Abstract
To learn a supervised classifier, traditional machine learning assumes plenty of labeled data. But what if labels are scarce or difficult to obtain? Scarcity of labeled data is a recurring problem in web-scale learning tasks where labeling billions of data points is tedious, time-consuming and costly. Imagine manually labeling one-by-one a billion images downloaded from the net.

Modern machine learning proposes to circumvent this labeling bottleneck in a variety of ways. Two common approaches, among others, are Active Learning and Transfer Learning. In this talk, I will introduce the notions of active and transfer learning and focus on one sub-problem that aims to merge the two.