Stabilizing Sharpness-aware Minimization Through A Simple Renormalization Strategy

Recently, sharpness-aware minimization (SAM) has attracted a lot of attention because of its surprising effectiveness in improving generalization performance.However, training neural networks with SAM can be highly unstable since the loss does not decrease along the direction of the exact gradient at the current point, but instead follows the direction of a surrogate gradient evaluated at another point nearby. To address this issue, we propose a simple renormalization strategy, dubbed StableSAM, so that the norm of the surrogate gradient maintains the same as that of the exact gradient. Our strategy is easy to implement and flexible enough to integrate with SAM and its variants, almost at no computational cost. With elementary tools from convex optimization and learning theory, we also conduct a theoretical analysis of sharpness-aware training, revealing that compared to stochastic gradient descent (SGD), the effectiveness of SAM is only assured in a limited regime of learning rate. In contrast, we show how StableSAM extends this regime of learning rate and when it can consistently perform better than SAM with minor modification. Finally, we demonstrate the improved performance of StableSAM on several representative data sets and tasks.

Trajectory-dependent Generalization Bounds for Deep Neural Networks via Fractional Brownian Motion

Despite being tremendously overparameterized, it is appreciated that deep neural networks trained by stochastic gradient descent (SGD) generalize surprisingly well. Based on the Rademacher complexity of a pre-specified hypothesis set, different norm-based generalization bounds have been developed to explain this phenomenon. However, recent studies suggest these bounds might be problematic as they increase with the training set size, which is contrary to empirical evidence. In this study, we argue that the hypothesis set SGD explores is trajectory-dependent and thus may provide a tighter bound over its Rademacher complexity. To this end, we characterize the SGD recursion via a stochastic differential equation by assuming the incurred stochastic gradient noise follows the fractional Brownian motion. We then identify the Rademacher complexity in terms of the covering numbers and relate it to the Hausdorff dimension of the optimization trajectory. By invoking the hypothesis set stability, we derive a novel generalization bound for deep neural networks. Extensive experiments demonstrate that it predicts well the generalization gap over several common experimental interventions. We further show that the Hurst parameter of the fractional Brownian motion is more informative than existing generalization indicators such as the power-law index and the upper Blumenthal-Getoor index.

Understanding Long Range Memory Effects in Deep Neural Networks

\textit{Stochastic gradient descent} (SGD) is of fundamental importance in deep learning. Despite its simplicity, elucidating its efficacy remains challenging. Conventionally, the success of SGD is attributed to the \textit{stochastic gradient noise} (SGN) incurred in the training process. Based on this general consensus, SGD is frequently treated and analyzed as the Euler-Maruyama discretization of a \textit{stochastic differential equation} (SDE) driven by either Brownian or L\'evy stable motion. In this study, we argue that SGN is neither Gaussian nor stable. Instead, inspired by the long-time correlation emerging in SGN series, we propose that SGD can be viewed as a discretization of an SDE driven by \textit{fractional Brownian motion} (FBM). Accordingly, the different convergence behavior of SGD dynamics is well grounded. Moreover, the first passage time of an SDE driven by FBM is approximately derived. This indicates a lower escaping rate for a larger Hurst parameter, and thus SGD stays longer in flat minima. This happens to coincide with the well-known phenomenon that SGD favors flat minima that generalize well. Four groups of experiments are conducted to validate our conjecture, and it is demonstrated that long-range memory effects persist across various model architectures, datasets, and training strategies. Our study opens up a new perspective and may contribute to a better understanding of SGD.

Discrete Cosine Transform Network for Guided Depth Map Super-Resolution

Guided depth super-resolution (GDSR) is a hot topic in multi-modal image processing. The goal is to use high-resolution (HR) RGB images to provide extra information on edges and object contours, so that low-resolution depth maps can be upsampled to HR ones. To solve the issues of RGB texture over-transferred, cross-modal feature extraction difficulty and unclear working mechanism of modules in existing methods, we propose an advanced Discrete Cosine Transform Network (DCTNet), which is composed of four components. Firstly, the paired RGB/depth images are input into the semi-coupled feature extraction module. The shared convolution kernels extract the cross-modal common features, and the private kernels extract their unique features, respectively. Then the RGB features are input into the edge attention mechanism to highlight the edges useful for upsampling. Subsequently, in the Discrete Cosine Transform (DCT) module, where DCT is employed to solve the optimization problem designed for image domain GDSR. The solution is then extended to implement the multi-channel RGB/depth features upsampling, which increases the rationality of DCTNet, and is more flexible and effective than conventional methods. The final depth prediction is output by the reconstruction module. Numerous qualitative and quantitative experiments demonstrate the effectiveness of our method, which can generate accurate and HR depth maps, surpassing state-of-the-art methods. Meanwhile, the rationality of modules is also proved by ablation experiments.

