Analysis of the Rise of Deep Learning

• In Event: Making Objectivity in Data-Centric Knowledge Practices

Abstract

In recent years, there has been a boom in deep learning, a term coined in 2006 for a family of machine learning models known as neural networks. The goal of this paper is to analyze the growing popularity of deep learning from a philosophy of science perspective. Our project is both descriptive and prescriptive. We show how the rise of deep learning can be described using traditional philosophy of science terminology, mainly drawing from Kuhn and Lakatos, but also evaluate the extent to which it was justified. In particular, we identify a pivotal event in the process of the acceptance of deep learning – the 2012 Imagenet challenge. Imagenet is a well-known computer vision competition. In 2012, a deep learning model did better than all the rest in the competition. After that point, deep learning quickly became the predominant framework for learning algorithms as more and more researchers adopted it. We argue that the 2012 Imagenet competition can be conceived as a “crucial experiment” and catalyzed a process reminiscent of a Kuhnian paradigm shift. We discuss what terms such as paradigm, auxiliary hypotheses, and experiment mean in the context of context of machine learning and the extent to which the transition into deep learning was justified.

Authors

• Ravit Dotan, UC Berkeley

• Smitha Milli, UC Berkeley