Sparse Deep Learning: A New Framework Immune to Local Traps and Miscalibration