Deep Learning

Our scientists research and develop novel methodologies, models, and algorithms to enhance the capabilities of deep learning systems and expand their application domains.

Through rigorous analysis and experimentation we strive to unravel the underlying principles governing the behavior and performance of deep neural networks. This includes investigating the effects of network depth, width, and connectivity patterns on learning dynamics, generalization, and expressiveness. These endeavors aim to provide deeper insights into the fundamental properties of deep learning systems and lay the groundwork for further advancements.

Another significant focus is the development of advanced techniques for training deep neural networks. This involves exploring innovative optimization algorithms, regularization methods, and initialization schemes that facilitate faster convergence, mitigate overfitting, and enhance the robustness and generalization capabilities of deep learning models.

We investigate novel architectures and mechanisms to improve the interpretability, explainability, and reasoning abilities of deep learning systems. The objective is to develop models that not only achieve exceptional performance but also provide insights and justifications for their decisions, enabling users to understand and trust the outputs of these models.

Our research delves into unsupervised and reinforcement learning, seeking innovative methods to leverage unlabelled data and reinforcement signals to drive learning and decision-making processes. This entails developing unsupervised learning algorithms that can autonomously discover meaningful representations and structures in data, as well as reinforcement learning techniques that enable intelligent agents to learn from interaction with the environment and optimize complex tasks.