An Empirical Study of Attention Networks for Semantic Segmentation

Semantic segmentation is a vital problem in computer vision. Recently, a common solution to semantic segmentation is the end-to-end convolution neural network, which is much more accurate than traditional methods.Recently, the decoders based on attention achieve state-of-the-art (SOTA) performance on various datasets. But these networks always are compared with the mIoU of previous SOTA networks to prove their superiority and ignore their characteristics without considering the computation complexity and precision in various categories, which is essential for engineering applications. Besides, the methods to analyze the FLOPs and memory are not consistent between different networks, which makes the comparison hard to be utilized. What's more, various methods utilize attention in semantic segmentation, but the conclusion of these methods is lacking. This paper first conducts experiments to analyze their computation complexity and compare their performance. Then it summarizes suitable scenes for these networks and concludes key points that should be concerned when constructing an attention network. Last it points out some future directions of the attention network.

Over-the-Air Multi-View Pooling for Distributed Sensing

Sensing is envisioned as a key network function of the 6G mobile networks. Artificial intelligence (AI)-empowered sensing fuses features of multiple sensing views from devices distributed in edge networks for the edge server to perform accurate inference. This process, known as multi-view pooling, creates a communication bottleneck due to multi-access by many devices. To alleviate this issue, we propose a task-oriented simultaneous access scheme for distributed sensing called Over-the-Air Pooling (AirPooling). The existing Over-the-Air Computing (AirComp) technique can be directly applied to enable Average-AirPooling by exploiting the waveform superposition property of a multi-access channel. However, despite being most popular in practice, the over-the-air maximization, called Max-AirPooling, is not AirComp realizable as AirComp addresses a limited subset of functions. We tackle the challenge by proposing the novel generalized AirPooling framework that can be configured to support both Max- and Average-AirPooling by controlling a configuration parameter. The former is realized by adding to AirComp the designed pre-processing at devices and post-processing at the server. To characterize the end-to-end sensing performance, the theory of classification margin is applied to relate the classification accuracy and the AirPooling error. Furthermore, the analysis reveals an inherent tradeoff of Max-AirPooling between the accuracy of the pooling-function approximation and the effectiveness of noise suppression. Using the tradeoff, we optimize the configuration parameter of Max-AirPooling, yielding a sub-optimal closed-form method of adaptive parametric control. Experimental results obtained on real-world datasets show that AirPooling provides sensing accuracies close to those achievable by the traditional digital air interface but dramatically reduces the communication latency.

Semantic Data Sourcing for 6G Edge Intelligence

As a new function of 6G networks, edge intelligence refers to the ubiquitous deployment of machine learning and artificial intelligence (AI) algorithms at the network edge to empower many emerging applications ranging from sensing to auto-pilot. To support relevant use cases, including sensing, edge learning, and edge inference, all require transmission of high-dimensional data or AI models over the air. To overcome the bottleneck, we propose a novel framework of SEMantic DAta Sourcing (SEMDAS) for locating semantically matched data sources to efficiently enable edge-intelligence operations. The comprehensive framework comprises new architecture, protocol, semantic matching techniques, and design principles for task-oriented wireless techniques. As the key component of SEMDAS, we discuss a set of machine learning based semantic matching techniques targeting different edge-intelligence use cases. Moreover, for designing task-oriented wireless techniques, we discuss different tradeoffs in SEMDAS systems, propose the new concept of joint semantics-and-channel matching, and point to a number of research opportunities. The SEMDAS framework not only overcomes the said communication bottleneck but also addresses other networking issues including long-distance transmission, sparse connectivity, high-speed mobility, link disruptions, and security. In addition, experimental results using a real dataset are presented to demonstrate the performance gain of SEMDAS.

In-situ Model Downloading to Realize Versatile Edge AI in 6G Mobile Networks

The sixth-generation (6G) mobile networks are expected to feature the ubiquitous deployment of machine learning and AI algorithms at the network edge. With rapid advancements in edge AI, the time has come to realize intelligence downloading onto edge devices (e.g., smartphones and sensors). To materialize this version, we propose a novel technology in this article, called in-situ model downloading, that aims to achieve transparent and real-time replacement of on-device AI models by downloading from an AI library in the network. Its distinctive feature is the adaptation of downloading to time-varying situations (e.g., application, location, and time), devices' heterogeneous storage-and-computing capacities, and channel states. A key component of the presented framework is a set of techniques that dynamically compress a downloaded model at the depth-level, parameter-level, or bit-level to support adaptive model downloading. We further propose a virtualized 6G network architecture customized for deploying in-situ model downloading with the key feature of a three-tier (edge, local, and central) AI library. Furthermore, experiments are conducted to quantify 6G connectivity requirements and research opportunities pertaining to the proposed technology are discussed.

Resource Allocation for Multiuser Edge Inference with Batching and Early Exiting (Extended Version)

The deployment of inference services at the network edge, called edge inference, offloads computation-intensive inference tasks from mobile devices to edge servers, thereby enhancing the former's capabilities and battery lives. In a multiuser system, the joint allocation of communication-and-computation ($\text{C}^\text{2}$) resources (i.e., scheduling and bandwidth allocation) is made challenging by adopting efficient inference techniques, batching and early exiting, and further complicated by the heterogeneity in users' requirements on accuracy and latency. Batching groups multiple tasks into one batch for parallel processing to reduce time-consuming memory access and thereby boosts the throughput (i.e., completed task per second). On the other hand, early exiting allows a task to exit from a deep-neural network without traversing the whole network to support a tradeoff between accuracy and latency. In this work, we study optimal $\text{C}^\text{2}$ resource allocation with batching and early exiting, which is an NP-complete integer program. A set of efficient algorithms are designed under the criterion of maximum throughput by tackling the challenge. Experimental results demonstrate that both optimal and sub-optimal $\text{C}^\text{2}$ resource allocation algorithms can leverage integrated batching and early exiting to achieve 200% throughput gain over conventional schemes.

Deep Unsupervised Learning for Joint Antenna Selection and Hybrid Beamforming

In this paper, we consider a massive multiple-input-multiple-output (MIMO) downlink system that improves the hardware efficiency by dynamically selecting the antenna subarray and utilizing 1-bit phase shifters for hybrid beamforming. To maximize the spectral efficiency, we propose a novel deep unsupervised learning-based approach that avoids the computationally prohibitive process of acquiring training labels. The proposed design has its input as the channel matrix and consists of two convolutional neural networks (CNNs). To enable unsupervised training, the problem constraints are embedded in the neural networks: the first CNN adopts deep probabilistic sampling, while the second CNN features a quantization layer designed for 1-bit phase shifters. The two networks can be trained jointly without labels by sharing an unsupervised loss function. We next propose a phased training approach to promote the convergence of the proposed networks. Simulation results demonstrate the advantage of the proposed approach over conventional optimization-based algorithms in terms of both achieved rate and computational complexity.