Document worth reading: “HARK Side of Deep Learning — From Grad Student Descent to Automated Machine Learning”

Recent developments in machine learning evaluation, i.e., deep learning, launched methods that excel normal algorithms as well as to folks in a quantity of superior duties, ranging from detection of objects in photographs and speech recognition to collaborating in troublesome strategic video video games. However, the current methodology of machine learning evaluation and consequently, implementations of the real-world features of such algorithms, seems to have a recurring HARKing (Hypothesizing After the Results are Known) topic. In this work, we elaborate on the algorithmic, monetary and social causes and penalties of this phenomenon. We present examples from current frequent practices of conducting machine learning evaluation (e.g. avoidance of reporting unfavourable outcomes) and failure of generalization potential of the proposed algorithms and datasets in exact real-life utilization. Furthermore, a attainable future trajectory of machine learning evaluation and progress from the angle of accountable, unbiased, ethical and privacy-aware algorithmic selection making is talked about. We would love to emphasize that with this dialogue we neither declare to current an exhaustive argumentation nor blame any specific institution or specific particular person on the raised factors. This is only a dialogue put forth by us, insiders of the machine learning space, reflecting on us. HARK Side of Deep Learning — From Grad Student Descent to Automated Machine Learning