Deep learning: Deep learning has recently been at the forefront of artificial intelligence (AI) research and has truly revolutionized several fields of AI. Data Science Group members are actively involved in developing methods that can help us both better understand the limitations of existing deep learning approaches and also improve various aspects of their performance. They have worked on a wide variety of topics from this area, exploring both fundamental questions and practical applications. This includes, to mention a few of these topics, exploring and proposing novel deep network architectures, developing new ways of effectively transferring knowledge between networks, revisiting weight parameterizations for deep networks in order to improve their generalization capabilities, proposing novel self-supervised and few-shot learning methods, devising and applying learning approaches that advance the state-of-the-art for fundamental problems from the areas of computer vision and image analysis, making use of attention in the context of knowledge distillation, proposing hybrid (scattering-based) convolutional network architectures that allow for better representation learning and more interpretable features, as well as properly adapting deep neural networks such that they can be applied to arbitrary graph-structured data directly and can also handle structured-prediction tasks.
Neural-based Speech Synthesis - Deep Learning in Speech Processing: Deep Neural Networks have taken the engineering community by storm. Data-rich areas such as image processing and speech processing have been transformed during the last years. The Data Science Group combines its expertise on speech processing and applies deep learning techniques to applications such as voice conversion, speech synthesis and speech enhancement.
Generative Adversarial Nets - Deep Generative Learning: Generative models based on Deep Neural Nets have shown unprecedented capabilities in sampling data from complex but unknown distributions. Researchers in the Data Science Group develop novel algorithms for training generative models, focusing on Generative Adversarial Networks (GANs). GANs have been used in data augmentation schemes to generate synthetic data that follow the same distributional characteristics as the original dataset (which may contain sensitive information or limited number of cases). The proposed methodology has been applied to identify dyslexia in children, using measurements from specialized eye trackers.