Intelligent reflecting surface (IRS) emerges as a promising technology to enhance wireless communication in recent years. However, the applications of deep learning algorithms within RIS-aided communication systems often suffer performance degradation under extreme conditions owing to a reliance on manual trial-and-error attempts. In this paper, the proposed beamforming neural network architecture search (BNAS) framework automates the design of of neural networks for the joint optimization of precoding vectors and IRS phase shift vectors. To improve robustness and performance, a specialized search space, incorporating two cascading supernets with selectable channel routes, diverse topological connections, and varied operations, is meticulously crafted for beamforming tasks. Meanwhile, the integration of neural tangent kernel theory, supported by alternative optimization guidance and bayesian optimization, not only enhances interpretability but also improves efficiency, thus enabling a more systematic and insightful search process compared to conventional approaches. Extensive numerical simulations confirm the applicability of BNAS, demonstrating superior performance compared to existing deep learning-based methods and traditional algorithms, particularly in challenging scenarios.

This paper investigates an intelligent reflecting surface (IRS)-aided wireless-powered communication network (WPCN) for maximizing the weighted sum rate (WSR). To reduce the complexity of traditional model-driven algorithms and improve convergence in data-driven deep learning approaches, a novel hybrid block coordinate descent (BCD) algorithm motivated by the dilation extraction and context attention (DECA) neural network (NN) is proposed. Specifically, the WSR maximization problem is firstly reformulated as a more tractable form, enabling the BCD algorithm to efficiently optimize the decoupled variables within the constraints. Meanwhile, at each BCD iteration, the DECA NN accelerates IRS phase shift optimization by facilitating the majorization-minimization (MM) algorithm to solve the computationally intensive fractional programming problem. Moreover, by leveraging dilation convolution and high-speed attention mechanisms, the DECA NN significantly outperforms existing deep learning benchmarks in both precision and convergence speed. Numerical results show that the proposed hybrid framework delivers performance comparable to the traditional BCD algorithm with dramatically reduced time consumption, while consistently maintaining robust performance under imperfect CSI and exhibiting strong transferability across diverse communication scenarios.

Accurate channel estimation is an essential prerequisite for achieving significant performance gains in intelligent reflecting surface (IRS)-aided communication systems. Recent studies have shown that deep neural network-based channel estimation holds promise as a competitive alternative to conventional methods. However, existing neural network-based approaches typically involve manual design of network architectures through a trial-and-error process, demanding extensive domain knowledge and human resources. In this paper, we propose an automatic approach to construct a high-performance neural network architecture for channel estimation. Our method, called the channel estimation neural network architecture search (CENAS), utilizes a truncated back-propagation optimization search strategy to explore a neural network tailored for channel estimation. By carefully designing a search space tailored to channel estimation tasks, the automatically constructed network surpasses both conventional and deep learning-based channel estimation algorithms. The convergence of our framework’s network construction process is comprehensively analyzed, providing formal evidence of its convergence properties. Additionally, the proposed framework exhibits good generalization and applicability by allowing flexible adjustment of hyperparameters to generate networks with varying scales. Empirical results show the stability and the improved performance of CENAS framework, validating its effectiveness and desirability.

Accurate channel estimation is an indispensable prerequisite for intelligent reflecting surface (IRS) aided communication systems to achieve huge system performance gains. Current works show that deep neural network-based channel estimation is a promising solution to achieve competitive performance compared with the conventional methods. However, neural network-based approaches generally realize the channel estimation by manually designing network architectures in a trial-and-error manner which need complex neural network domain knowledge and tremendous computation resource. This paper automatically constructs a high-performance neural network architecture to obtain dedicated channel estimation schemes intelligently. Specifically, we propose a channel estimation neural network architecture search (CENAS) method based on gradient alternatively search strategy to search a channel estimation neural network. With the search space designed meticulously for the channel estimation task, the network searched by the proposed method outperforms the conventional and deep learning-based channel estimation algorithms.

Deep learning (DL) is an emerging paradigm for accurate channel estimation for intelligent reflecting surface (IRS)-aided wireless communication systems. It has been proven to be a promising way to achieve better channel estimation performance for the IRS-aided wireless communication system than traditional methods (e.g., least-square algorithm). However, existing DL-based methods rely on ground truth (labels of the true channels) which is difficult to obtain in real networks. In this paper, we propose a self-supervised learning (SSL) method for the IRS channel estimation problem. No ground truth channel is needed in the training, while a simple and novel self-supervised denoising formula without a clean reference signal is presented. Particularly, in the training phase, the self-supervised signal and the input are the received signal vector and its noisy version, respectively. While in the inference phase the input is the estimated channel by using the least-square method and the output is the refined channel estimation. That is, our neural network-based channel estimation algorithm is not reciprocal for training and testing. We demonstrate that the proposed SSL solution has good convergence performance and generalization ability through numerical simulations. Interestingly, we find a “double descent” phenomenon in the learning curve during the test phase, i.e., when we gradually increase the number of training epochs, the performance first gets better, then becomes worse, and further gets better again. Besides, we propose to analyze SSL using the loss landscape and centered kernel alignment method. The results show that the self-supervised model has a similar loss landscape and representational similarity to the supervised model. We explored the effects of different signal-to-noise ratios (SNRs), different neural network sizes, and different training data volumes on our algorithm through numerical simulations. Extensive numerical simulation results show that our SSL algorithm is still competitive without ground truth. We also show that the developed scheme exhibits robustness to SNR ratio mismatch.

Image-level weakly supervised semantic segmentation is a challenging task. As classification networks tend to capture notable object features and are insensitive to over-activation, class activation map (CAM) is too sparse and rough to guide segmentation network training. Inspired by the fact that erasing distinguishing features force networks to collect new ones from non-discriminative object regions, we using relationships between CAMs to propose a novel weakly supervised method. In this work, we apply these features, learned from erased images, as segmentation super-vision, driving network to study robust representation. In specifically, object regions obtained by CAM techniques are erased on images firstly. To provide other regions with seg-mentation supervision, Erased CAM Supervision Net (ECS-Net) generates pixel-level labels by predicting segmentation results of those processed images. We also design the rule of suppressing noise to select reliable labels. Our experiments on PASCAL VOC 2012 dataset show that without data annotations except for ground truth image-level labels, our ECS-Net achieves 67.6% mIoU on test set and 66.6% mIoU on val set, outperforming previous state-of-the-art methods.