[aabajian]

This is a fascinating paper. It demonstrates how rooted deep learning is in classical mathematics, and how libraries like TensorFlow/Keras/PyTorch/etc. have both helped and hurt its progress. A software developer can build and train a neural network in under an hour with these libraries, with little knowledge of the underlying math. We take for granted that the implementation is optimized. Imagine if the built-in Python sorted function took O(n^3) instead of O(n log n). It wouldn't take long for someone to point out a better approach. There's a huge need for people who excel at math and can program and have time to do open-source.