Current state-of-the-art models for automatic FER are based on very deep neural networks that are difficult to train. This makes it challenging to adapt these models to changing conditions, a requirement from FER models given the subjective nature of affect perception and understanding. In this paper, we address this problem by formalizing the FaceChannel, a light-weight neural network that has much fewer parameters than common deep neural networks. We perform a series of experiments on different benchmark datasets to demonstrate how the FaceChannel achieves a comparable, if not better, performance, as compared to the current state-of-the-art in FER.