Traps, pitfalls and misconceptions of machine learning applied to scientific disciplines

Abstract

In the last decade, Machine Learning has experienced a dramatic increase in performance on a wide variety of tasks, including computer vision, speech recognition, text parsing, and language translation, just to name a few. This has corresponded to an understandable hype especially for the remarkable results achieved in some cases. Therefore, practitioners of Scientific Disciplines have become interested in utilizing new Machine Learning techniques, and have sometimes started doing so with mixed success. The purpose of this paper is to describe some of the common Traps, Pitfalls and Misconceptions of Machine Learning as relevant to the Scientific Discipline, and how to avoid them. In fact, Machine Learning and Deep Learning are fast evolving fields, and some of the astonishing results achieved recently sit on small but important details which have become the state of the art. Some of these details are not broadly known by the scientific community. No new scientific result is presented in this paper, which is a survey and a summary of the best of the field, for the benefit of researchers with limited experience. It is not the intention of the authors to provide any criticism to the work of experienced practitioners, particularly not to the ones working on the cutting edge of what is currently possible: in these cases expert researchers may well be doing exactly what we recommend here to avoid, and for a good reason. However we believe that the advice provided here will be useful, and perhaps even a reference, for the newcomers of the field.