Deep Convolutional Sparse Coding Network for Pansharpening with Guidance of Side Information

Pansharpening is a fundamental issue in remote sensing field. This paper proposes a side information partially guided convolutional sparse coding (SCSC) model for pansharpening. The key idea is to split the low resolution multispectral image into a panchromatic image related feature map and a panchromatic image irrelated feature map, where the former one is regularized by the side information from panchromatic images. With the principle of algorithm unrolling techniques, the proposed model is generalized as a deep neural network, called as SCSC pansharpening neural network (SCSC-PNN). Compared with 13 classic and state-of-the-art methods on three satellites, the numerical experiments show that SCSC-PNN is superior to others. The codes are available at https://github.com/xsxjtu/SCSC-PNN.

Deep Gradient Projection Networks for Pan-sharpening

Pan-sharpening is an important technique for remote sensing imaging systems to obtain high resolution multispectral images. Recently, deep learning has become the most popular tool for pan-sharpening. This paper develops a model-based deep pan-sharpening approach. Specifically, two optimization problems regularized by the deep prior are formulated, and they are separately responsible for the generative models for panchromatic images and low resolution multispectral images. Then, the two problems are solved by a gradient projection algorithm, and the iterative steps are generalized into two network blocks. By alternatively stacking the two blocks, a novel network, called gradient projection based pan-sharpening neural network, is constructed. The experimental results on different kinds of satellite datasets demonstrate that the new network outperforms state-of-the-art methods both visually and quantitatively. The codes are available at https://github.com/xsxjtu/GPPNN.

FGF-GAN: A Lightweight Generative Adversarial Network for Pansharpening via Fast Guided Filter

Pansharpening is a widely used image enhancement technique for remote sensing. Its principle is to fuse the input high-resolution single-channel panchromatic (PAN) image and low-resolution multi-spectral image and to obtain a high-resolution multi-spectral (HRMS) image. The existing deep learning pansharpening method has two shortcomings. First, features of two input images need to be concatenated along the channel dimension to reconstruct the HRMS image, which makes the importance of PAN images not prominent, and also leads to high computational cost. Second, the implicit information of features is difficult to extract through the manually designed loss function. To this end, we propose a generative adversarial network via the fast guided filter (FGF) for pansharpening. In generator, traditional channel concatenation is replaced by FGF to better retain the spatial information while reducing the number of parameters. Meanwhile, the fusion objects can be highlighted by the spatial attention module. In addition, the latent information of features can be preserved effectively through adversarial training. Numerous experiments illustrate that our network generates high-quality HRMS images that can surpass existing methods, and with fewer parameters.

Domain Adaptive Object Detection via Feature Separation and Alignment

Recently, adversarial-based domain adaptive object detection (DAOD) methods have been developed rapidly. However, there are two issues that need to be resolved urgently. Firstly, numerous methods reduce the distributional shifts only by aligning all the feature between the source and target domain, while ignoring the private information of each domain. Secondly, DAOD should consider the feature alignment on object existing regions in images. But redundancy of the region proposals and background noise could reduce the domain transferability. Therefore, we establish a Feature Separation and Alignment Network (FSANet) which consists of a gray-scale feature separation (GSFS) module, a local-global feature alignment (LGFA) module and a region-instance-level alignment (RILA) module. The GSFS module decomposes the distractive/shared information which is useless/useful for detection by a dual-stream framework, to focus on intrinsic feature of objects and resolve the first issue. Then, LGFA and RILA modules reduce the distributional shifts of the multi-level features. Notably, scale-space filtering is exploited to implement adaptive searching for regions to be aligned, and instance-level features in each region are refined to reduce redundancy and noise mentioned in the second issue. Various experiments on multiple benchmark datasets prove that our FSANet achieves better performance on the target domain detection and surpasses the state-of-the-art methods.

MFIF-GAN: A New Generative Adversarial Network for Multi-Focus Image Fusion

Multi-Focus Image Fusion (MFIF) is one of the promising techniques to obtain all-in-focus images to meet people's visual needs and it is a precondition of other computer vision tasks. One of the research trends of MFIF is to solve the defocus spread effect (DSE) around the focus/defocus boundary (FDB). In this paper, we present a novel generative adversarial network termed MFIF-GAN to translate multi-focus images into focus maps and to get the all-in-focus images further. The Squeeze and Excitation Residual Network (SE-ResNet) module as an attention mechanism is employed in the network. During the training, we propose reconstruction and gradient regularization loss functions to guarantee the accuracy of generated focus maps. In addition, by combining the prior knowledge of training conditon, this network is trained on a synthetic dataset with DSE based on an {\alpha}-matte model. A series of experimental results demonstrate that the MFIF-GAN is superior to several representative state-of-the-art (SOTA) algorithms in visual perception, quantitative analysis as well as efficiency.