A fusion ARTMAP is an online incremental supervised learning algorithm with multiple input channels. Whenever the ARTMAP receives labeled data, it can learn the data instantly. The fusion ARTMAP, however, is not robust to noise, which means the network predicts the wrong classes from noisy inputs. To solve this problem, we propose a multi-channel classification resonance network (MCRN). MCRN consists of two phases. In the first phase, the network maintains multiple channels without concatenating the inputs. In the second phase, the network identifies the inputs near the decision boundaries and reclassifies them by employing multi-layer perceptron (MLP) networks of which weights are trained by a back-propagation algorithm. A parallel match tracking process in MCRN finds the inputs near the decision boundaries. Two-channel classification simulations are carried out to demonstrate the effectiveness of MCRN for multi-channel cases. The simulation results show that the performance of MCRN is better than that of the fusion ARTMAP for artificial data sets.