Core Concepts
ANNs can learn human and non-human concepts but may not represent them in individual units.
Abstract
The article explores the narrative that ANNs learn human concepts and store them in individual units.
It delves into three key assumptions: ANNs work well, learn human concepts, and represent them in units.
Evidence is presented for the first assumption, mixed evidence for the second, and questionable evidence for the third.
Various techniques like activation maximization and network dissection are discussed.
The importance of selectivity in ANNs and the implications for performance are highlighted.
The article concludes with a call for skepticism and the need for further research to validate the narrative.
Stats
ANNs are solid predictors in static data settings.
ANNs may learn both human and non-human concepts.
ANNs do not necessarily represent learned concepts in individual units.
Quotes
"One’s skepticism should be proportional to the feeling of intuitiveness." - Leavitt and Morcos