Event
Lernen durch Fehlerrückführung
On 9 October 1986 David E. Rumelhart, Geoffrey E. Hinton and Ronald J. Williams published a Nature paper on learning in neural networks through back-propagation. In simplified terms, differences between actual and desired outputs are traced backwards through the network to adjust connection weights. The authors showed that hidden units can thereby learn useful internal representations. Original paper
As the authors noted, the procedure could become stuck in local minima. The publication helped popularize the method; it followed earlier work and does not establish the first invention of back-propagation. Research history
Sources
- Nature-Aufsatz und Forschungsgeschichte: Verfahren, Beispiele, Grenzen und Vorgeschichte.
Offen
Frühere Arbeiten zur automatischen Differentiation und zum Netztraining anhand ihrer Originaltexte gesondert zuordnen. / Attribute earlier work on automatic differentiation and network training separately using the original papers.