Face modeling and biometric anti-spoofing using probability distribution transfer learning
Rodrigues, Ricardo Nagel
MetadataShow full item record
Transfer learning, as applied in machine learning, transfers knowledge between similar learning tasks with the objective of improving performance. It has been successfully applied in several classification, regression and reinforcement learning applications. In this thesis, we investigate how transfer learning can be applied to the task of estimating probability distribution functions (pdfs). Our main objective is to apply transfer learning concepts to improve the estimation of probability distributions when very few training samples are available. We focus mainly in the exponential family of pdfs and show applications in document topic modeling, face modeling and biometric fusion. The first contribution of this thesis is to investigate how transfer learning is being applied in some statistical methods based on hierarchical Bayes modeling, like Latent Dirichlet Allocation (LDA). These methods can be seen as implementing pdf transfer learning by using a common prior distribution over a group of pdfs, but usually no explicit relation to transfer learning concepts is drawn. The other contributions of this thesis are in the development and application of three new methods that explicitly implement transfer learning for estimation of pdfs in the exponential family. The first method transfers information between tasks by constructing a prior distribution for the pdf parameters of a given task based on similar tasks. We apply this method for learning the pdf of different, but related, text document topics. The second method is based on the idea that similar tasks have similar pdf parameters. A Markov Random Field is used to implement this idea, inducing a joint prior distribution over all tasks parameters. We apply this method for learning pdfs of faces with similar poses. The third method reduces the number of parameters that need to be learned in a given pdf, allowing it to be efficiently learned with fewer samples. We use this method to model biometric matching scores for spoofed fingerprints.