DonYin/orthographic-dnn-priming

orthographic DNN priming dataset

replication data for the paper:

Yin, D. and Davelaar, E.J. (2023). convolutional neural networks trained to identify words provide a good account of visual form priming effects. Computational Brain & Behavior. doi:10.1007/s42113-023-00172-7

dataset description

this dataset contains pre-rendered word stimulus images used to evaluate how well visual DNN models (CNNs and ViTs) predict human orthographic priming patterns from the form priming project (adelman et al., 2014).

each image is a 224x224 black-background PNG with white text, rendered in arial at size 22, centred.

example stimuli for the target word "design"

ID (identity)	TL12 (transposed 1-2)	DL-1M (deleted middle)	SN-M (substituted middle)	RF (reversed full)	ALD-ARB (all different)
DESIGN	EDSIGN	DSIGN	DESIHN	NGISE	CBHAUX

what's included

prime_data/           <- 11,760 prime stimulus images (420 targets x 28 conditions)
  {target_word}/
    {condition}.png
metadata/
  2014-prime-types.txt   <- 28 prime condition labels
  2014-targets.txt       <- 420 target words
  2014-prime-data.json   <- prime string for each target x condition
  normalization-stats.json <- channel-wise mean/std for the training set

prime conditions (28 types from the form priming project)

code	description
ID	identity (e.g., prime and target are both "design")
TL12	transposed letters positions 1-2
TL-I	transposed letters internal
TL56	transposed letters positions 5-6
NATL2	non-adjacent transposition (2 letters)
NATL3	non-adjacent transposition (3 letters)
DL-1M	deleted letter (1, middle)
DL-1F	deleted letter (1, final)
DL-2M	deleted letters (2, middle)
T-All	all letters transposed
TH	transposed halves
SUB3	subset of 3 letters
RH	reversed halves
IH	interleaved halves
RF	reversed full
SN-I	single substitution (initial)
SN-M	single substitution (middle)
SN-F	single substitution (final)
N1R	neighbours at distance 1 (random)
DSN-M	double substitution (middle)
IL-1M	inserted letter (1, middle)
IL-2M	inserted letters (2, middle)
EL	extra letter
IL-1I	inserted letter (1, initial)
IL-1F	inserted letter (1, final)
IL-2MR	inserted letters (2, middle random)
ALD-ARB	all-letter-different arbitrary
ALD-PW	all-letter-different pseudoword

quick start

from datasets import load_dataset

dataset = load_dataset("donyin/orthographic-dnn-priming")

or load images directly:

from pathlib import Path
from PIL import Image

prime_dir = Path("prime_data")
target = "design"
condition = "TL12"

img = Image.open(prime_dir / target / f"{condition}.png")

reproducing the main result

the core analysis computes kendall's tau between model cosine-similarity patterns and human priming scores across the 28 conditions. see the source code repository for the full pipeline:

1. fine-tune pretrained torchvision models on word classification (training images not included here; generate with generate_data.py)
2. extract layer-wise activations for each prime image pair (identity vs. condition)
3. compute cosine similarity at each layer
4. correlate with human priming scores using kendall's tau

training data

training images (800k+) are not included due to size. they are fully reproducible:

cd src && python generate_data.py

this requires the font files (not redistributable) and generates images with configurable rotation, translation, font-size variation, and spacing jitter. see src/utils/data_generate/main.py for parameters.

models evaluated

alexnet, densenet169, efficientnet-b1, resnet50, resnet101, vgg16, vgg19, vit-b/16, vit-b/32, vit-l/16, vit-l/32, all initialised from imagenet pretrained weights via torchvision.

citation

@article{yin2023cnn,
  title={Convolutional Neural Networks Trained to Identify Words Provide a Good Account of Visual Form Priming Effects},
  author={Yin, Don and Davelaar, Eddy J.},
  journal={Computational Brain \& Behavior},
  year={2023},
  publisher={Springer},
  doi={10.1007/s42113-023-00172-7}
}

license

MIT