Department of Computer Engineering
CS 590 SEMINAR
Exploiting Unsupervised Data for Zero Shot Learning
Mert Bülent Sarıyıldız
Computer Engineering Department
Deep neural networks have shown remarkable performance in numerous domains, including computer vision, speech recognition, language processing and many more. Most deep learning approaches rely on training over large-scale datasets. However, carefully annotating sufficiently many samples can be prohibitively expensive when the number of object categories is large. In addition, finding training instances for rare classes is a laborious task on its own. In order to reduce the dependency on supervised training datasets, zero-shot learning (ZSL) aims to build models for unseen classes without zero training examples. Arguably, there are two main streams actively being studied in ZSL: (i) learning a better model of semantic relationship between embeddings of classes and domain specific features extracted by a deep neural network, (ii) utilizing unlabeled test samples to reduce the domain-shift problem, which roots from the fact that train and test set can differ significantly in ZSL. Following the latter trend, we propose a novel method that models underlying data manifold of the whole dataset as embedding space of that dataset. We benefit from recent advances in generative models in machine learning, particularly variational auto-encoders. Our proposed solution outperforms current state-of-the-art techniques on several benchmark datasets, namely CUB, SUN and AWA.
DATE: 06 November, 2017, Monday @ 15:40