A key design principle I follow in libraries (e.g. Keras) is "progressive disclosure of complexity". Make it easy to get started, yet make it possible to handle arbitrarily flexible use cases, only requiring incremental learning at each step.
Like zooming in a complex landscape.
Like zooming in a complex landscape.
When you zoom in, the incremental complexity is still part of *the same picture*. Everything you're learning in the beginning will still be relevant once you've become an expert. You will be gradually diving into workflows where you're writing more and more logic from scratch.
So with something like Keras, you won't have to switch to an entire different framework as you go from student to researcher, or from data scientist to deep learning engineer.
I think thoughtful design can dramatically broaden the impact and reach of software. There are no complicated ideas in deep learning, only bad interfaces.
