received the BSc and the MSc degrees in Mechatronics Engineering at University of Sao Paulo, Brazil, in 2001 and 2003 respectively, and the PhD degree at the Australian Centre for Field Robotics, University of Sydney, Australia, in 2008. My Yoda Master was Hugh Durrant-Whyte.
In January, 2011, I commenced as a Senior Lecturer in machine learning and robotics at the School of Information Technologies, University of Sydney, where I lead the Learning and Reasoning Group. Before, I was an Australian Research Council (ARC) research fellow at the Australian Centre for Field Robotics. My research interests include Bayesian statistics, stochastic processes for spatial modelling, probabilistic networks, and multi-sensor perception. Over the last ten years I have applied these techniques to mining and exploration, environment monitoring, agriculture and healthcare.
