This is the official code for TMLR paper, "Dual Cognitive Architecture: Incorporating Biases and Multi-Memory Systems for Lifelong Learning" by Shruthi Gowda, Elahe Arani and Bahram Zonooz.

Methodology

DUCA framework

Dual Cognitive Architecture (DUCA) : The explicit module features a working model that is dedicated to learning direct sensory data. In the implicit module, the inductive bias learner encodes prior shape-related knowledge, while the semantic memory submodule consolidates information originating from the explicit module. During the inference process, a single network, specifically the semantic memory, is employed, as it contains consolidated knowledge encompassing all tasks.

New Dataset - "DN4IL"

We introduce a new dataset DN4IL for the Domain-IL setting. More details are provided here : https://github.com/NeurAI-Lab/DN4IL-dataset

Requirements

• python==3.8.0
• torch==1.10.0
• torchvision==0.8.0

Setup

DUCA is trained and tested on four different datasets using ResNet18 architecture. The learning rate is set to 0.03, batch size to 32, and epochs to 50 respectively for all the datasets. The decay factor d is always set to 0.999.

Running

Train DUCA - DN4IL

best_params_dn4il = {
    200: {'lr': 0.05,
          'batch_size': 32,
          'n_epochs': 50,
          'aux': 'shape',
          'img_size': 64,
          'shape_filter': 'sobel',
          'lamda': 0.1,
          'gamma': 0.01,
          'd': 0.999,
          'r': 0.06
          },
    500: {'lr': 0.05,
          'batch_size': 32,
          'n_epochs': 50,
          'aux': 'shape',
          'img_size': 64,
          'shape_filter': 'sobel',
          'lamda': 0.1,
          'gamma': 0.01,
          'd': 0.999,
          'r':0.08
          },
}

buffer_size = 200 (or 500)
train_params = best_params_dn4il[buffer_size]
python main.py 
    --experiment_id domainIL_exp_dn4il_dataset \
    --seed 0 \
    --model duca \
    --dataset dn4il \
    --buffer_size {buffer_size} \
    --aux {train_params['aux']} \
    --lr {train_params['lr']} \
    --n_epochs {train_params['n_epochs']} \
    --img_size {train_params['img_size']} \
    --shape_filter {train_params['shape_filter']} \
    --batch_size {train_params['batch_size']} \
    --output_dir /output/ \
    --loss_wt {train_params['lamda']} {train_params['gamma']}
    --ema_alpha {d} \
    --ema_update_freq {r} \
    --tensorboard \
    --csv_log \

Cite Our Work

@article{
  gowda2023dual,
  title={Dual Cognitive Architecture: Incorporating Biases and Multi-Memory Systems for Lifelong Learning},
  author={Shruthi Gowda and Bahram Zonooz and Elahe Arani},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2023},
  url={https://openreview.net/forum?id=PEyVq0hlO3}
}

License

This project is licensed under the terms of the MIT license.