GRADIENT DESCENT PROVABLY OPTIMIZES OVER-PARAMETERIZED NEURAL NETWORKS

this work: two layer fc + ReLU able to achieve a globally optimal at linear rate using GD

• observation: zero error on random label
• proof…