François CholletVerified account

In the context of machine learning research, science and engineering are not distinct concepts. You don't do "science" by thinking very hard about platonic ML concepts and then publishing your thoughts. You do science by engineering systems that test small ideas, iteratively.

Also, the general lack of scientific rigor in the field is more often than not coming from people with little engineering experience. If anything, knowing how to deal with system complexity makes your more rigorous (in the sense of the scientific method)

Buggy code is bad science. Poorly tuned benchmarks are bad science. Poorly factored code is bad science (hinders reproducibility, increases chances of a mistake). If your field is all about empirical validation, then your code *is* a large part of your scientific output.

Similarly, deployment of ML systems is not a completely independent branch with a one-way connection to research. Remember that the purpose of machine learning research is to *generate knowledge about how to build ML systems that work in the real world*.

In the same way that the physical reality we can measure is the ultimate referential for physics research, the real-world effectiveness of our algorithms and systems is the ultimate referential for machine learning research.