This study reveals that functional neural networks (FNNs) suffer from a “curse of smoothness,” where the smoothness of hidden state functions causes gradient vanishing and hampers optimization in deep architectures, highlighting the limitations of directly applying standard deep learning methods to functional data.