[Re] Latent Weights Do Not Exist: Rethinking Binarized Neural Network Optimization

Nik Vaessen, Peter Mastnak, Sri Datta Budaraju

Abstract: This report examines the reproducibility of the paper "Latent Weights Do Not Exist: Rethinking Binarized Neural Network Optimization". This paper proposes a novel view on how binary neural networks are trained and proposes a new optimisation algorithm called Bop (Binary optimisation) implemented in TensorFlow. By re-implementing Bop in PyTorch, we reproduce experiments that study the effect of the hyperparameters tau and gamma used by the algorithm. We also reproduce training a binary neural network on CIFAR-10 with Bop and achieve similar accuracy. Our code repository can be found at https://github.com/nikvaessen/Rethinking-Binarized-Neural-Network-Optimization

Track: Replicability

NeurIPS Paper Id: https://openreview.net/forum?id=BJe_nNrgIS