In this project, we answer two main questions:
- Bad global minimum do exist (here we refer the bad minimum as the model that can perfectly fit the training data, yet it generatizes poorly).
- We can construct such bad global minimum using unlabeled training data only, without the knowledge of the loss-landscape.
A toy example is given as follows, for the complete results, feel free to check out the full paper.
wget -q –retry-connrefused –waitretry=10 https://repo.continuum.io/archive/Anaconda2-4.3.1-Linux-x86_64.sh
chmod 777 *
./Anaconda2-4.3.1-Linux-x86_64.sh -b -p ./anaconda > /dev/null
chmod 777 *
conda install --yes pyyaml > /dev/null
conda install --yes HDF5 > /dev/null
conda install --yes h5py > /dev/null
conda install --yes -c rdonnelly libgpuarray > /dev/null
conda install --yes -c rdonnelly pygpu > /dev/null
conda install --yes pytorch=0.3.1 torchvision -c soumith > /dev/null
chmod 777 -R ./anaconda
@article{liu2020bad,
title={Bad global minima exist and sgd can reach them},
author={Liu, Shengchao and Papailiopoulos, Dimitris and Achlioptas, Dimitris},
journal={Advances in Neural Information Processing Systems},
volume={33},
year={2020}
}