The recent advancements in the research area primarily focus on optimizing and enhancing the efficiency of deep learning models across various applications. A significant trend is the integration of semantic constraints and knowledge transfer mechanisms to improve model performance, particularly in tasks like stereo matching and semantic segmentation. Innovations in pruning techniques, such as semi-structured and spatial-aware methods, are being developed to create more efficient models without compromising accuracy, especially for resource-constrained environments like UAVs and edge devices. Additionally, there is a growing emphasis on adversarial robustness and task-adaptive communication strategies, which aim to secure and optimize neural networks for real-world applications. Notably, the introduction of novel architectures and pruning strategies that leverage semantic information and adaptive mechanisms are pushing the boundaries of what is achievable in terms of both performance and efficiency. For instance, the development of semantic-constrained stereo matching networks and adaptive structural pruning methods for remote sensing image classification are particularly noteworthy for their innovative approaches and state-of-the-art results.