Among all the methods that can alleviate computational workload during the training process, BinaryConnect is especially effective by binarizing the weights to +1 and -1 during the training process. However, there has been debate about which binarization method is the best over time. Therefore, we reproduced the work of the original BinaryConnect paper and examined the authors’ results. We also proposed two new binarization methods and discussed their performance.

Report

Source code

https://github.com/HappyWalkers/ESE_539_Project.git