In this study, we investigate a neuro-control scheme proposed in the literature, which uses techniques from variable structure systems (VSS) theory in order to robustify learning dynamics' for control of nonlinear systems. Gaussian Radial Basis Function Neural Network (GRBFNN) is chosen as the neural net-work architecture because of its strong adaptation capabilities. By means of an instability analysis, it is shown that this scheme leads to unbounded evolution of the controller parameters in steady state due to presence of noise and uncertainties. A modification on the original adaptation algorithm is proposed in order to alleviate this problem. The simulation studies on a nonlinear cement mill circuit model show that the modified update rule stabilizes the learning dynamics and closed loop system becomes insensitive to parametric changes.