Virtual furniture synthesis, a critical task in image composition, aims to seamlessly integrate reference objects into indoor scenes while preserving geometric coherence and visual realism. Despite its significant potential in home design applications, this field remains underexplored due to two major challenges: the absence of publicly available and ready-to-use benchmarks hinders reproducible research, and existing image composition methods fail to meet the stringent fidelity requirements for realistic furniture placement. To address these issues, we introduce RoomBench, a ready-to-use benchmark dataset for virtual furniture synthesis, comprising 7,298 training pairs and 895 testing samples across 27 furniture categories. Then, we propose RoomEditor, a simple yet effective image composition method that employs a parameter-sharing dual U-Net architecture, ensuring better feature consistency by sharing weights between dual branches. Technical analysis reveals that conventional dual-branch architectures generally suffer from inconsistent intermediate features due to independent processing of reference and background images. In contrast, RoomEditor enforces unified feature learning through shared parameters, thereby facilitating model optimization for robust geometric alignment and maintaining visual consistency. Experiments show our RoomEditor is superior to state-of-the-arts, while generalizing directly to diverse objects synthesis in unseen scenes without task-specific fine-tuning. Our dataset and code are available at https://github.com/stonecutter-21/roomeditor.

The vision of 6G mobile networks aims to connect intelligent machines to humans to provide the latter with cooperation, care, and assistance. The mainstream approach for human-to-machine (H2M) semantic communication is to map words into (word) embedding vectors which are clustered according to their semantic similarity to facilitate machines’ interpretation of human languages. The computation-intensive tasks of text-to-embedding mapping are usually delegated to an edge server that senses human commands, maps them into embedding vectors, and then transmits the vectors to a machine over a wireless link. In this work, we propose a quantization framework customized for embedding vectors, called semantic-topology preserving VQ (SemTop-VQ), to overcome the communication bottleneck due to the vectors’ high dimensionality. While traditional VQ focuses on minimizing the distortion of individual vectors, SemTop-VQ aims to minimize the distortion of the topology of embedding matrix, referring to the vectors’ relative positions that represent semantics. To this end, we adopt a topology-distortion metric, termed pointwise-inner-product (PIP) loss, a hierarchical VQ architecture targeting high-dimensional VQ. In this architecture, an embedding vector is decomposed into blocks; the norm and shape (normalized vector) are quantized separately using a scalar and a Grassmannian quantizers, respectively. The main feature of SemTop-VQ lies in deriving from the PIP loss a set of so-called semantic-importance indicators, which reflect the level of influences of individual blocks’ quantization errors on the topology distortion. Then the indicators are applied to optimize quantization-bit allocation for decomposed vector blocks under the criterion of PIP-loss minimization. In practice, the usage probabilities of embedding vectors for a specific machine task are highly skewed and the task is time-varying. We exploit this fact to further develop SemTop-VQ to feature task adaptation that can attain a higher communication efficiency. The task-adaptive VQ is realized via the use of a frequently used (quantization) codebook that is much smaller in size than the original codebook and continuously updated via estimation of embedding-usage distribution. Our experiments using real embedding datasets, namely Word2Vec and Glove, demonstrate the effectiveness of SemTop-VQ as a goal-oriented technique for efficient H2M communications.

Recently, pre-trained vision-language models (e.g., CLIP) have shown great potential in few-shot learning and attracted a lot of research interest. Although efforts have been made to improve few-shot ability of CLIP, key factors on the effectiveness of existing methods have not been well studied, limiting further exploration of CLIP's potential in few-shot learning. In this paper, we first introduce a uni-fied formulation to analyze CLIP-based few-shot learning methods from a perspective of logit bias, which encourages us to learn an effective logit bias for further improving per-formance of CLIP-based few-shot learning methods. To this end, we disassemble three key components involved in computation of logit bias (i.e., logit features, logit predictor, and logit fusion) and empirically analyze the effect on per-formance of few-shot classification. Based on analysis of key components, this paper proposes a novel AMU-Tuning method to learn effective logit bias for CLIP-based few-shot classification. Specifically, our AMU-Tuning predicts logit bias by exploiting the appropriate Auxiliary features, which are fed into an efficient feature-initialized linear clas-sifier with Multi-branch training. Finally, an Uncertainty-based fusion is developed to incorporate logit bias into CLIP for few-shot classification. The experiments are con-ducted on several widely used benchmarks, and the re-sults show AMU-Tuning clearly outperforms its counter-parts while achieving state-of-the-art performance of CLIP-based few-shot learning without bells and whistles.

