“Misconfigured neural networks often continue to train without throwing errors, but will perform worse than correctly configured models.”