Long-tailed data distribution (i.e., minority classes occupy most of the data, while most classes have very few samples) is a common problem in image classification. In this paper, we propose a novel multimodal framework for long-tailed data recognition. In the first stage, long-tailed data are used for visual-semantic contrastive learning to obtain good features, while in the second stage, class-balanced data are used for classifier training. The proposed framework leverages the advantages of multimodal models and mitigates the problem of class imbalance in long-tailed data recognition. Experimental results demonstrate that the proposed framework achieves competitive performance on the CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist2018 datasets for image classification.
Multi-agent communication allows agents to share local information or their own intention with other agents, thus enhancing the collaborative performance of the multi-agent system. Despite its significance, previous multi-agent communication methods typically assume agents communicate in a perfect, interference-free environment without any communication disturbance. However, in real-world scenarios, communication within multi-agent systems often faces environmental noise, and even in adversarial settings, agents may encounter communication attacks from adversaries. How to maintain the collaborative stability within the multi-agent system under communication attack remains crucial yet inadequately researched. In this paper, we build a robust communication mechanism that utilizes the denoising process of diffusion models to recover the original communicated messages, thus preventing the influence of communication interference. Besides, to reduce the computational burden brought by diffusion models, we design a detection module to recognize when the messages have been distorted, thus adopting the diffusion model only when necessary. The experiments show that, with this diffusion-based message denoising mechanism, our approach shows significantly superior communication robustness than existing baselines. Our approach exhibits excellent communication robustness in the presence of various types and different degrees of noise attacks. Visualization experiments and ablation studies validate the effectiveness of our diffusion model in recovering the original messages.
Industrial private networks (IPNs), having enhanced communication characteristics in a specific area, emerge to fulfill the demanding industrial use cases. However, users' demands for flexibility and low latency in the B5G era are in inevitable conflict with limited and isolated resources in IPNs; thus, scalability and flexibility of resources are required, which can be realized by jointly funneling users' traffic and resource allocation spanning IPN s to promote load balancing and resource efficiency. In this paper, we devise a cost-aware resource allocation (CARA) approach embedded with a proba-bilistic latency guarantee for resource efficiency achievement and low latency fulfillment. Specifically, we first establish a unified cost model for coupling funneled traffic and resource allocated in each lPN, avoiding the optimization penalty of alternating them. Then, to solve the conflict between limited computing and communication resources, we propose the CARA approach based on the non-dominated sorting genetic algorithm-III (NSGA-III). Furthermore, a probabilistic latency guarantee sub-algorithm is embedded in CARA to fulfill the latency constraint and relax it for advanced industrial implementation. Additionally, compared with other existing algorithms, simulation results reveal that our proposed algorithm not only globally minimizes unified cost across IPNs, but also individually balances the funneled traffic.