BNN

BNN is the implementation of a binary-native and gradient-free learning algorithm for Binary Neural Networks. BNN has been introduced in the paper: "Training Multi-Layer Binary Neural Networks With Random Local Binary Error Signals", published in the journal 'IOP Machine Learning: Science and Technology'.

Disclaimer: we are working on the topic, so, expect changes in this repo.

External libraries used

You can install them using uv. First of all you need to install uv (check here for instructions).

Then, you can install all you need using:

uv sync

How to use

Clone the repo, install the dependencies and run the src/train.py script.

python3 src/train.py \
    --algo-layer ["our"/"baldassi"] \
    --dataset ["prototypes"/"fmnist"/"cifar10tl"/"imagenettetl"/"cifar100tl"] \
    --binarize-dataset/--no-binarize-dataset \
    --test-dim [int] \
    --layers [str] \
    --freeze-first/--no-freeze-first \
    --freeze-last/--no-freeze-last \
    --group-size [int] \
    --bs [int] \
    --epochs [int] \
    --prob-reinforcement [float] \
    --rob [float] \
    --seed [int] \
    --n-runs [int] \
    --device [int] \
    --log/--no-log \

Comparisons with SGD and Adam

python3 {sgd/adam}_torch.py \
    --dataset ["prototypes"/"fmnist"/"bfmnist"/"cifar10tl"/"bcifar10tl"/"imagenettetl"/"bimagenettetl"/"cifar100tl"/"bcifar100tl"] \
    --layers [str] \
    --lr [float] \
    --seed [int] \
    --n-runs [int] \
    --device [int] \

Results

To visualize the results as in the paper, look at the VisualizeResults.ipynb notebook.

Authors and Contacts

If you have questions, suggestions or problems, feel free to open an Issue. You can contact us at:

• Luca Colombo: luca2.colombo@polimi.it
• Fabrizio Pittorino: fabrizio.pittorino@polimi.it
• Manuel Roveri: manuel.roveri@polimi.it