Beating the Perils of Non-Convexity in Neural Nets