Active learning: theoretical aspects and application to multimedia annotation

Description

Active Learning is a machine learning method that allows an algorithm to interact with the user to ask him/her information about the elements considered as most relevant to discriminate classes in a database. These methods are thus theoretically more efficient than passive learning algorithms, which ask information about randomly chosen elements. Active learning is particularly well suited for sizeable databases, in which information about class labels is difficult to obtain [1]. One application of this work is sound annotation. Active learning allows in this case to decrease the number of individual sounds that the user must listen to, which allows considerable gains in time. Industrial applications of sound classification are numerous, in domains such as identification of sounds in industrial contexts, or classification of sound effects for film production. This project aims at studying the theoretical advantage of various active learning methods [2, 3], and to apply these results to the annotation of sound databases.

[1] T. S. Huang, C. K. Dagli, S. Rajaram, E. Y. Chang, M. I. Mandel, G. E. Poliner, and D. P. W. Ellis. Active learning for interactive multimedia retrieval. Proceedings of the IEEE, 96(4), 2008. .

[2] S. Dasgupta. Two faces of active learning. Theoretical Computer Science, (412):1767– 1781, 2011.

[3] M.-F. Balcan, S. Hanneke, and J. W. Vaughan. The true sample complex- ity of active learning. Machine learning, 80(2-3):111–139, 2010.