Скачать или смотреть Andrés Paez - Idealization and non-factive understanding in machine learning

Proxy or surrogate models are used in explainable artificial intelligence (XAI) to provide some degree of understanding of opaque machine learning systems. In this talk I explore the nature of these models. In particular, I inquire whether they are akin to the idealizations used in science to understand complex phenomena. I argue that although proxy models differ in significant ways from idealizations, they cannot be understood in factive or quasi-factive terms either. The peculiar nature of proxy models, and the epistemic role they play, provide an argument in support of three different but interconnected theses. First, I argue that proxy models show that (i) non-factive understanding is a legitimate and unavoidable form of understanding. This view is closely tied to (ii) the idea that many models are epistemic tools that transcend their representational nature. Proxy models present a vivid example of that pragmatic thesis. Finally, I argue that (iii) proxy models provide objectual understanding of the target system, and that the understanding they provide cannot be reduced to a functional kind of understanding based only on inputs and outputs.