Learning matrix
The learning matrix is a special type of artificial neural network (ANN) , which was invented by computer science and ANN pioneer Karl Steinbuch around 1960 .
application
With this technical model for adaptive systems, similar to the conditional Pavlovian reflexes in living beings, complex links between certain sets of properties (e.g. letters of an alphabet) and associated meanings can be established. With this fundamental invention, technical character recognition, shape perception, automatic language translation, the decoding of disturbed messages in radio and data transmission, and also today's copier and fax technology were made possible.
function
The learning matrix generally consists of n “lines of properties” and m “lines of meaning”, with each line of properties being linked to each line of meaning, similar to how the neurons in the brain are linked by synapses . (This can be realized in different ways - according to Steinbuch, this can be done as a pure hardware solution instead of as a computer program). By suitably interconnecting several learning matrices, a switching system can be set up which, after certain training phases have been completed, is able to automatically determine the most likely associated meaning for an entered sequence of features. Conditional complex links between certain sets of properties (e.g. letters of an alphabet, or points in certain colors) and associated meanings (words formed from those letters or figures from those points) can be established.
A learning matrix must always be “trained”; for this purpose, values are given on the corresponding property and meaning lines (binary or real); then the connections between all pairs of traits and lines of meaning are strengthened. (The Hebb rule is also used to calculate this gain .) When the learning phase is completed, the “can phase” begins: When a certain input is made on the property lines, the learning matrix activates the corresponding meaning lines.
By suitably interconnecting several learning matrices, a switching system can be set up which, after certain training phases have been completed, is finally able to automatically determine the most likely associated meaning for an entered sequence of features.
literature
- Karl Steinbuch : Automat and Human. 1st edition Springer 1961.
Web links
- A new theoretical framework for the Steinbuch's learning matrix, bibcode : 2005SPIE.5916..233S
- Pattern recognition and classification using weightless neural networks (WNN) and Steinbuch Lernmatrix, bibcode : 2005SPIE.5916..247A
- DARPA project will study neural network processes
- Discussion in the newsgroup de.sci.informatik.ki
- Wolfgang Hilberg: "Karl Steinbuch, an unjustifiably forgotten pioneer of artificial neural systems", communication from the Institute for Data Technology at the Technical University of Darmstadt, 1995, PDF file (size approx. 4 MB), accessed on June 17, 2017
Individual evidence
- ^ Karl Steinbuch - computer scientist from the very beginning at edoc.hu-berlin.de, accessed on March 1, 2015.