The Costs and Benefits of Goal-Directed Attention in Deep Convolutional Neural Networks

[Original paper]

Set up environment

conda create --name myenv --file conda_env/attn_tf22_py37.txt, pip install tensorflow-gpu==2.2.0
or conda install tensorflow-gpu==2.2.0

Train models

• To learn corresponding attention weights, python META_train --option attn_layer
  
• To learn corresponding final layer weights, python META_train --option last_layer

Evaluate models

• To evaluate trained models on tests and generate intermediate results, such as probabilities, python META_eval --option eval --exp <num> --task <task code>
  
• To plot the final results, python META_eval --option plot --exp <num> --task <task code>.
  
• <num> is a number among 1, 2 or 3. <task code> is a string between EXP or retrain where EXP corresponds to models trained using attention layer and retrain corresponds to training the last layer without attention.

Repo organisation

• Model definition, custom data generator, fitting function can be found in top_down_attention/keras_custom/
• Model training code can be found in top_down_attention/TRAIN/
• Model evaluation and results plotting code can be found in top_down_attention/EVAL/

Attribution

@article{Luo2021TheNetworks,
    title = {{The costs and benefits of goal-directed attention in deep convolutional neural networks}},
    year = {2021},
    journal = {Computational Brain {\&} Behavior},
    author = {Luo, Xiaoliang and Roads, Brett D. and Love, Bradley C.},
    month = {2},
    pages = {1--18},
    url = {https://doi.org/10.1007/s42113-021-00098-y},
    doi = {10.1007/s42113-021-00098-y},
    issn = {23318422},
}