Videos inherently contain complex temporal dynamics across various spatial directions, often entangled in ways that obscure effective dynamic extraction. Previous studies typically process video spatiotemporal features without disentangling, which hampers their ability to extract dynamic information. Additionally, the extraction of dynamics is disrupted by transient high-dynamic information in video sequences, e.g., noise or flicker, which has received limited attention in the literature. To tackle those problems, this paper proposes the Disentangling and Filtering Dynamics Network (DFDNet). Firstly, to disentangle the interwoven dynamics, DFDNet decomposes the spatially encoded video sequences into lower dimensional sequences. Secondly, a learnable threshold filter is proposed to eliminate the transient high-dynamic information. Thirdly, the model incorporates an MLP to extract the temporal dependencies from the disentangled and filtered sequences. DFDNet demonstrates competitive performance across four chosen datasets, including both low and high-resolution videos. Specifically, on the low-resolution Moving MNIST dataset, DFDNet achieves a 19% improvement on MSE over the previous state-of-the-art model. On the high-resolution SJTU4K dataset, it outperforms the previous state-of-the-art model by 10% on the LPIPS metric under similar inference time.

Video prediction is an important yet challenging task that generates future frames based on previous observations. Despite recent progress, existing methods still suffer from motion blur, due to weak motion perception capabilities leading to uncertainty in motion direction. To address this, we propose a Motion Direction Awareness (MDA) mechanism inspired by the direction-selective mechanism in animal visual systems. Specifically, MDA can decompose complex motions into horizontal and vertical components, allowing dimension reduction and independent processing, thereby effectively enhancing motion perception and reducing uncertainty in predicted motion directions. Based on MDA, we design a multi-scale feature fusion network named MDANet for video prediction, which incorporates different scales of spatially encoded features in conjunction with MDA mechanism to extract the temporal evolution information of global and local spatial features. Extensive experiments on representative datasets demonstrate that MDANet can alleviate motion blurring, improving prediction accuracy and temporal consistency over state-of-the-art models. Furthermore, we validate the generalizability and effectiveness of our MDA mechanism by integrating it into other advanced models. The code is available at supplementary.

Routing optimization within a low Earth orbit (LEO) satellite broadband network (LSBN) has seen advancements through deep reinforcement learning (DRL) approaches in academia. Nonetheless, a crucial aspect often overlooked in these approaches pertains to the inference time of deep neural network (DNN) models during the routing of packets. Our investigation reveals that this oversight can significantly impair routing throughput in LSBN. In response, this paper innovatively proposes a decentralized flow-centric DRL approach, shifting the focus from routing individual packets to entire traffic flows. To align with the large-scale feature of LSBN, we embrace a fully-distributed architecture for flow-centric routing, which is modeled as a partially observable Markov decision process. In this construct, each satellite operates as an independent agent, locally classifying flows following a tailor-designed definition, and is responsible for forwarding a flow to an adjacent satellite based on its internal policy. Notably, the DNN inference is conducted only once on each agent to determine the route for the initial packet of a specific flow; subsequent packets are directed along the same route. Recognizing the potential impact of dynamic LSBN topologies on routing performance, we also introduce an adaptive flow routing update scheme. This scheme is completely free from LSBN environment modelling and aims to bolster the efficacy of the flow-centric approach. Comparative experiments showcase the superiority of the proposed approach over baseline algorithms across various metrics. Consequently, the flow-centric DRL approach can enable high-throughput traffic transmission for LSBN.

Video prediction is a challenging spatiotemporal prediction task that generates future frames based on historical observations. Although recently proposed deep learning-based methods significantly outperform legacy approaches, there still exist gaps between prediction and ground truth, primarily rooted in edge and motion blurring. On the one hand, since conventional performance metrics like Mean Square Error (MSE) and Structure Similarity Index Measure (SSIM) cannot decently evaluate this deficiency, we design a 3D Frequency Loss (3DFL) metric to better assess the similarity of predicted video frames. On the other hand, edge and motion blurring is mainly attributed to the predictive model's insufficient attention to high spatial frequency arising from rapid pixel value variations at object edges, and it is observed that shallow networks are more adept at capturing high spatial frequency information. Therefore, aiming to alleviate edge and motion blurring, we propose a novel video prediction model termed SDFNet that can extract and integrate both spatially encoded shallow and deep-level features. To accommodate SDFNet's multi-branch input structure, a frequency adaptive translator (FATranslator) is derived, which leverages involution operators to adaptively extract inter-frame temporal dependencies from different spatial encoding layers, and further mitigates motion blurring. Extensive experiments demonstrate that our proposed model achieves significant improvements in prediction accuracy and temporal consistency over the current state-of-the-art models on various benchmarks. The results highlight the importance of spatial frequency modeling for enhancing video prediction performance, contributing to the advancement of multimedia technologies.

Legacy computing offloading approaches are originally designed for the terrestrial networks with rather static topologies, and may not be appropriate for the next-generation LEO satellite broadband networks (LSBNs) featured with high dynamicity. This paper presents a multi-agent deep reinforcement learning (MADRL) algorithm for making edge computing multi-level offloading decisions in the LSBNs. Particularly, computing offloading is formulated as a partially observable Markov decision process (POMDP) based multi-agent decision problem. Each LEO satellite is an intelligent agent, either conducting a received edge computing task or forwarding it to its four neighboring satellite or the nearest cloud node on the ground. These agents are fully cooperative and their deep neural network models used to make offloading decisions share the same parameter values and are trained by the same replay buffer. A centralized training and distributed execution framework is utilized to ensure that globally optimized offloading decisions can be achieved based on local observations. Comparative simulation experiments for six representative offloading approaches show that the proposed MADRL aided approach outperforms the others regarding to decreasing edge computing task processing delay and increasing onboard compute resource utilization ratio. In addition, the convergence of this MADRL aided approach is also the best among the three DRL-based approaches.

Network traffic prediction can serve as a proactive approach for network resource planning, allocation, and management. Besides, it can also be applied for load generation in digital twin networks (DTNs). This paper focuses on background traffic prediction of typical local area networks (LANs), which is vital for synchronous traffic generation in DTN. Conventional traffic prediction models are firstly reviewed. The challenges of DTN traffic prediction are analyzed. On that basis, a spatiotemporal-enhanced recurrent neural network (RNN) based approach is elaborated to accurately predict the background traffic matrices of target LANs. Experiments compare this proposed model with four baseline models, including LSTM, CNN-LSTM, ConvLSTM, and PredRNN. The results turn out that the spatiotemporal-enhanced RNN model outperforms the baselines on accuracy. In particular, it can decrease the MSE of PredRNN more than 18%, with acceptable efficiency degradation.

Conventional computation offloading approaches are originally designed for ground networks, and are not effective for low earth orbit (LEO) satellite networks. This paper proposes a multi-agent deep reinforcement learning (MADRL) algorithm for making multi-level offloading decisions in LEO satellite networks. Offloading is formulated as a partially observable Markov decision process based multi-agent decision problem. Each satellite as an agent either conducts a received task, forwards it to neighbors, or sends it to ground clouds based on its own policy. These agents are independent and their deep neural networks to make offloading decisions share identical parameter values and are trained by using the same replay buffer. A centralized training and distributed executing mechanism is adopted to ensure that agents can make globally optimized offloading decisions. Comparative experiments demonstrate that the proposed MADRL algorithm outperforms the five baselines in terms of task processing delay and bandwidth consumption with acceptable computational complexity.