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