Relevant Component Analysis

Daphna Weinshall (The Hebrew University of Jerusalem, Israel)

In the continuum between clustering and classification, between algorithms that use only labeled data and algorithms that use only unlabeled data, lies a neglected niche: learning from unlabeled data (clustering?) by automatically generating labels for classification. We focus on one observation: in a variety of unsupervised data collection processes there exists natural dependency between temporally successive datapoints, which immediately provides equivalence ``labels''. Equivalence constraints are relations between pairs of data points which indicate whether the pair was generated by the same source or not. Thus in this talk I will concentrate on the classification problem of learning from equivalence constraints between pairs of datapoints, and its application to computer vision problems such as image retrieval.