Multi-level projection with exponential parallel speedup; Application to sparse auto-encoders neural networks