Pattern recognition and data compression in neural networks have been investigated and it has been realised that biology related work can be conducted using concept similar to those used for the analysis of artificial neural networks.
In the field of pattern recognition the following topics have been investigated:
design of neural networks for the classification of static patterns;
investigation of learning rules and of the efficiency of retrieval for static and dynamic patterns;
storage and retrieval of structured patterns in neural networks;
investigation of the accuracy required for efficient data storage and retrieval, in the context of the design of neural silicon chips.
Data compression is an engineering problem important in high speed transmission of very large amounts of data, in high definition television or voice transmission over telephone lines, for example. The possibility of using Hopfield networks for data compression for picture transmission has been studied as has voice coding for voice transmission.
Studies on artificial neural networks have given insights into neurobiologically plausible models. An in depth study of the immune system, in neural network terms, aims to predict the global dynamical properties of a model immune system.
The aim of the project is to improve the prospects of applying neural network concepts to specific problems; we shall attempt to capitalize on the huge number of facts which are already known from biology and/or from cognitive science, in order to extract the overall features that are characteristic of the high-level operations of the brain, relevant to the specific problems under investigation; given the current state of solid-state technology, trying to mimic neurons in any detailed way is just unthinkable, anyway. Therefore, the general philosophy of our investigations will be: get from neurosciences some hints on how problems have been solved by nature, then investigate the properties of feasible architectures that might be derived from these considerations. As mentioned above, we shall focus on the applicability of neural networks to real problems in data processing; there are essentially three issues involved: coding and representation of data, architectures and learning, and dynamics of retrieval.
Funding SchemeCSC - Cost-sharing contracts
5600 JM Eindhoven