Disentangling Neural Architectures and Weights: A Case Study in Supervised Classification