For human-like models, train on human-like tasks.

Journal: The Behavioral and brain sciences
Published Date:

Abstract

Bowers et al. express skepticism about deep neural networks (DNNs) as models of human vision due to DNNs' failures to account for results from psychological research. We argue that to fairly assess DNNs, we must first train them on more human-like tasks which we hypothesize will induce more human-like behaviors and representations.

Authors

  • Katherine Hermann
    Google DeepMind, Mountain View, CA, USA hermannk@google.com.
  • Aran Nayebi
    Neurosciences PhD Program, Stanford University, Stanford, CA 94305.
  • Sjoerd van Steenkiste
    IDSIA, SUPSI & USI, Via Cantonale 2C, 6928 Manno, Switzerland. Electronic address: sjoerd@idsia.ch.
  • Matt Jones
    Google Research, Mountain View, CA, USA sjoerdvansteenkiste@gmail.com https://www.sjoerdvansteenkiste.com/.