Yang Luo's research while affiliated with University of Electronic Science and Technology of China and other places

Federated learning (FL) is widely regarded as highly promising because it enables the collaborative training of high-performance machine learning models among a large number of clients while preserving data privacy by keeping the data local. However, many existing FL frameworks have a two-layered architecture, thus requiring the frequent exchange of large-scale model parameters between clients and remote cloud servers over often unstable networks and resulting in significant communication overhead and latency. To address this issue, we propose to introduce edge servers between the clients and the cloud server to assist in aggregating local models, thus combining asynchronous client–edge model aggregation with synchronous edge–cloud model aggregation. By leveraging the clients’ idle time to accelerate training, the proposed framework can achieve faster convergence and reduce the amount of communication traffic. To make full use of the grouping properties inherent in three-layer FL, we propose a similarity matching strategy between edges and clients, thus improving the effect of asynchronous training. We further propose to introduce model-contrastive learning into the loss function and personalize the clients’ local models to address the potential learning issues resulting from asynchronous local training in order to further improve the convergence speed. Extensive experiments confirm that our method exhibits significant improvements in model accuracy and convergence speed when compared with other state-of-the-art federated learning architectures.

This paper proposes a novel approach for optimizing power allocation in multi-cell multi-user (MCMU) networks using a hypergraph neural network (HGNN). In MCMU networks, each base station (BS) serves multiple user equipments (UEs). This multivariate and implicit connection introduces computational overheads and is unsuitable for pairwise relationship modeling. To address this challenge, we first propose a hypergraph structure that represents BSs as hyperedges, capturing the complex interactions and multiple dependencies within the network. Second, a synchronous training loss is developed, which includes negative weighted sum rate and parameter regularity terms. The first term can learn the distribution without relying on labeled data. The second term avoids overfitting and improves scalability. Third, power constraints are embedded into the network architecture to ensure the feasibility of the power allocation. Extensive simulations demonstrate that our proposed HGNN achieves higher sum rate than the baselines and exhibits its excellent scalability with the increase of complexity in future networks.

In model-driven deep learning (DL)-based channel estimation methods for orthogonal frequency division multiplexing (OFDM) systems, all input features obtained through processed raw signals are considered to have equal weights, regardless of their significance or correlation to the channel predictions, limiting the estimation performance that can be achieved. This letter proposes an attention-aided residual channel estimation network, namely AttenReEsNet, for enhanced estimation performance. AttenReEsNet utilizes the channel attention module embedded in each attention residual block (AttenResBlock) to successively reweight the channel features of the least square estimation input with a learned attention map and combine them with the residual input, improving the learning efficiency and estimation accuracy of the overall network. Numerical results based on the mean square error metric demonstrate that the proposed method outperforms the benchmark methods. Furthermore, we propose to compress the AttenReEsNet model’s parameters by 74% without significantly affecting the estimation performance and also implement the model pruning method to further reduce its complexity, to provide insight for efficient deployment in mobile devices.

In the real world, relationships between objects are often complex, involving multiple variables and modes. Hypergraph neural networks possess the capability to capture and represent such intricate relationships by deriving and inheriting their graph-based counterparts. Nevertheless, both graph and hypergraph neural networks suffer from the problem of over-smoothing when multiple graph convolution layers are stacked. To address this issue, this paper introduces the Multi-modal Hypergraph Neural Network with Parametric Filtering and Feature Sampling (MHNet) to encode complex hypergraph features and mitigate over-smoothing. The proposed approach uses hypergraph structures to model high-order and multi-modal data correlations, a polynomial hypergraph filter to dynamically extract multi-scale node features through parametric polynomial fitting, and a feature sampling strategy to learn from sparse and labeled samples while avoiding overfitting. Experimental results on four hypergraph datasets and two multi-modal visual datasets demonstrate that the proposed MHNet outperforms state-of-the-art algorithms.

This paper proposes an autoencoder-based method to enhance the information interaction between in-phase/quadrature (I/Q) channels of the input data for automatic modulation recognition (AMR). The proposed method utilizes an autoencoder built by fully-connected layers to correlate the features of I/Q data and obtain the interaction feature from the intermediate layer, which is concatenated together with the original I/Q data as model inputs. To accommodate the new data dimensions, a modification scheme for the existing representative deep learning based AMR (DL-AMR) models is presented. Experimental results show that our method can improve the recognition accuracy of the state-of-the-art baseline models, and has a smaller time overhead compared with complex-valued neural networks.

Learning based on multimodal data has attracted increasing interest recently. While a variety of sensory modalities can be collected for training, not all of them are always available in practical scenarios, which raises the challenge to infer with incomplete modality. This article presents a general framework termed multimodal hallucination (MMH) to bridge the gap between ideal training scenarios and real-world deployment scenarios with incomplete modality data by transferring the complete multimodal knowledge to the hallucination network with incomplete modality input. Compared with the modality hallucination methods that restore privileged modalities information for late fusion, the proposed framework not only helps to preserve the crucial cross-modal cues but relates the study in complete modalities and in incomplete modalities. Then, we introduce two strategies called region-aware distillation and discrepancy-aware distillation to transfer the response-based and joint-representation-based knowledge of pre-trained multimodal networks, respectively. Region-aware distillation establishes and weights knowledge transferring pipelines between the response of multimodal and hallucination networks at multiple regions, which guides the hallucination network to focus on discriminative regions and avoid wasted gradients. Discrepancy-aware distillation guides the hallucination network to mimic the local inter-sample distance of multimodal representations, which enables the hallucination network to acquire the inter-class discrimination refined by multimodal cues. Extensive experiments on multimodal action recognition and face anti-spoofing demonstrate the proposed multimodal hallucination framework can overcome the problem of incomplete modality input in various scenes and achieve state-of-the-art performance. https://github.com/shicaiwei123/TMM-MMH</uri