This paper proposes a novel supervised feature selection method named NeuroFS. NeuroFS introduces dynamic neuron evolution in the training process of a sparse neural network to find an informative set of features. By evaluating NeuroFS on real-world benchmark datasets, we demonstrated that it achieves the highest ranking-based score among the considered state-of-the-art supervised feature selection models. However, due to the general lack of knowledge on optimally implementing sparse neural networks during training, NeuroFS does not take full advantage of its theoretical high computational and memory advantages. We let the development of this challenging research direction for future work, hopefully, in a greater joint effort of the community.