We know from software engineering that striving to make our software simpler can make it better. It makes it easier to reason about debugging, it increases architectural agility, it makes it function better in the real world.
Machine learning software is no different, simpler is often better. It also increases interpretability, it lets your system operate at larger scale, and gives you additional ways to incorporate humans in the loop.
What are some practical steps you can take to make your machine learning code simpler? And what trade-offs are you making by making your system simpler?