This is the repository for the paper: “Visual competition: similarities and differences between deep neural networks and humans”. It contains the following:
-
‘human experiment’ folder: codes and stimuli used in the human behavioral experiment
a. Mixed images can be found in ‘img’ folder.
b. For a ‘dummy’ version of the experiment open index.html in any browser.
-
‘results’ folder: csv output files of the human’s answers
a. ‘personal’ – subjects’ details (condition, gender, etc.)
b. ‘subject_answers’ – raw subjects’ text answers
c. ‘results_cond_#’ – encoded subjects’ answers (0=none, 1=chose first,2=chose second,3=chose both).
-
‘analysis’ folder: Example codes for data generation and analysis. These codes demonstrate the major analysis we performed (pretrained models were downloaded from: https://github.com/BVLC/caffe/wiki/Model-Zoo, ImageNet dataset was downloaded from: http://image-net.org/download-images).
a. ‘morph_img_pair.m’- code for mixing two images in different methods.
b. ‘extractNetworkChoice.m’ – used after running the images in the network. This code loads the network’s output and define the network choice (see text). This script uses three utils functions: ‘LoadNetResults.m’, ‘GetIndicesAccordingToNet.m’, ‘GetIndicesAccordingToExp.m’
c. ‘CalcImageParameters.m’ – Calculates the image parameters used in the paper (see Table 1).
d. ‘RunGLMOnImageFeatures.m’ – Trains a generalized linear model to predict networks’/humans’ choices using different features combinations.
For further details please contact: liron.gruber@weizmann.ac.il