Efficient and Robust Deep Learning through Approximate Computing