Recent developments in RDMA networks are leading to the trend of memory disaggregation. However, the performance of each compute node is still limited by the network, especially when it needs to perform a large number of concurrent fine-grained remote accesses. According to our evaluations, existing IOPS-bound disaggregated applications do not scale well beyond 32 cores, and therefore do not take full advantage of today's many-core machines.

This work focuses on the sample collection in reinforcement learning (RL), where the interaction with the environment is typically time-consuming and extravagantly expensive. In order to collect samples in a more valuable way, we propose a confidence-based sampling strategy based on the optimal computing budget allocation algorithm (OCBA), which actively allocates the computing efforts to actions with different predictive uncertainties. We estimate the uncertainty with ensembles and generalize them from tabular representations to function approximations. The OCBA-based sampling strategy could be easily integrated into various off-policy RL algorithms, where we take Q-learning, DQN, and SAC as examples to show the incorporation. Besides, we provide the theoretical analysis towards convergence and evaluate the algorithms experimentally. According to the experiments, the incorporated algorithms obtain remarkable gains compared with modern ensemble-based RL algorithms.Note to Practitioners—Reinforcement learning is a powerful tool for handling sequential decision-making problems, e.g., autonomous driving and robotics control, where the behaviors typically have a long-term effect on future events. However, although RL achieves human-level control in some tasks, it severely suffers from low sample efficiency. Therefore, implementing RL in some practical areas, e.g., healthcare and rescue, is extremely hard due to the requirement of massive samples. This work aims to enhance the exploration of RL by incorporating OCBA, which provides an asymptotically optimal data-collection strategy for simulation-based optimization. Based on ensemble-based uncertainty estimation and OCBA-based action selection, the incorporated RL algorithms show competitive performance on many benchmarks and significantly reduce the sampling efforts during iterations.

3D visual grounding aims to locate the referred target object in 3D point cloud scenes according to a free-form language description. Previous methods mostly follow a two-stage paradigm, i.e., language-irrelevant detection and cross-modal matching, which is limited by the isolated architecture. In such a paradigm, the detector needs to sample keypoints from raw point clouds due to the inherent properties of 3D point clouds (irregular and large-scale), to generate the corresponding object proposal for each keypoint. However, sparse proposals may leave out the target in detection, while dense proposals may confuse the matching model. Moreover, the language-irrelevant detection stage can only sample a small proportion of keypoints on the target, deteriorating the target prediction. In this paper, we propose a 3D Single-Stage Referred Point Progressive Selection (3D-SPS) method, which progressively selects keypoints with the guidance of language and directly locates the target. Specifically, we propose a Description-aware Keypoint Sampling (DKS) module to coarsely focus on the points of language-relevant objects, which are significant clues for grounding. Besides, we devise a Target-oriented Progressive Mining (TPM) module to finely concentrate on the points of the target, which is enabled by progressive intra-modal relation modeling and inter-modal target mining. 3D-SPS bridges the gap between detection and matching in the 3D visual grounding task, localizing the target at a single stage. Experiments demonstrate that 3D-SPS achieves state-of-the-art performance on both ScanRe-fer and Nr3D/Sr3D datasets.

Vision-language navigation is the task of directing an embodied agent to navigate in 3D scenes with natural language instructions. For the agent, inferring the long-term navigation target from visual-linguistic clues is crucial for reliable path planning, which, however, has rarely been studied before in literature. In this article, we propose a Target-Driven Structured Transformer Planner (TD-STP) for long-horizon goal-guided and room layout-aware navigation. Specifically, we devise an Imaginary Scene Tokenization mechanism for explicit estimation of the long-term target (even located in unexplored environments). In addition, we design a Structured Transformer Planner which elegantly incorporates the explored room layout into a neural attention architecture for structured and global planning. Experimental results demonstrate that our TD-STP substantially improves previous best methods' success rate by 2% and 5% on the test set of R2R and REVERIE benchmarks, respectively. Our code is available at https://github.com/YushengZhao/TD-STP.

Trajectory prediction of multiple agents in a crowded scene is an essential component in many applications, including intelligent monitoring, autonomous robotics, and self-driving cars. Accurate agent trajectory prediction remains a significant challenge because of the complex dynamic interactions among the agents and between them and the surrounding scene. To address the challenge, we propose a decoupled attention-based spatial-temporal modeling strategy in the proposed trajectory prediction method. The past and current interactions among agents are dynamically and adaptively summarized by two separate attention-based networks and have proven powerful in improving the prediction accuracy. Moreover, it is optional in the proposed method to make use of the road map and the plan of the ego-agent for scene-compliant and accurate predictions. The road map feature is efficiently extracted by a convolutional neural network, and the features of the ego-agent’s plan is extracted by a gated recurrent network with an attention module based on the temporal characteristic. Experiments on benchmark trajectory prediction datasets demonstrate that the proposed method is effective when the ego-agent plan and the the surrounding scene information are provided and achieves state-of-the-art performance with only the observed trajectories.

