Paper

Learning Rate Decay Can Exponentially Accelerate SGD for Global Optimization of Nonconvex Functions · ByteTrail