Unreasonable effectiveness of learning neural networks: From accessible states and robust ensembles to basic algorithmic schemes