Noise label learning has attracted considerable attention owing to its ability to leverage large amounts of inexpensive and imprecise data. Sharpness aware minimization (SAM) has shown effective improvements in the generalization performance in the presence of noisy labels by introducing adversarial weight perturbations in the model parameter space. However, our experimental observations have shown that the SAM generalization bottleneck primarily stems from the difficulty of finding the correct adversarial perturbation amidst the noisy data. To address this problem, a theoretical analysis of the mismatch in the direction of the parameter perturbation between noise and clean samples during the training process was conducted. Based on these analyses, a clean aware sharpness aware minimization algorithm known as CA-SAM is proposed. CA-SAM dynamically divides the training data into possible likely clean and noisy datasets based on the historical model output and uses likely clean samples to determine the direction of the parameter perturbation. By searching for flat minima in the loss landscape, the objective was to restrict the gradient perturbation direction of noisy samples to align them while preserving the clean samples. By conducting comprehensive experiments and scrutinizing benchmark datasets containing diverse noise patterns and levels, it is demonstrated that our CA-SAM outperforms certain innovative approaches by a substantial margin.

Many studies in recent years have focused on social image understanding due to the increasing number of shared images from social networks and online communities. However, previous work in social image understanding fails to learn an effective feature representation because of a large amount of missing and irrelevant tags, though matrix completion techniques are frequently utilized for this purpose. Autoencoder models have been validated to be effective in learning latent low-dimensional representations in unsupervised learning. In this paper, we propose a new social image understanding model based on deep autoencoders, which can learn the shared latent codes of social images and tags as supervision information in the deep autoencoders. First, social images are extracted with multi-modal features, which provides a comprehensive characterization to image semantic understanding. And, the social image understanding problems are transformed into the problem of minimizing an optimization objective. Second, multi-layered autoencoders with weak supervision integration are employed to learn an efficient low-dimensional representation from the multi-view feature sources that can make up the semantic gap between image features and tags through minimizing the problem formulation. Finally, we design a new balanced loss function based on binary cross entropy, in which we address highly sparse inputs for a better optimization performance. Extensive experiments on several real-world social image datasets confirm the effectiveness and robustness of the proposed model compared with the state-of-the-art methods.

Coherent frequency agile radar has a strong anti-jamming ability, but it may encounter difficulty in computational effort while reconstructing target using compressed sensing. In this paper, an acceleration method and implementation of Orthogonal Matching Pursuit algorithm (OMP) based on CUDA are developed to meet the needs of real-time application of sparse reconstruction algorithm of coherent frequency agile radar target. The algorithms are tested on a computer with a NVIDIA Rtx2070 Super GPU and an Inter-i5-10400f CPU. Computational results indicate that the proposed parallelization method has a good acceleration effect on compressed sensing reconstruction calculation.