Lu Wang

Yuhong He, Long Peng, Qiaosi Yi et al.

Image restoration endeavors to reconstruct a high-quality, detail-rich image from a degraded counterpart, which is a pivotal process in photography and various computer vision systems. In real-world scenarios, different types of degradation can cause the loss of image details at various scales and degrade image contrast. Existing methods predominantly rely on CNN and Transformer to capture multi-scale representations. However, these methods are often limited by the high computational complexity of Transformers and the constrained receptive field of CNN, which hinder them from achieving superior performance and efficiency in image restoration. To address these challenges, we propose a novel Multi-Scale State-Space Model-based (MS-Mamba) for efficient image restoration that enhances the capacity for multi-scale representation learning through our proposed global and regional SSM modules. Additionally, an Adaptive Gradient Block (AGB) and a Residual Fourier Block (RFB) are proposed to improve the network's detail extraction capabilities by capturing gradients in various directions and facilitating learning details in the frequency domain. Extensive experiments on nine public benchmarks across four classic image restoration tasks, image deraining, dehazing, denoising, and low-light enhancement, demonstrate that our proposed method achieves new state-of-the-art performance while maintaining low computational complexity. The source code will be publicly available.

Howe Tissue, Venus Wang, Lu Wang

We find that the cross-entropy loss curves of neural language models empirically adhere to a scaling law with learning rate (LR) annealing over training steps: $$L(s) = L_0 + A\cdot S_1^{-α} - C\cdot S_2,$$ where $L(s)$ is the validation loss at step $s$, $S_1$ is the area under the LR curve, $S_2$ is the LR annealing area, and $L_0$, $A$, $C$, $α$ are constant parameters. This formulation takes into account two factors: (1) power-law scaling over data size, and (2) the additional loss reduction during LR annealing. Therefore, this formulation can describe the full loss curve at each step, rather than the single loss point at the end of training. Applying the scaling law with LR annealing and fitting only one or two training curves, we can accurately predict the loss at any given step across any learning rate scheduler (LRS). This approach significantly reduces computational cost in formulating scaling laws while providing more accuracy and expressiveness for training dynamics. Extensive experiments demonstrate that our findings hold across a range of hyper-parameters and model architectures, and our equation can extend to scaling effect of model sizes. Moreover, our formulation provides accurate theoretical verification and explanation for empirical results observed in numerous previous studies, particularly those focusing on LR schedule and annealing. We believe that this work is promising to enhance the understanding of LLM training dynamics while greatly democratizing scaling laws, and it can guide researchers in refining training strategies (e.g. critical LRS) for further LLMs.

Ruolin Li, Lu Wang, Tingting Yang et al.

Micro-expressions (MEs) are spontaneous, unconscious facial expressions that have promising applications in various fields such as psychotherapy and national security. Thus, micro-expression recognition (MER) has attracted more and more attention from researchers. Although various MER methods have emerged especially with the development of deep learning techniques, the task still faces several challenges, e.g. subtle motion and limited training data. To address these problems, we propose a novel motion extraction strategy (MoExt) for the MER task and use additional macro-expression data in the pre-training process. We primarily pretrain the feature separator and motion extractor using the contrastive loss, thus enabling them to extract representative motion features. In MoExt, shape features and texture features are first extracted separately from onset and apex frames, and then motion features related to MEs are extracted based on the shape features of both frames. To enable the model to more effectively separate features, we utilize the extracted motion features and the texture features from the onset frame to reconstruct the apex frame. Through pre-training, the module is enabled to extract inter-frame motion features of facial expressions while excluding irrelevant information. The feature separator and motion extractor are ultimately integrated into the MER network, which is then fine-tuned using the target ME data. The effectiveness of proposed method is validated on three commonly used datasets, i.e., CASME II, SMIC, SAMM, and CAS(ME)3 dataset. The results show that our method performs favorably against state-of-the-art methods.