Distinct contributions of functional and deep neural network features to representational similarity of scenes in human brain and behavior

  1. Iris IA Groen  Is a corresponding author
  2. Michelle R Greene
  3. Christopher Baldassano
  4. Li Fei-Fei
  5. Diane M Beck
  6. Chris I Baker
  1. National Institutes of Health, United States
  2. New York University, United States
  3. Bates College, United States
  4. Princeton University, United States
  5. Stanford University, United States
  6. University of Illinois, United States
8 figures and 1 additional file

Figures

Models and predicted stimulus dissimilarity.

(A) Stimuli were characterized in three different ways: functions (derived using human-generated action labels), objects (derived using human-generated object labels) and DNN features (derived using …

https://doi.org/10.7554/eLife.32962.002
Behavioral multi-arrangement paradigm and results.

(A) Participants organized the scenes inside a large white circle according to their perceived similarity as determined by their own judgment, without receiving instructions as to what information …

https://doi.org/10.7554/eLife.32962.004
RDMs and model comparisons for fMRI Experiment 1 (n = 20).

(A) RDMs displaying average dissimilarity between categories in multi-voxel patterns in PPA, OPA and MPA, ordered as in Figure 1B (rank-ordered for visualization only). (B) Average (bar) and …

https://doi.org/10.7554/eLife.32962.005
Correlations and variance partitioning of behavioral measurements of scene categorization and similarity of fMRI responses.

(A) Correlations of three measures of behavioral categorization (see Results section for details) with fMRI response patterns in PPA, OPA and MPA. See legend of Figure 2B for explanation of the …

https://doi.org/10.7554/eLife.32962.006
RDMs and model comparisons for Experiment 2 (n = 8, covert naming task).

(A) Average dissimilarity between categories in multi-voxel patterns measured in PPA, OPA and MPA (rank-ordered as in Figure 1B). (B) Correlations between the ROIs in A) and the model RDMs for …

https://doi.org/10.7554/eLife.32962.007
Medial (left) and lateral (right) views of group-level searchlights for (A) the DNN and (B) function model, overlaid on surface reconstructions of both hemispheres of one participant.

Each map was created by submitting the partial correlation maps for each model and hemisphere to one-sample tests against a mean of zero, cluster-corrected for multiple comparisons using …

https://doi.org/10.7554/eLife.32962.008
Multi-arrangement behavior searchlights and post-hoc analysis of functional clusters. 

(A) Searchlight result for behavioral scene categorization. Maps reflect correlation (Pearson’s r) of the group-average behavior in the multi-arrangement task from the participants of Experiment 1. …

https://doi.org/10.7554/eLife.32962.009
Figure 8 with 2 supplements
DNN layer and DNN training comparisons in terms of correlation with fMRI responses in scene-selective cortex.

Panels show convolutional and fully-connected (FC) layer-by-layer RDM correlations between (A) an object-trained (ReferenceNet) and a scene-trained (Places) DNN; (B) both DNNs and the a priori

https://doi.org/10.7554/eLife.32962.010
Figure 8—video 1
Layer-by-layer searchlight results for the object-trained DNN (ReferenceNet).

The first half of the movie shows group-average correlation maps for layer 1–8, cluster-corrected for multiple comparisons using Threshold-Free Cluster Enhancement (thresholded on z = 1.64, …

https://doi.org/10.7554/eLife.32962.011
Figure 8—video 2
Layer-by-layer searchlight results for the scene-trained DNN (Places).

See legend of Figure 8—video 1 for details.

https://doi.org/10.7554/eLife.32962.012

Additional files

Download links