Learning paradigms in dynamic environments

Motivation

Machine learning techniques such as neural networks, Support Vector Machines and Bayesian methods constitute particularly successful tools with numerous applications ranging from industrial tasks up to the investigation and simulation of biological systems. Unlike their biological paragon, however, they are often restricted to very narrow settings such as dedicated classification and regression tasks and simple data structures such as real vectors. In biological networks we find a rich repertoire of complex dynamic behaviour, hierarchical organization, and implicit evolution of structures. The learning scenarios which are usually addressed in technological settings are too restricted to investigate these principles, not only concerning recurrence, but concerning the integration and emergence of structure, cooperation, and autonomous behaviour. Although promising approaches can be found in the literature, we are lacking general learning paradigms for complex dynamical systems which go beyond the standard comparably narrow setting as usually formalized and successfully tackled in classical learning scenarios.

The goal of the seminar is to foster the understanding of complex dynamic learning scenarios and to bridge the gap from powerful biological systems towards learning in technical applications by bringing together researchers in the fields of autonomous learning, structure learning, and biological or psychological investigations of dynamical behaviour, structure formation and structure perception.

A principled difference of machine learning and learning in biological systems concerns the autonomy of the learning system. Although some aspects of autonomy and self-organization have lead to remarkable technical achievements in particular in the areas of robotics and knowledge discovery / data mining, the majority of Machine Learning methods concern narrow supervised settings where explicit teacher signals are available or carefully designed by humans. Two problem areas which focus on the problem complex and can be fruitfully addressed within the limited time span of a seminar are the autonomous development and learning of structures and modularity, and the investigation of implicit biases, priors and problem shaping in natural problems which allow humans to efficiently tackle dynamic problems. Based on these two fundamental questions, the following aspects will be discussed in the seminar:

Investigation of biological and cognitive models:

• How does structure emerge in human perception and how do humans process structures?
• What are natural biases, priors and constraints in biological systems e.g. due to the underlying sensors or effectors?
• How do humans shape complex problems such that sub-goals arise which can be solved more efficiently?

• How can structure be represented and adapted in connectionist systems or graphical models without an explicit teacher?
• How can structures of increasing complexity and modularity be integrated in such systems?
• Can structure learning and organization of structures be accompanied by theoretical models?

• What are natural and technically realizable constraints or priors which help to shape complex and possibly ill-posed problems?
• How can natural constraints posed e.g. by human perception be transferred to technical domains such as robotics?
• How can reasonable priors be included when the language and structure resentation of connectionist systems is changed at a higher level?