Precision Gating: Improving Neural Network Efficiency with Dynamic Dual-Precision Activations
This work addresses efficiency improvements for deep learning practitioners by reducing computational costs in neural networks, though it is incremental as it builds on existing quantization methods.
The paper tackles the problem of reducing computational cost in deep neural networks by introducing precision gating, a dynamic dual-precision quantization technique that computes most features in low precision and important ones in high precision. It achieves results such as 2.4× less compute on ImageNet with same or higher accuracy compared to state-of-the-art methods and 2.7× computational cost reduction on LSTMs with a 1.2% perplexity improvement.
We propose precision gating (PG), an end-to-end trainable dynamic dual-precision quantization technique for deep neural networks. PG computes most features in a low precision and only a small proportion of important features in a higher precision to preserve accuracy. The proposed approach is applicable to a variety of DNN architectures and significantly reduces the computational cost of DNN execution with almost no accuracy loss. Our experiments indicate that PG achieves excellent results on CNNs, including statically compressed mobile-friendly networks such as ShuffleNet. Compared to the state-of-the-art prediction-based quantization schemes, PG achieves the same or higher accuracy with 2.4$\times$ less compute on ImageNet. PG furthermore applies to RNNs. Compared to 8-bit uniform quantization, PG obtains a 1.2% improvement in perplexity per word with 2.7$\times$ computational cost reduction on LSTM on the Penn Tree Bank dataset. Code is available at: https://github.com/cornell-zhang/dnn-gating