An implementation of the Binarized Neural Networks

Dataset:

MNIST dataset

Method:

There were two kinds of models named binarized neural networks (1-bit) and full-precision model (16-bits). Each model has three building blocks.

Results:

1. Loss curve

2. Macro F1 curve

3. Test confusion matrix of full-precision model

4. Test confusion matrix of binarized neural networks

5. Test result

Model	Epoch	Macro F1	Loss
Full-precision model	33	0.991	0.0382
Binarized neural networks	35	0.978	0.2589

6. Visualization of some kernels of the models

Model	Kernel 0	Kernel 1	Kernel 2
Full-precision model
Binarized neural networks

7. Weights

• Full-precision model: link

• Binarized neural networks: link

References:

Simons, Taylor, and Dah-Jye Lee. "A review of binarized neural networks." Electronics 8.6 (2019): 661.