Dynamic Auto-Sizing

Implementation of the dynamic auto-sizing technique which we presented in the paper Adapting the Size of Artificial Neural Networks Using Dynamic Auto-Sizing. Dynamic auto-sizing allows artificial neural networks to automatically adapt their size to the problem domain. Besides the TensorFlow implementation of dynamic auto-sizing, this repository additionally contains the complete Jupyter notebooks with experiments.

Repository structure

• nets - Python package with the implementations
• notebooks - Jupyter notebooks with experiments
  • standalone - Experiments with dynamic auto-sizing
  • anytime - Experiments with dynamic auto-sizing as an underlying technique for anytime algorithms
  • misc - Supporting code, e.g. for the rendering of figures

How to run

1. Clone the repository, cd into it.
2. Install required packages by running pip install -r requirements.txt (Python version 3.9 is recommended).
3. Open Jupyter Lab using jupyter lab.

Citing

In case you find this repository dataset helpful, feel free to cite our related publication Adapting the Size of Artificial Neural Networks Using Dynamic Auto-Sizing:

@inproceedings{10000471,
    title        = {Adapting the Size of Artificial Neural Networks Using Dynamic Auto-Sizing},
    author       = {Cahlik, Vojtech and Kordik, Pavel and Cepek, Miroslav},
    year         = 2022,
    booktitle    = {2022 IEEE 17th International Conference on Computer Sciences and Information Technologies (CSIT)},
    volume       = {},
    number       = {},
    pages        = {592--596},
    doi          = {10.1109/CSIT56902.2022.10000471}
}