Fast and robust visual object recognition in young children

Fast and robust visual object recognition in young children Data for "Fast and robust visual object recognition in young children" Ayzenberg, V., Sener, S.B., Novick, K., & Lourenco, S. F. (2025, accepted). Fast and robust visual object recognition in young children. Science Advances. The data, code, stimuli, and figures can also be found at: https://github.com/vayzenb/kornet The kid_object_recognition-data_stimuli_code.zip contains all the files necessary for replicating the results in the paper The kornet_train_image.zip file contains all the images that were used for training model classifiers Naming Conventions Slightly different naming conventions were used as shorthand for the tasks stimuli and analyses. These do not always line up with the labels used in the manuscript. The repo name KorNet stands for "Kid Object Recognition Networks" -- a play on the CORnet family of models Tasks Complete contour conditions is sometimes referred to as 'Outline' Perturbed contour condition is sometimes referred to as 'Pert' or 'ripple' Deleted contour condition is sometimes referred to as 'IC' for illusory contours Model Names Models were often renamed for the main-text to more precisly specify their architecture and training or for readability. See the model_comparison.ipynb notebook and the modelling/model_loader.py for all model info Folder Structure analyses Contains all files for conducting all high-level statistical analyses presented in the main-text Human only data figures can be recreated using the sub_analysis.ipynb notebook Human model comparison figures can be recreated from the model_comparison.ipynb notebook data Contains all individual participant data files for both children and adults. The sub_info.csv file contains subject demographic information for children figures Contains all figures presented in the main-text and supplemental materials modelling Contains scripts necessary for extracting model activations and conducting classifcation (i.e., decoding). Below I highlight the files most relevant to the main-text model_loader.py: A general script for loading model architectures, their transforms, and weights train.py: The training script used for training VoneNet models extract_acts.py/extract_acts_layerwise.py: Scripts for extracting feature activations for all training and test stimuli, either from the top layer or from pre-specified layers defined in "all_model_layers.csv" decode_image.py/decode_image_layerwise.py: Scripts for predicting the category of a test image after training on naturalistic images using activations from the extract_acts script train_sizes.xlsx: An excel sheet summarizing the dataset size and total experience of each model as measured by data_set size x eopochs. Information is drawn from papers or model cards describing each model results Summary files that are used in subsequent analyses. group_data folder: this contains human only summaries seperated by condition, and summaries containing model performance model folder: contains performance of each model on the test stimuli using each classifier, seperated by # of training images used natural_image_decoding: performance of each model on the training images The top-level results folder also has various summary files for generating figures. stim Contains all the test stimuli used in the current study (see naming conventions). Original stimuli was numerically labelled. See kornet_classes for the corresponding object name for each number The completed_silhouette folder contains a silhouette versions of the complete contour condition stimuli. These were used to compute curvature and shape envelope statistics