Learning Simple Relations in High Dimensional Signal Spaces

---

In systems of high complexity it is often necessary to employ simulation. It is, however, usually hard to simulate certain subparts of a system, in particular when they exhibit a complex dynamic

The long term objective of this project is to broaden the application area of learning systems in industry. The main difficulty in achieving efficient learning systems suitable for a wider range of industrial problems lies in the required level of adaptability.

Recent developement in the field of neural computation provides consistent evidence that incorporation of knowledge gained in more mature fields of research can significantly benefit understanding of learning processes. Much of the knowledge developed within the areas of information theory , signal theory , control theory and computer science is in fact at the core of learning and the `large scale' strategy in the project is to integrate pertinent theory and principles from these areas.

The search for methods that will allow a sufficient level of adaptivity for learning systems will be based on two main principles:

1.  Simple local models

2.  Adaptive model distribution

Along with these principles the following three basic guide lines will be of general importance:

• Reinforcement principle - This principle allows for a most general form of learning requiring only a measure of success or failure.

• Neural design - implying dynamic self organization, `fuzzy' signal representations and robustness against malfunction of individual units.

• Modularity - implying that a complete system consists of a number of separate subsystems (`neuron clusters'). This requires that the subsystems have the ability to individually adapt to different sub tasks and that they cooperate to produce the actions of the system.

The project is described in greater detail in the documentation below.

PhD theses

• Learning Multidimensional Signal Processing
• Reinforcement Learning and Distributed Local Model Synthesis

Conference papers

• Automated Generation of Representations in Vision
• Learning Visual Operators from Examples: A New Paradigm in Image Processing
• Learning Multidimensional Signal Processing

The scientific manager for this project is ass. prof. Hans Knutsson.
Technical managers for the project are Dr. Magnus Borga and Dr. Tomas Landelius.

The project is sponsored by TFR (Swedish Research Council for Engineering Sciences).