Recently, efficient fine-tuning of large-scale pre-trained models has attracted increasing research interests, where linear probing (LP) as a fundamental module is involved in exploiting the final representations for task-dependent classification. However, most of the existing methods focus on how to effectively introduce a few of learnable parameters, and little work pays attention to the commonly used LP module. In this paper, we propose a novel Moment Probing (MP) method to further explore the potential of LP. Distinguished from LP which builds a linear classification head based on the mean of final features (e.g., word tokens for ViT) or classification tokens, our MP performs a linear classifier on feature distribution, which provides the stronger representation ability by exploiting richer statistical information inherent in features. Specifically, we represent feature distribution by its characteristic function, which is efficiently approximated by using first- and second-order moments of features. Furthermore, we propose a multi-head convolutional cross-covariance (MHC3) to compute second-order moments in an efficient and effective manner. By considering that MP could affect feature learning, we introduce a partially shared module to learn two recalibrating parameters (PSRP) for backbones based on MP, namely MP+. Extensive experiments on ten benchmarks using various models show that our MP significantly outperforms LP and is competitive with counterparts at lower training cost, while our MP+achieves state-of-the-art performance.

Distributed optimization finds a wide range of applications ranging from machine learning to vehicle platooning. To overcome the bottleneck caused by the required extensive message exchange, we propose in this work the framework of distributed over-the-air computing (AirComp) to realize a one-step aggregation for distributed optimization. Equivalently, the technique superimposes multiple instances of conventional AirComp processes, giving rise to the challenge of jointly designing multicast beamforming at devices to rein in errors due to interference and channel distortion. We consider two design criteria. One is to minimize the sum AirComp error (i.e., sum mean-squared error (MSE)) with respect to the desired average-functional values. An efficient solution approach is proposed by transforming the non-convex beamforming problem into an equivalent concave-convex fractional program and solving it by nesting convex programming into a bisection search. The other one, called zero-forcing (ZF) multicast beamforming, is to force the received over-the-air aggregated signals at devices to be equal to the desired functional values, where the optimal beamforming admits closed form. Last, the convergence of a classic distributed optimization algorithm is analyzed. The distributed AirComp is found experimentally to accelerate convergence by dramatically reducing communication latency.

The deployment of federated learning in a wireless network, calledfederated edge learning(FEEL), exploits low-latency access to distributed mobile data to efficiently train an AI model while preserving data privacy. In this work, we study the spatial (i.e., spatially averaged) learning performance of FEEL deployed in a large-scale cellular network with spatially random distributed devices. Both the schemes of digital and analog transmission are considered, providing support of error-free uploading and over-the-air aggregation of local model updates by devices. The derived spatial convergence rate for digital transmission is found to be constrained by a limited number of active devices regardless of device density and converges to the ground-true rate exponentially fast as the number grows. The population of active devices depends on network parameters such as processing gain and signal-to-interference threshold for decoding. On the other hand, the limit does not exist for uncoded analog transmission. In this case, the spatial convergence rate is slowed down due to the direct exposure of signals to the perturbation of inter-cell interference. Nevertheless, the effect diminishes when devices are dense as interference is averaged out by aggressive over-the-air aggregation. In terms of learning latency (in second), analog transmission is preferred to the digital scheme as the former dramatically reduces multi-access latency by enabling simultaneous access.

Self-Organizing Networks (SON) is an automation technology making the planning, deployment, operation, optimization, and healing of networks simpler and faster. Legacy SON is targeted at network automation and network optimization through certain optimization rules and policies which are globally applied in networks. However, scalable and targeted optimization is not considered yet in 3GPP. Furthermore, SON is driven by performance optimization rather than ultimately improving user Quality of Experience (QoE). The impact of application characteristics on network performance and further on QoE are also not considered in 3GPP SON. This paper presents an application characteristics-driven SON system (APP-SON) to optimize 4G/5G network performance and user Quality of Experience. APP-SON leverages a scalable big data platform for targeted optimization through profiling cell application characteristics in an incremental manner in temporal space. A Hungarian Algorithm Assisted Clustering (HAAC) algorithm and a deep learning-assisted regression algorithm are developed to profile the cell application characteristics and find the targeted KPIs to be optimized for each cell. A similarity-based, parametertuning algorithm is developed to tune the corresponding engineering parameters to optimize the targeted KPIs which further improve QoE. Experimental results demonstrate that the APP-SON system can precisely profile cell traffic and application characteristics to find the targeted KPIs for optimization for each cell. APP-SON can also automatically tune the corresponding engineering parameters to improve the corresponding KPIs, ultimately improving QoE. APP-SON has been successfully implemented in production and applied in a tier-1 operator's 4G network and as a universal SON solution it will be smoothly transitioned and applied in 5G networks for this operator.

The well-established and extensively validated atmospheric radiative transfer model MODTRAN®1is widely used by the remote sensing community to define the mapping from surface spectral reflectance to solar scatter radiance as a function of atmosphere definition. This mapping is subsequently used in atmospheric correction/compensation models to define the reverse mapping from multi- and hyperspectral images to ground surface spectral reflectances. Since MODTRAN employs a stratified spherical shell atmosphere and accounts for spherical refraction, its applicability extends to low sun and highly off-nadir scenarios. The discrete ordinate module DISORT solves the RTE to compute multiple scattering solar path radiances for MODTRAN. However, there is a problem: DISORT is a plane-parallel atmosphere model. In this paper, a method for modifying DISORT to compute the curved-path segment radiance contributions required by MODTRAN is described.