Index
Technical Facilities
Neural Networks
Since about 1990 artificial Neural Networks have been used at the
Institute for four purposes:
- to study their mathematics and behavior
- to apply them in recognition tasks especially to recognize speech
patterns such as phonemes, words, and prosodic characteristics
- to investigate the usability of recurrent NN to spot keywords in
continuous speech
- to use them in psycholinguistically motivated modeling
A number of different standard Neural Networks have been investigated:
- spreading activation networks of different sorts
- simple feed-forward networks trained with the back-propagation
algorithm for static pattern classification
- recurrent neural networks trained with an extended back-propagation
algorithm and home-made fast optimization modules for dynamic pattern
classification
- the standard Kohonen algorithm for static self-organization tasks
Further some new algorithms and architectures were developed at the MPI
- dynamic extensions to the Kohonen algorithm for the
self-organization of dynamic signals
- a special architecture which exists of chains of specially connected
neurons which are capable of storing sequential order and, in so far
as they are able word patterns
- a kind of thermodynamic neural network where weight vectors randomly
shift and under certain conditions form stable clusters to simulate
developmental processes
|