A Deep Learning Account of How Language Affects Thought

[Original paper]

Environment setup

conda env create -f conda_envs/env.yml

Reproducing results

• This command uses pre-computed intermediate results stored in resources_val_white/.
• label_type should be set to finegrain or coarsegrain.

python main_eval.py -l <label_type> -f simclr -v v3.1.run12 -p True -gpu <num_gpu>

Reproducing results from scratch

1. Download pre-trained SimCLR (i.e., unsupervised front end) model

gsutil -m cp -r \
  "gs://simclr-checkpoints-tf2/simclrv2/pretrained/r50_1x_sk0/" \
  .

• If you get AttributeError: module 'pyparsing' has no attribute 'downcaseTokens' from running the above code,
  see this issue for a workaround.
• More info about SimCLR model can be found at the Official repo for SimCLR

2. Prepare dataset for training (a full path to ImageNet-2012 should be set in TRAIN/utils/data_utils.py)

python data.py --model simclr

3. Train the models

python main_train.py -l <label_type> -f simclr -v v3.1.run12 -r True -gpu <num_gpu>

4. Evaluate trained models and plot results

python main_eval.py -l <label_type> -f simclr -v v3.1.run12 -s True -m True -p True -gpu <num_gpu>

Attribution

@article{Luo2021DeepLangThought,
    author = {Xiaoliang Luo and Nicholas J. Sexton and Bradley C. Love},
    title = {A deep learning account of how language affects thought},
    journal = {Language, Cognition and Neuroscience},
    volume = {38},
    number = {4},
    pages = {499-508},
    year  = {2023},
    publisher = {Routledge},
    doi = {10.1080/23273798.2021.2001023},
}