Parameter Disentanglement for Diverse Representations

Recent advances in neural network architectures reveal the importance of diverse representations. However, simply integrating more branches or increasing the width for the diversity would inevitably increase model complexity, leading to prohibitive inference costs. In this paper, we revisit the learnable parameters in neural networks and showcase that it is feasible to disentangle learnable parameters to latent sub-parameters, which focus on different patterns and representations. This important finding leads us to study further the aggregation of diverse representations in a network structure. To this end, we propose Parameter Disentanglement for Diverse Representations (PDDR), which considers diverse patterns in parallel during training, and aggregates them into one for efficient inference. To further enhance the diverse representations, we develop a lightweight refinement module in PDDR, which adaptively refines the combination of diverse representations according to the input. PDDR can be seamlessly integrated into modern networks, significantly improving the learning capacity of a network while maintaining the same complexity for inference. Experimental results show great improvements on various tasks, with an improvement of 1.47% over Residual Network 50 (ResNet50) on ImageNet, and we improve the detection results of Retina Residual Network 50 (Retina-ResNet50) by 1.7% Mean Average Precision (mAP). Integrating PDDR into recent lightweight vision transformer models, the resulting model outperforms related works by a clear margin.