Optical communication networks use electrical constellation converters requiring optical-electrical-optical conversions and expensive symbol-rate limiting electronics. In this paper, a generic method for all-optical silicon-photonic conversion of amplitudephase modulation formats is proposed. The method is based on the implementation of single-layer radial basis function neural networks. A mathematical model is developed, and the parameters influencing the performance of the method are analyzed. Full optical simulation of a four-symbol constellation conversion was performed, resulting in errorfree converted constellation that has an error vector magnitude lower than 2.5%.