Given just a few glimpses of a scene, can you imagine the movie playing out as the camera glides through it? That’s the lens we take on sparse-input novel view synthesis, not only as filling spatial gaps between widely spaced views, but also as completing a natural video unfolding through space. We recast the task as test-time natural video completion, using powerful priors from pretrained video diffusion models to hallucinate plausible in-between views. Our zero-shot, generation-guided framework produces pseudo views at novel camera poses, modulated by an uncertainty-aware mechanism for spatial coherence. These synthesized frames densify supervision for 3D Gaussian Splatting (3D-GS) for scene reconstruction, especially in under-observed regions. An iterative feedback loop lets 3D geometry and 2D view synthesis inform each other, improving both the scene reconstruction and the generated views. The result is coherent, high-fidelity renderings from sparse inputs without any scene-specific training or fine-tuning. On LLFF, DTU, DL3DV, and MipNeRF-360, our method significantly outperforms strong 3D-GS baselines under extreme sparsity. Our project page is at https://decayale.github.io/project/SV2CGS.

Applying self-supervised deep learning improves the processing speed and accuracy of automatic modulation recognition (AMR). It reduces the dependence of previous deep networks on many labeled samples. However, affected by an incomplete signal representation modes set, previous models do not fully utilize the multiview property of signals in self-supervised learning. To deal with this issue, a hybrid-view contrastive model for AMR is proposed in this article based on self-supervised learning framework. First, star video is proposed to complete the set of signal representation modes. Next, a self-supervised learning framework based on hybrid-view contrastive learning, hybrid-view self-supervised framework (HVSF), is established to fully extract the signal features, where signals are augmented across views, including the discrete sequence, image, and video format. Considering the view-exclusive information loss and the model complexity, a weakly contrastive strategy and a Transformer-based view-shared feature extractor are finally constructed. Evaluation on four standard datasets demonstrates that the proposed model, HVSF, outperforms both the self-supervised models and supervised models, affirming its superior performance and stability.

Fast charging has attracted increasing attention from the battery community for electrical vehicles (EVs) to alleviate range anxiety and reduce charging time for EVs. However, inappropriate charging strategies would cause severe degradation of batteries or even hazardous accidents. To optimize fast-charging strategies under various constraints, particularly safety limits, we propose a novel deep Bayesian optimization (BO) approach that utilizes Bayesian recurrent neural network (BRNN) as the surrogate model, given its capability in handling sequential data and providing uncertainty quantifications for the output. In addition, a combined acquisition function of expected improvement (EI) and upper confidence bound (UCB) is developed to better balance the exploitation and exploration. The effectiveness of the proposed approach is demonstrated on the PETLION, a porous electrode theory-based battery simulator. Our method is also compared with the state-of-the-art BO methods that use Gaussian process (GP) and non-recurrent network as surrogate models. The results verify the superior performance of the proposed fast charging approaches, which mainly results from that: 1) the BRNN-based surrogate model provides a more precise prediction of battery lifetime than that based on GP or non-recurrent network; and 2) the combined acquisition function outperforms traditional EI or UCB criteria in exploring the optimal charging protocol that maintains the longest battery lifetime. Note to Practitioners—This study is motivated by the need to develop fast-charging strategies for batteries to reduce the charging time while maintaining the safety and longevity of batteries. Traditional methods to optimize battery fast-charging protocols require either solving complex battery models or conducting tremendous repetitive and costly cycling experiments. In addition, the majority of works in this direction mainly focus on minimizing the charging time without considering the battery degradation caused by the designed charging protocol. To address these issues, we present a data-driven optimization technique to rapidly discover the optimal fast-charging strategies that not only decrease the charging time but also slow down the battery degradation. Our method is efficient and can find the optimal solution within tens of iterations (or repetitive experiments), thus greatly reducing the cost and efforts in optimizing battery charging protocols.

