KERNEL: A Matlab toolbox for knowledge extraction and refinement by neural learning. In this paper we present KERNEL, a neuro-fuzzy system for the extraction of knowledge directly from data, and a toolbox developed in the Matlab environment for its implementation. The KERNEL system belongs to the novel approach which concerns the use and representation of explicit knowledge within the neurocomputing paradigm: the Knowledge Based Neurocomputing. A specific neural network is designed that reflects in its topology the structure of the fuzzy inference model on which the KERNEL system is based. A well-known system identification benchmark is used as an illustrative example.