Avoiding Spurious Local Minima in Deep Quadratic Networks

In this repository, we provide the codes used to create the results in the following paper:

Abbas Kazemipour, Brett W. Larsen, and Shaul Druckmann. Avoiding Spurious Local Minima in Deep Quadratic Networks. arXiv:2001.00098, 2019. https://arxiv.org/abs/2001.00098

Figures 2 and 3: Two-Layer Networks

• varyHidden.py - Main script with data generation
• varyHidden_load.py - Main script for loading dataset
• networkFiles.py - Network classes

Figure 4 and 5: Deep Networks

• varyHidden_DeepSynthetic.py - Main script for experiments in Figure 4
• varyHidden_DeepMNIST.py - Main script for experiments in Fgiure 5
• networkFiles.py - Network classes
• TensorGenerator.py - Helper function BestLinearMap
• ExpData/dataset38.pth.tar - Dataset of MNIST 3 and 8 digits
• ExpData/dataset47.pth.tar - Dataset of MNIST 4 and 7 digits