Considering the «data classification» problem it is known that efficient classifiers only consider a few (pertinent) parameters. This seems in contradiction with usual biological plausible models, based on neuronal networks, which intrinsically have a lot of parameters. Here, we propose to solve this apparent contradiction, building a link between biological plausible models and classifiers with low Vapnik-Chernovenkis dimension. The -somehow very simple- key idea is to consider piece-wise linear classifiers of minimal dimension, as a generalization of support-vector machine. This allows to solve the previous dilemma at both a theoretical and computational levels, including some elements of biological plausibility. Experimentation of a small interactive toy demonstration to analyze the performances of these mechanisms is reported, while the methodology is validated on a real experimental problems.