A-SOINN+

The associative SOINN+ (A-SOINN+) experiments and the v-NICO-World-LL dataset for continual learning proposed in Learning Then, Learning Now, and Every Second in Between: Lifelong Learning With a Simulated Humanoid Robot.

v-NICO-World-LL

Original Dataset

The original v-NICO-World-LL dataset can be downloaded with the command:

wget -r --no-parent -nH --cut-dirs=2 https://www2.informatik.uni-hamburg.de/wtm/datasets/20220419_v_NICO_World_LL/

Note that the files are compressed and have to be decompressed with the command:

tar -zxvf s<i>.tgz
# Example: tar -zxvf 20220419_v_NICO_World_LL/s0.tgz

Feature Vectors

You don't have to download the dataset to start the experiments. For the experiments, use v-NICO-World-LL-feature-vectors. The folder contains feature vectors for each image of the original v-NICO-World-LL dataset. These features are created using a pre-trained VGG16 model as described in our paper.

Experiments

Requirements

• Python 3.8.10+

pip install -r requirements.txt

Usage

Both the A-SOINN+ and the Growing-Dual Memory (GDM), proposed by German I. Parisi et al., can be trained with the following command:

./run_training.sh -c <path/to/model/config.yml> -d <path/to/dataset>
# Example A-SOINN+: ./run_training.sh -c src/configs/asoinn_plus/config.yml -d data/v_NICO_World_LL_feature_vectors
# Example GDM: ./run_training.sh -c src/configs/gdm/config.yml -d data/v_NICO_World_LL_feature_vectors

The config files of the GDM and A-SOINN+ approaches contain hyperparameters that can be adjusted. The default values are the same as in our paper.

Note

The GDM reimplementation is based on the original GDM implementation of German I. Parisi et al. https://github.com/giparisi/GDM, proposed in:

G. I. Parisi, J. Tani, C. Weber, and S. Wermter, “Lifelong Learning of Spatiotemporal Representations With Dual-Memory Recurrent Self-Organization,” Front. Neurorobot., vol. 12, 2018, doi: 10.3389/fnbot.2018.00078.

License

The v-NICO-World-LL dataset and the A-SOINN+ approach are distributed under the Creative Commons CC BY-NC-ND 4.0 license. If you use them, you agree (i) to use them for research purposes only, and (ii) to cite the following reference in any works that make any use of the dataset or the approach.

Citation

A. Logacjov, M. Kerzel, and S. Wermter, “Learning Then, Learning Now, and Every Second in Between: Lifelong Learning With a Simulated Humanoid Robot,” Frontiers in Neurorobotics, vol. 15, p. 78, 2021, doi: 10.3389/fnbot.2021.669534.