DECOLLE

Deep Continuous Local Learning (DECOLLE) is a surrogate gradient (SG) algorithm leveraging per-layer local errors for online learning of deep SNNs, for which details can be found in reference Kaiser et al., 2020.

Each layer of the SNN is associated to a fixed, random projection layer to compute the real-valued errors.

This library is meant as an add-on to Intel's Lava framework. It is based on the authors' own implementation, which can be found on their GitHub repository.

Dependencies

lava-dl library (https://github.com/lava-nc/lava-dl/)

Installation

• Follow lava-dl install instruction here.
• cd lava-decolle/src
• python -c "import os; from distutils.sysconfig import get_python_lib; open(get_python_lib() + os.sep + 'lava.pth', 'a').write(os.getcwd())"

Getting Started

NMNIST digit classification tutorial

Comparison with SLAYER

License

This project is licensed under the GPLv3 License - see the LICENSE.txt file for details

Author

Nicolas Skatchkovsky (nskat)