This paper shows that neural networks which use continuous activation
functions have VC dimension at least as large as the square of the num
ber of weights w. This results settles a long-standing open question,
namely whether the well-known O(w log w) bound, known for hard-thresho
ld nets, also held for more general sigmoidal nets. Implications for t
he number of samples needed for valid generalization are discussed. (C
) 1997 Academic Press.