Deep learning has spawned a wide range of AI applications that are changing our lives. However, deep neural networks are both computationally and memory intensive, thus they are power hungry when deployed on embedded systems and data centers with a limited power budget. To address this problem, I will present an algorithm and hardware co-design methodology for improving the efficiency of deep learning.

I will first introduce "Deep Compression", which can compress deep neural network models by 10–49× without loss of prediction accuracy for a broad range of CNN, RNN, and LSTMs. The compression reduces both computation and storage. Next, by changing the hardware architecture and efficiently implementing Deep Compression, I will introduce EIE, the Efficient Inference Engine, which can perform decompression and inference simultaneously, saving a significant amount of memory bandwidth. By taking advantage of the compressed model and being able to deal with an irregular computation pattern efficiently, EIE achieves 13× speedup and 3000× better energy efficiency over GPU. Finally, I will revisit the inefficiencies in current learning algorithms, present DSD training, and discuss the challenges and future work in efficient methods and hardware for deep learning.