By integrating the software-defined networking (SDN) architecture with the machine learning-based knowledge plane, knowledge-defined networking (KDN) is revolutionizing established traffic engineering (TE) methodologies. This paper investigates the challenging joint routing and caching problem in KDN-based networks, managing multiple traffic flows to improve long-term quality-of-service (QoS) performance. This challenge is formulated as a computationally expensive non-convex mixed-integer non-linear programming (MINLP) problem, which exceeds the capacity of heuristic methods to achieve near-optimal solutions. To address this issue, we present DRL-JRC, an actor-critic deep reinforcement learning (DRL) algorithm for adaptive joint routing and caching in KDN-based networks. DRL-JRC orchestrates the optimization of multiple QoS metrics, including end-to-end delay, packet loss rate, load balancing index, and hop count. During offline training, DRL-JRC employs proximal policy optimization (PPO) to smooth the policy optimization process. In addition, the learned policy can be seamlessly integrated with conventional caching solutions during online execution. Extensive experiments demonstrate the comprehensive superiority of DRL-JRC over baseline methods in various scenarios. Meanwhile, DRL-JRC consistently outperforms the heuristic baseline under partial policy deployment during execution. Compared to the average performance of the baseline methods, DRL-JRC reduces the end-to-end delay by 51.14% and the packet loss rate by 40.78%.

This paper presents an empirical study on deep reinforcement learning (DRL) based probabilistic cognitive routing using the OMNeT++ framework and programming protocol-independent packet processors (P4). The proposed algorithm combines the power of DRL and cognitive routing to achieve efficient and adaptive probabilistic routing in software-defined networking (SDN) environments. To facilitate the research, we develop a dedicated network simulation environment using the OMNeT++ framework and a self-developed SDN platform based on P4. The empirical study highlights the importance of a comprehensive training and validation process in both simulation and real-world SDN environments. Through closed-loop training, the cognitive routing framework provides real-time feedback from the actual network environment to the simulation environment, allowing the agent to excel in real-world network environments. Meanwhile, the results demonstrate that solely testing the algorithm in either environment is inadequate for evaluating its performance accurately.

Large-scale capture of human motion with diverse, complex scenes, while immensely useful, is often considered prohibitively costly. Meanwhile, human motion alone contains rich information about the scene they reside in and interact with. For example, a sitting human suggests the existence of a chair, and their leg position further implies the chair’s pose. In this paper, we propose to synthesize diverse, semantically reasonable, and physically plausible scenes based on human motion. Our framework, Scene Synthesis from HUMan MotiON (SUMMON), includes two steps. It first uses ContactFormer, our newly introduced contact predictor, to obtain temporally consistent contact labels from human motion. Based on these predictions, SUMMON then chooses interacting objects and optimizes physical plausibility losses; it further populates the scene with objects that do not interact with humans. Experimental results demonstrate that SUMMON synthesizes feasible, plausible, and diverse scenes and has the potential to generate extensive human-scene interaction data for the community.

Plasma cell-free DNA (cfDNA) has certain fragmentation patterns, which can bring non-random base content curves of the sequencing data's beginning cycles. We studied the patterns and found that we could determine whether a sample is cfDNA or not by just looking into the first 10 cycles of its base content curves. We analysed 3189 FastQ files, including 1442 cfDNA, 1234 genomic DNA, 507 FFPE tumour DNA and 6 urinary cfDNA. By deep analysing these data, we find the patterns are stable enough to distinguish cfDNA from other kinds of DNA samples. Based on this finding, we build classification models to recognise cfDNA samples by their sequencing data. Pattern recognition models are then trained with different classification algorithms like k-nearest neighbours (KNN), random forest and support vector machine (SVM). The result of 1000 iteration .632+ bootstrapping shows that all these classifiers can give an average accuracy higher than 98%, indicating that the cfDNA patterns are unique and can make the dataset highly separable. The best result is obtained using random forest classifier with a 99.89% average accuracy (σ = 0.00068). A tool called CfdnaPattern (http://github.com/OpenGene/CfdnaPattern) has been developed to train the model and to predict whether a sample is cfDNA or not.