Network quantization is promising to alleviate the storage and deployment burden for deep neural networks. However, existing methods focus on directly quantizing the network parameters but ignore the correlations between the initialized parameters (generated by random seeds) and the optimized parameters (obtained when the network training converges). In this paper, we reveal the redundancy between the initialized and optimized parameters and propose a novel quantization scheme for network parameters with rescaled initialization. We develop a novel quantization scheme for quantizing the delta parameters, i.e. , the difference between the optimized parameters and the rescaled initialization, as shown in Figure 1 . The scale is learned to adapt the weight regularization in training by minimizing the rounding error and penalizing the long-tailed distribution of the delta parameter. Adaptive rounding is further presented to calibrate the rounding direction of the delta parameter via a learnable bias. The proposed method can achieve high-efficiency prediction of network parameters and consistent performance gains with a trivial overhead of single random seed. Extensive experimental results demonstrate its superior performance on various popular network architectures, as exemplified in Table 1 .

End-to-end learned image compression (LIC) has become promising alternatives for lossy image compression. However, deployments of LIC models are limited in real-world applications due to excessive network parameters and high computational complexity. Existing LIC models realized with integer networks are significantly de-graded in rate-distortion (R-D) performance. In this paper, we propose a novel fully integerized LIC model that simultaneously achieves channel-wise weight and 8-bit activation quantization for alleviating the loss of R-D performance. For weight quantization, we develop an internal bit width increment (IBWI) via nonlinear logarithmic mapping to convert the stored low-precision integer weights to high bit-width weights for inference. Moreover, outlier channel splitting (OCS) [1] is employed to address large outliers of weight distribution by duplicating channels and constrain the integer weights within the scope of INT8. For activation quantization, we leverage activation equalization to balance the channel-wise distribution of activations. Experimental results demonstrate that the proposed method achieves a reduction of 75% storage cost with subtle performance loss compared to full-precision pre-trained models, and outperforms existing integer-only networks, as shown in Table 1.

Quantizing one single deep neural network into multiple compression rates (precisions) has been recently considered for flexible deployments in real-world scenarios. In this paper, we propose a novel scheme that achieves progressive bit-width allocation and joint training to simultaneously optimize mixed-precision quantized networks under multiple compression rates. Specifically, we develop a progressive bit-width allocation with switchable quantization step size to enable mixed-precision quantization based on analytic sensitivity of network layers under multiple compression rates. Furthermore, we achieve joint training for quantized networks under different compression rates via knowledge distillation to exploit their correlations based on the shared network structure. Experimental results show that the proposed scheme outperforms AdaBits [1] in various networks on CIFAR-10 and ImageNet.