A controller-peripheral architecture and costly energy principle for learning

Description

This is a complementary repo to which focuses on DNN model finetuning which then used as part of the proposed architecture in this research paper.

Structure and main components

This repo follows a fairly generic ML-project organisation:

1. main.py is the top executable that runs the finetuning process. <br > Depending on the --mode specified (see Example usage), the script will either be running training by first preparing datasets from raw data and training models and saving them to disk

if mode == 'train':
    dataset_dir = f'stimuli/{preprocessed_dir}/{model_name}/' \
              f'{layer}_reprs/task{stimulus_set}'
    # if no dataset, create dataset 
    if not os.path.exists(dataset_dir):
        print('[Check] Preparing dataset..')
        data.execute(config)

    print('[Check] Start training..')
    train.execute(config_version)

or the script will be evaluating trained models on tests

elif mode == 'eval':
    evaluations.execute(
        config_version, 
        full_test, 
        heldout_test)

2. models.py contains model definitions of the candidate DNNs.
3. train.py contains training logics for finetuning specified models.
4. evaluations.py contains evaluation routines for trained models.
5. utils.py contains general utilites and data loading mechanisms.
6. data.py contains code for raw data preprocessing and dataset preparation. Prepared datasets will be saved to disk.
7. config_generator.py produces different model configurations in batch.
8. keras_custom contains customised TF/keras functionalities, specifically

.
├── keras_custom
│   ├── generators          # base classes with customised components of TF generators which are called by data loaders in `utils.py`
│   ├── callbacks.py        # customised TF callbacks.
│   └── regularizers.py     # customised TF regularizers.

Example usage

1. Run finetuning on task2 on GPU-0

python main.py --mode train --task 2 --gpu 0

Attribution

@article {Luo2023.01.16.524194,
    author = {Xiaoliang Luo and Robert M. Mok and Brett D. Roads and Bradley C. Love},
    title = {A controller-peripheral architecture and costly energy principle for learning},
    elocation-id = {2023.01.16.524194},
    year = {2023},
    doi = {10.1101/2023.01.16.524194},
    publisher = {Cold Spring Harbor Laboratory},
}