Grant: Discovery Grants Program - Individual

Supervised machine learning techniques are built by giving training data, and describing a desired concept, so that the computer can produce a model that replicates such concept. The development of powerful supervised machine learning techniques often requires large amounts of labelled multimodal data (e.g., text, images, videos, etc.), collected for instance from different IoT devices, which is crucial for learning high-quality models. This is especially true in the case of deep learning approaches. However, compiling large-scale annotated datasets is a tedious and extremely costly operation which may cause serious privacy threats. Hence, in many cases, practitioners are faced with the absence of training data. In addition, although this form of supervised learning has provided good results, there are classes of concepts (e.g., activities which may differ across cultures, gender, individuals and life trajectory) that cannot be learned efficiently with such rigid approach. Moreover, the majority of existing approaches are not explainable and do not consider the human component while building ("teaching") the model. The goal of the proposed research program is to develop a unified framework and a prototype of a system for the modeling and analysis of unlabeled data. The system outputs should be explainable and adaptative to different contexts. This will be done via the development of explainable unsupervised learning approaches and unsupervised model adaptation techniques. The system should also allow to open a communication channel between the developed models and the user (i.e. domain expert for the application) which shall lead to unification between the human intelligence and the machine intelligence. This will be performed via the development of interactive learning techniques that should not assume that the user has prior knowledge of machine learning. The approaches and techniques that will be developped will be validated via a challenging application that concerns human activities recognition using different data modalities in different contexts/scenarios. The proposed research will broaden the scope of applicability of machine learning by putting it into synergy with human expertise and expectations, which will engender trust on the developed models.