September 2019
Volume 19, Issue 10
Open Access
Vision Sciences Society Annual Meeting Abstract  |   September 2019
Relating category-selective regions in biological and artificial neural networks
Author Affiliations & Notes
  • Jacob S Prince
    Department of Psychology, Harvard University
  • Talia Konkle
    Department of Psychology, Harvard University
Journal of Vision September 2019, Vol.19, 60d. doi:
  • Views
  • Share
  • Tools
    • Alerts
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Jacob S Prince, Talia Konkle; Relating category-selective regions in biological and artificial neural networks. Journal of Vision 2019;19(10):60d.

      Download citation file:

      © ARVO (1962-2015); The Authors (2016-present)

  • Supplements

In the occipitotemporal cortex of the human brain, a few focal regions respond relatively selectively to some categories—e.g. faces and houses. However, recent empirical work has shown that the representational geometries of these regions can be quite strongly correlated with each other and with occipitotemporal cortex outside of these regions (Cohen et al., 2017). Here, we leverage deep convolutional neural networks to provide some insight into this unexpected empirical result, defining and probing deepnet category-selective “regions”. Data acquisition occurred by extracting artificial neuron activations from a pretrained convolutional neural network (Alexnet) in response to the image set used in Cohen et al. Within each layer, non-overlapping subsets of neurons responding preferentially to faces and places were defined using a selectivity contrast (faces>all other categories; houses>all other categories). Deepnet face- and place-selective regions defined in the last convolutional layer (after pooling) had moderate-to-highly correlated geometries, mirroring the human brain data. For example, face- and place-selective regions of 100 neurons had representational geometries that correlated with human fusiform face area (FFA, r=0.66) and parahippo-campal place area (PPA, r=0.81), and with each other (deepnet-FFA to deep-net-PPA r=0.56). Also like the human brain data, these deep net regions were correlated with the geometry of the “non-selective” layer neurons: r=0.61 and r=0.54). Even though these deepnet neurons did not learn features specifically to do place or face recognition, face and place “regions” were present with correlated geometries similar to human FFA and PPA. Given that the features of deepnet neurons are optimized to discriminate all categories from all other categories, our results are consistent with the broad possibility that occipitotemporal cortex as a whole participates as one massive discriminative feature bank, where face and place regions reflect a topographic mapping of this feature space onto the cortex.

Acknowledgement: Harvard Brain Science Initiative Collaborative Seed Grant 

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.