Current machine learning based AI systems have limited abilities to transfer knowledge among tasks. In contrast, humans are highly adaptive, being able to learn new skills with few examples. We believe that meta learning techniques, i.e., techniques for learning how to learn, are a key ingredient to provide current AI systems with enhanced capabilities to learn new tasks or acquire new abilities. In this context, our current research focuses on 3 main areas:
i) Develop new training schemes that incorporate meta learning strategies,
ii) Incorporate auxiliary tasks as a way to constraint the internal representation of deep learning models,
iii) Create new learning architectures that improve generalization among different tasks.