Speaker: Jedrzej Kozerawski

Title: SVM Transfer Learning for Object Recognition

 

Learning object representation given many positive samples for training is nowadays a fairly known task. Achieving the same goal given only few examples is a much harder problem, but we know that human brain tends to adapt very well information from one domain to another and transfer the knowledge it already possesses to learn more efficiently in the future. 

Same principle is a motor for current “Learning to learn” methods for transferring knowledge from already known object categories to facilitate learning the representation of completely novel and previously unseen objects.

I will describe methods involving pure SVM approaches and deep learning model transfer networks where prior knowledge can be used to boost recognition of novel object instances.