An implementation of the Hedge Backpropagation(HBP) proposed in Online Deep Learning: Learning Deep Neural Networks on the Fly

@inproceedings{sahoo2018online,
  title     = {Online Deep Learning: Learning Deep Neural Networks on the Fly},
  author    = {Doyen Sahoo and Quang Pham and Jing Lu and Steven C. H. Hoi},
  booktitle = {Proceedings of the Twenty-Seventh International Joint Conference on
               Artificial Intelligence, {IJCAI-18}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},             
  pages     = {2660--2666},
  year      = {2018},
  month     = {7},
  doi       = {10.24963/ijcai.2018/369},
  url       = {https://doi.org/10.24963/ijcai.2018/369},
}

Link to publication

• Theano 0.8.2
• Keras 1.2.1

To install HBP, you need to replace the Keras's keras/engine/training.py file with our modified training.py. this doesn't affect normal projects that don't use HBP. Note that as the current HBP implementation only supports Keras 1.

• To run HBP on the sample Higgs dataset, first download the data:

wget -O data/higgs.mat https://www.dropbox.com/s/fvqnhe34cf0mlz9/higgs_100k.mat

• To train HBP, run:

cd src/hbp
python main.py -c hb19.yaml

• To train other baseline models, run:

cd src/baselines
./run.sh

The data used in our experiments are available at https://drive.google.com/drive/folders/1fNZHK2NYbgfz27PPdSSA6lZTkoFakH28?usp=sharing

We provide a sample script in src/train.py to train HBP on a new dataset. Feel free to modify the code to suit your experiments.

This is an independent pytorch implementation, please note that this is unofficial and not yet tested by us.