Pruning and quantization for deep neural network acceleration: A survey

Deep neural networks have been applied in many applications exhibiting extraordinary abilities the field of computer vision. However, complex network architectures challenge efficient real-time deployment and require significant computation resources energy costs. These challenges can be overcome through optimizations such as compression. Network compression often realized with little loss accuracy. In some cases accuracy may even improve. This paper provides a survey on two types compression: pruning quantization. Pruning categorized static if it is performed offline or dynamic at run-time. We compare techniques describe criteria used to remove redundant computations. discuss trade-offs element-wise, channel-wise, shape-wise, filter-wise, layer-wise network-wise pruning. Quantization reduces computations by reducing precision datatype. Weights, biases, activations quantized typically 8-bit integers although lower bit width implementations are also discussed including binary networks. Both quantization independently combined. current techniques, analyze their strengths weaknesses, present compressed results number frameworks, provide practical guidance for compressing