Deep neural networks predict category typicality ratings for images

Abstract

The latest generation of neural networks has made major performance\nadvances in object categorization from raw images.\nIn particular, deep convolutional neural networks currently\noutperform alternative approaches on standard benchmarks by\nwide margins and achieve human-like accuracy on some tasks.\nThese engineering successes present an opportunity to explore\nlong-standing questions about the nature of human concepts\nby putting psychological theories to test at an unprecedented\nscale. This paper evaluates deep convolutional networks\ntrained for classification on their ability to predict category\ntypicality – a variable of paramount importance in the\npsychology of concepts – from the raw pixels of naturalistic\nimages of objects. We find that these models have substantial\npredictive power, unlike simpler features computed from the\nsame massive dataset, showing how typicality might emerge\nas a byproduct of a complex model trained to maximize classification\nperformance

Ähnliche Arbeiten