Title :
Univalent Neural Nets and Shape Detection
Author_Institution :
Inst. of Comput. Sci., Czech Acad. of Sci., Prague
Abstract :
We introduce a notion of so called univalent neural networks realizing injective mapping and sharing the input and output space. First, we postulate necessary and sufficient conditions of univalence and derive several models of univalent nets. Then explore learning algorithms that could be used for the defined network class - special variants of backpropagation learning.
Keywords :
"Artificial neural networks","Training","Classification algorithms","Matrix decomposition","Sufficient conditions","Linear matrix inequalities","Approximation algorithms"
Conference_Titel :
Signal-Image Technologies and Internet-Based System, 2007. SITIS ´07. Third International IEEE Conference on
Print_ISBN :
978-0-7695-3122-9
DOI :
10.1109/SITIS.2007.126