Video instance segmentation (VIS) is the task that requires simultaneously classifying, segmenting and tracking object instances of interest in video. Recent methods typically develop sophisticated pipelines to tackle this task. Here, we propose a new video instance segmentation framework built upon Transformers, termed VisTR, which views the VIS task as a direct end-to-end parallel sequence decoding/prediction problem. Given a video clip consisting of multiple image frames as input, VisTR outputs the sequence of masks for each instance in the video in order directly. At the core is a new, effective instance sequence matching and segmentation strategy, which supervises and segments instances at the sequence level as a whole. VisTR frames the instance segmentation and tracking in the same perspective of similarity learning, thus considerably simplifying the overall pipeline and is significantly different from existing approaches.Without bells and whistles, VisTR achieves the highest speed among all existing VIS models, and achieves the best result among methods using single model on the YouTube-VIS dataset. For the first time, we demonstrate a much simpler and faster video instance segmentation framework built upon Transformers, achieving competitive accuracy. We hope that VisTR can motivate future research for more video understanding tasks.Code is available at: https://git.io/VisTR

Trajectory prediction is a fundamental task in many real applications such as autonomous robotics and video surveillance. In this paper, we propose a novel vision-based trajectory prediction method which is able to extract the interactive features by active global interaction-aware drivable space learning. The learned global interaction-aware drivable spaces denote the areas with low occupation probabilities, which provide the regions and directions that the agents can move into. Specifically, our method describes a sequence of motion states, i.e. the location, the velocity and the acceleration, as occupancy grid maps, and then use them to train the deep learning model in the supervised manner. Moreover, an interactive loss for training the inference net of drivable spaces and trajectory prediction net simultaneously is introduced. Comparisons with state-of-the-art methods on benchmark datasets demonstrate the effectiveness of the proposed method.

Motion prediction of multiple agents in a dynamic scene is a crucial component in many real applications, including intelligent monitoring and autonomous driving. Due to the complex interactions among the agents and their interactions with the surrounding scene, accurate trajectory prediction is still a great challenge. In this paper, we propose a new method for robust trajectory prediction of multiple intelligent agents in a dynamic scene. The input of the method includes the observed trajectories of all agents, and optionally, the planning of the ego-agent and the surrounding high definition map at every time steps. Given observed trajectories, an efficient approach in a star computational topology is utilized to compute both the spatiotemporal interaction features and the current interaction features between the agents, where the time complexity scales linearly to the number of agents. Moreover, on an autonomous vehicle, the proposed prediction method can make use of the planning of ego-agent to improve the modeling of the interaction between surrounding agents. To increase the robustness to upstream perception noises, at the training stage, we randomly mask out the input data, a.k.a. the points on the observed trajectories of agents and the lane sequence. Experiments on autonomous driving and pedestrian-walking datasets demonstrate that the proposed method is not only effective when the planning of ego-agent and the high definition map are provided, but also achieves state-of-the-art performance with only the observed trajectories.

Very recently, a variety of vision transformer architectures for dense prediction tasks have been proposed and they show that the design of spatial attention is critical to their success in these tasks. In this work, we revisit the design of the spatial attention and demonstrate that a carefully devised yet simple spatial attention mechanism performs favorably against the state-of-the-art schemes. As a result, we propose two vision transformer architectures, namely, Twins- PCPVT and Twins-SVT. Our proposed architectures are highly efficient and easy to implement, only involving matrix multiplications that are highly optimized in modern deep learning frameworks. More importantly, the proposed architectures achieve excellent performance on a wide range of visual tasks including image-level classification as well as dense detection and segmentation. The simplicity and strong performance suggest that our proposed architectures may serve as stronger backbones for many vision tasks.

Pedestrian trajectory prediction is crucial for many important applications. This problem is a great challenge because of complicated interactions among pedestrians. Previous methods model only the pairwise interactions between pedestrians, which not only oversimplifies the interactions among pedestrians but also is computationally inefficient. In this paper, we propose a novel model StarNet to deal with these issues. StarNet has a star topology which includes a unique hub network and multiple host networks. The hub network takes observed trajectories of all pedestrians to produce a comprehensive description of the interpersonal interactions. Then the host networks, each of which corresponds to one pedestrian, consult the description and predict future trajectories. The star topology gives StarNet two advantages over conventional models. First, StarNet is able to consider the collective influence among all pedestrians in the hub network, making more accurate predictions. Second, StarNet is computationally efficient since the number of host network is linear to the number of pedestrians. Experiments on multiple public datasets demonstrate that StarNet outperforms multiple state-of-the-arts by a large margin in terms of both accuracy and efficiency.

Emerging large-scale scientific applications require to access large data objects in high and robust performance. We propose RobuSTore, a storage architecture that combines erasure codes and speculative access mechanisms for parallel write and read in distributed environments. The mechanisms can effectively aggregate the bandwidth from a large number of distributed disks and statistically tolerate pear-disk performance variation. Our simulation results affirm the high and robust performance of RobuSTore in both write and read operations compared to traditional parallel storage systems. For example, for a 1GB data access using 64 disks, RobuSTore achieves average bandwidth of 186MBps for write and 400MBps for read, nearly 6x and 15x that achieved by a RAID-0 system. The standard deviation of access latency is only 0.5 second, about 9% of the write latency and 20% of the read latency, and a 5-fold improvement from RAID-0. The improvements are achieved at moderate cost: about 40% increase in I/O operations and 2x-3x increase in storage capacity utilization.