This is the project page for the manuscript "Human-like dissociations between confidence and accuracy in convolutional neural networks". Codes for training the models and performing all analyses are available under 'Files'. All codes are written in Python and Matlab. If you have any questions, write to Medha Shekhar at medha@gatech.edu **Below is a brief description of the files** The folder 'codes' contains all the codes for generating imgaes and training and testing the models. The folder contains the following files: 1. Generate_images.ipynb: Generates 10,000 training and 1000 testing images for each of the 4 experiments 2. Train_models.ipynb: Trains 25 instances of each of the 3 model architectures: 4-layer CNN, VGG-19 and ResNet-50 3. Fix_stim_parameters.ipynb: The stimulus is simultaneously manipulated along two features such as contrast and noise. This code searches for the stimulus parameter combinations that will give 70% accuracy across 3 conditions. This file requires the models to have been trained previously. 4. Simulate_network_responses.ipynb: Simulates the network and plots the accuracy, confidence and the internal activations of the network. To run this file, you will require pre-trained models and the saved stimulus parameters. However, you can also generate the results without training the models using pre-generated data from model simulations contained in the folder 'data'. 5. Simulate_models_meta_dprime_analysis.ipynb: Generates model responses for the task paradigm (Figure 5 in the manuscript) where one stimulus class is held constant and the other is varied. You will require pre-trained models to run this code. However, to generate the results without the trained models, you can run the file 'analysis.m' in the folder 'meta-dprime analysis' which already contains the simulations. This file computes d' and meta-d' from the simulated data. 6. There are 5 python scripts in the folder which contain helper functions to generate the stimuli, generate the image sets, compute the networks' activations, confidence and accuracy and plot the results. Due to size restrictions, the pre-trained models could not be uploaded here. However, if you want access to them, please write to me.