Naslov
Selecting neural network architecture for investment profitability predictions

Autori
Kišasondi, Tonimir ; Lovrenčić, Alen

Izvornik
Journal of information and organizational sciences (1846-3312) 30 (2006), 1; 93-103

Vrsta, podvrsta i kategorija rada
Radovi u časopisima, članak, znanstveni

Ključne riječi
neural networks; MLP training algorithms; neural network ADT; neural network graph modelling

Sažetak
In this paper we present a modified neural network architecture and an algorithm that enables neural networks to learn vectors in accordance to user designed sequences or graph structures. This enables us to use the modified network algorithm to identify, generate or complete specified patterns that are learned in the training phase. The algorithm is based on the idea that neural networks in the human neurocortex represent a distributed memory of sequences that are stored in invariant hierarchical form with associative access. The algorithm was tested on our custom built simulator that support usage of our ADT neural network with standard backpropagation and our custom built training algorithms, and it proved to be useful and successful in modelling graphs.

Izvorni jezik
Engleski

Znanstvena područja
Informacijske i komunikacijske znanosti

POVEZANOST RADA