I.3.1. Transfer learning and meta learning

IDUB Programme Objective

I. Increasing the impact of the scientific activity of the University on the development of world science

Name of POB/activity group

III. The Challenge of Petabytes

Description

Objectives

Creating a problem-oriented research team in the following areas of inquiry:

Knowledge transfer in the area of image processing – the models developed so far, learned on specific sets of training images, do not allow for precise estimations of prediction uncertainty in a different distribution of test data. Our goal is to propose a better mechanism for estimating such uncertainty, and to build a model that would allow objects to be recognised in a universal way, as this is a key issue in the development of “perception” in robotics.
Knowledge transfer in the field of robotics, with an emphasis on the manipulation and control of drones – Industrial robots are able to make very fast, precise, and repetitive movements. The main challenge of modern robotics is to develop a methodology that allows robots to perform tasks in increasingly diverse environments. Examples of such tasks include controlling a drone in a new space or lifting previously-unencountered objects from a previously-unencountered container.
Meta-learning in the area of learning with reinforcement – Methods such as Deep Q Learning (DQN) or the Asynchronous Advantage Actor Critic (A3C) enable the learning of agent behaviour strategies in individual environments, such as simulators or computer games. Our goal is to build tools that allow for the transfer of strategies among different environments, while allowing for the option of short “training” using meta-learning techniques.

Justification

In recent years, we have witnessed many breakthroughs in the field of artificial intelligence, made possible by the use of deep neural networks. Using this knowledge, we want to focus on one of the most challenging aspects of artificial intelligence, which is taking the knowledge and experience gained by a trained model and transferring it to different areas of application.

Tasks and projects within the tasks

A new robotics laboratory will be created in order to carry out research on teaching robots. The laboratory will be managed by a European Research Council (ERC) grant winner – Dr hab. Marek Cygan.

In addition to carrying out local research, the Laboratory will offer special packages (through competitions) to enable external collaborators to access its infrastructure (robots, servers), finance research trips, and purchase a license for the necessary software.

The objectives of Measure I.3.1 will be achieved by:

recruiting staff, by means of open competitions, for scientific and technical positions (programmers);
establishing and developing international cooperation with leading research centres in the field (a programme of bilateral visitation);
financing participation by group members in conferences, seminars and thematic workshops;
the purchase of essential research equipment, such as manipulators and computing servers.

Coordinating unit

Faculty of Mathematics, Informatics and Mechanics

Entities involved in implementation

Faculty of Mathematics, Informatics and Mechanics