Publication View

On the Computational Power of Faulty and Asynchronous Neural Networks (2007)

Abstract
This paper deals with finite size recurrent neural networks which consist of general (possibly with cycles) interconnections of evolving processors. Each neuron may assume real activation value. We provide the first rigorous foundations for recurrent networks which are built of unreliable analog devices and present asynchronicity in their updates. The first model considered incorporates unreliable devices (either neurons or the connections between them) which assume fixed error probabilities, independent of the history and the global state of the network. The model corresponds to the random-noise philosophy of Shannon. Another model allows the error probabilities to depend both on the global state and the history. Next, we change the various faulty nets to update in a total asynchronity. We prove all the above models to be computationally equivalent and we express their power. In particular, we see that for some constrained models of networks, the random behavior adds nonunifromity to ...

Publication details
Download http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.55.7158
Source ftp://ftp.cis.ohio-state.edu/pub/neuroprose/siegelmann.prob.ps.Z
Contributors CiteSeerX
Repository CiteSeerX - Scientific Literature Digital Library and Search Engine (United States)
Type text
Language English
Relation 10.1.1.47.8383, 10.1.1.28.2074, 10.1.1.40.7980, 10.1.1.30.1282, 10.1.1.38.920