부산대학교 도서관

Representation learning involves learning models that process images in a way that makes them easier to use in down stream tasks, such as image classification, retrieval or robotics. Generative models learn representations by synthesising (consistent) novel image samples. I propose four models for learning representations both from unlabelled and partially labelled data using generative models. I being by proposing Sketch-GAN (Generative Adversarial Network, GAN), a model that learns representations for a small dataset of hand drawn marks (similar to sketches). The representations are used for retrieving marks that are visually similar to a query mark. Qualitative retrieval results demonstrate that the learned representation is suitable for retrieval. Following on from this, I propose a denoising adversarial autoencoder, DAAE and an integrating DAAE, iDAAE. These models incorporate two state-of-the-art regularisation techniques, that had previously only been investigated in isolation. I demonstrate that these models learn representations that encode attributes of the data with a near-linear structure. This allows us to use linear models on top of representations to perform down stream tasks, for example classification. The final model that I propose, unlike the previous three, utilises small amounts of labelled data and learns to separately encode the identity and attribute of an object, in this case a face. This model (a) achieves classification scores that are competitive with a state-of-the-art classification model and (b) may be used for facial attribute editing. The ability to edit attributes is evidence for the identity and attributes being encoded independently. Generative models may be evaluated on a specific task, for example attribute editing, retrieval or classification. I am also interested in evaluating generative models in a task agnostic way. Towards this, I propose an evaluation method that is task agnostic.