Part of Advances in Neural Information Processing Systems 3 (NIPS 1990)
Eduardo Sontag
We consider different types of single-hidden-Iayer feedforward nets: with or without direct input to output connections, and using either thresh(cid:173) old or sigmoidal activation functions. The main results show that direct connections in threshold nets double the recognition but not the interpo(cid:173) lation power, while using sigmoids rather than thresholds allows (at least) doubling both. Various results are also given on VC dimension and other measures of recognition capabilities.