Towards Ultra-Efficient Machine Learning For Edge Inference