We discuss an idea of granular computing regarded as a development environm
ent of neural networks and leading to the emergence of a new class of granu
lar neural networks, Such networks are viewed as new computing architecture
s that are focused on processing information granules rather than being gea
red towards plain numeric processing as usually encountered in most neural
networks, The considered information granules are represented as constructs
that may be formalized in the setting of set theory, fuzzy sets, rough set
s or specified within a probabilistic environment. We discuss several main
approaches to the design of information granules. A number of fundamental i
ssues are tackled including specificity of information granules vis-a-vis l
earning complexity in the neural networks along with their generalization f
eatures. We also provide with a list of architectures of granular neural ne
tworks and elaborate on the associated training (learning) scenarios. (C) 2
001 Elsevier Science B.V. All rights reserved.