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.
— François Chollet (@fchollet) July 15, 2018
My contribution to the “science” vs engineering debate. I think papers should be about explaining reproducible ideas clearly, not about winning datascience competitions. (In the ideal world ;) it should be okay to sacrifice performance for clarity, simplicity and reproducibility. https://t.co/xiBIuspCys
— hardmaru (@hardmaru) July 17, 2018
Not saying SOTA isn’t important, but for example a much simpler algorithm that gets say 85% on CIFAR-100, 60 PTB perplexity, or 900 for CarRacing-v0 is more useful than overly complicated methods that get 89%, 56 or 920, esp for adapting them for other completely different tasks.
— hardmaru (@hardmaru) July 17, 2018
