We present a semi-supervised learning algorithm robust to label errors in training data. Our method employs the mode filter used for smoothing noisy images. We extend it from images to functions on graphs for regression of classification functions on an undirected graph. Our contribution in this paper lies in the introduction of nonlinearity in the regression in contrast to linear interpolation used in previous graph-based semi-supervised learning algorithms. Error-correcting effect of mode filters is demonstrated and the classification rates of the present learning method is evaluated with experiments for the UCI benchmark datasets contaminated with label errors.