Email to connect:


Title: How are animate and inanimate categories differentiated in a pre-trained artificial neural network


  • A central goal of social neuroscience is to identify neural processes that support social interaction. Numerous analytical methods have been used to isolate regions of the brain sensitive to social and nonsocial images. However, confirming the validity of these findings requires studying the effects of lesions on the identified brain regions, a goal not possible in human participants and not well matched to animal behavior. Using recently established similarities between artificial neural networks (ANNs) and human brain representations, we seek to identify groups of neurons central to a social function by studying the effects of specific lesions in an ANN which (1) have a similar activation profile (2) produce a category specific deficit in performance when lesioned. Beginning with MobileNetV2, a multilayer ANN trained to recognize objects, we clustered neurons based on their activation to images of animate and inanimate objects. We then lesion each identified cluster to measure the specificity of the resulting performance deficit. Our results show that inanimate-specific neurons have similar functional activations and when lesioned, produce an inanimate selective deficit (inanimate deficit 79.49% vs. animate deficit 25.64%). We were also able to identify a group of units whose lesion produced an animate selective deficit (animate deficit 71.8% vs. inanimate deficit -5.13%). However, these groups of units produced a dissimilar activation profile. These results provide evidence of inanimate functional selectivity that is not present for animate objects. This approach provides grounds for improved understanding of cognitive processes central to social tasks.
  • ICS program: Triple PhD in Computer Science, Neuroscience and Computer Science
  • Advisor: James H Martin
  • Home degree department: Computer Science
  • Name of research lab: Social Neuroscience and Games Lab