Generative Continual Learning

Hybrid Intelligence (HI) is the combination of human and machine intelligence, expanding human intellect instead of replacing it. HI takes human expertise and intentionality into account when making meaningful decisions and perform appropriate actions, together with ethical, legal and societal values. Our goal is to design Hybrid Intelligent systems, an approach to Artificial Intelligence that puts humans at the centre, changing the course of the ongoing AI revolution.

The Hybrid Intelligence project is a collaboration of top AI researchers from the VU Amsterdam, the University of Amsterdam, the TU Delft, and the Universities of Groningen, Leiden, and Utrecht, in areas such as machine learning, knowledge representation, natural language understanding & generation, information retrieval, multi-agent systems,  psychology, multimodal interaction, social robotics, AI & law and ethics of technology. The HI poroject will create a national and international focus point for research on all aspects of Hybrid Intelligent systems.

Project description

Our project “Continual learning and deep generative modeling for adaptive systems’’ focuses on fundamental research into combining various learning paradigms for building intelligent systems capable of learning in a continuous manner and evaluating uncertainty of the surrounding environment.

Adaptivity is a crucial capability of living organisms. Current machine learning systems are not equipped with tools that allow them to adjust to new situations and understand their surroundings (e.g., observed data). For instance, a robot should be able to adapt to new environment or task and assess whether the observed reality is known (i.e., likely events) or it should contact a human operator due to unusual observations (i.e., high uncertainty). Moreover, we claim that uncertainty assessment is crucial for communicating with human beings and for decision making.

In this project, we aim at designing new models and learning algorithms by combining multiple machine learning methods and developing new ones. In order to quantify uncertainties, we prefer to use deep generative modeling paradigm and frameworks like Variational Autoencoders and flow-based models. However, we believe that standard learning techniques are insufficient to update models and, therefore, continual learning (a.k.a. life-long learning, continuous learning) should be used. Since this is still an open question how continual learning ought to be formulated, we propose to explore different directions that could include, but are not limited to Bayesian nonparametrics and (Bayesian) model distillation. Moreover, a combination of continual learning and deep generative modeling entails new challenges and new research questions.

Partner

Team

Anna Kuzina
PhD Student
 

Anna Kuzina

Jakub M. Tomczak
Assistant Professor
 

Jakub M. Tomczak