Epistemology and ethics of deep learning intertwined

Blame for various missteps of deep networks is often assigned to either programmers or users. The discovery of adversarial attacks and the complex epistemology behind supervised neural networks leads us to scepticism toward their capacity to work efficiently in domains where robust human categories are central. Will will illustrate how end-to-end optimization that networks are trained to achieve contravene the employment of human categorization when attempting to understand networks‘ inner operations. As ethics is one important domain where various humanly robust categories are central, deep networks might be particularly unsuitable for the replacement of various operations in this domain.