The CSA approach to knowledge representation in neural networks

摘要

New generation computers are to a large extent dependent on thenprogress of neural network research. The biggest problem of neuralnnetworks is the lack of representational power and incompatibility withnthe conventional AI. We propose to analyze neural networks using thenCalculus of Self-Modifiable Algorithms which is more general than neuralnnetworks. We demonstrate why neural networks can be interpreted as ansubclass of self-modifiable algorithms and how they work asnself-modifiable algorithms. For illustration, basic neural nets modelsnare described in a uniform way using this new approach