Learning to play table tennis is a challenging task for robots, due to the variety of the strokes required. Current advances in deep Reinforcement Learning (RL) have shown potential in learning the optimal strokes. However, the large amount of explor
ation still limits the applicability when utilizing RL in real scenarios. In this paper, we first propose a realistic simulation environment where several models are built for the balls dynamics and the robots kinematics. Instead of training an end-to-end RL model, we decompose it into two stages: the balls hitting state prediction and consequently learning the racket strokes from it. A novel policy gradient approach with TD3 backbone is proposed for the second stage. In the experiments, we show that the proposed approach significantly outperforms the existing RL methods in simulation. To cross the domain from simulation to reality, we develop an efficient retraining method and test in three real scenarios with a success rate of 98%.
We establish in this paper sharp lower bounds for the $2k$-th moment of the derivative of the Riemann zeta function on the critical line for all real $k geq 0$.
TODO comments are very widely used by software developers to describe their pending tasks during software development. However, after performing the task developers sometimes neglect or simply forget to remove the TODO comment, resulting in obsolete
TODO comments. These obsolete TODO comments can confuse development teams and may cause the introduction of bugs in the future, decreasing the softwares quality and maintainability. In this work, we propose a novel model, named TDCleaner (TODO comment Cleaner), to identify obsolete TODO comments in software projects. TDCleaner can assist developers in just-in-time checking of TODO comments status and avoid leaving obsolete TODO comments. Our approach has two main stages: offline learning and online prediction. During offline learning, we first automatically establish <code_change, todo_comment, commit_msg> training samples and leverage three neural encoders to capture the semantic features of TODO comment, code change and commit message respectively. TDCleaner then automatically learns the correlations and interactions between different encoders to estimate the final status of the TODO comment. For online prediction, we check a TODO comments status by leveraging the offline trained model to judge the TODO comments likelihood of being obsolete. We built our dataset by collecting TODO comments from the top-10,000 Python and Java Github repositories and evaluated TDCleaner on them. Extensive experimental results show the promising performance of our model over a set of benchmarks. We also performed an in-the-wild evaluation with real-world software projects, we reported 18 obsolete TODO comments identified by TDCleaner to Github developers and 9 of them have already been confirmed and removed by the developers, demonstrating the practical usage of our approach.
The recently proposed Detection Transformer (DETR) model successfully applies Transformer to objects detection and achieves comparable performance with two-stage object detection frameworks, such as Faster-RCNN. However, DETR suffers from its slow co
nvergence. Training DETR from scratch needs 500 epochs to achieve a high accuracy. To accelerate its convergence, we propose a simple yet effective scheme for improving the DETR framework, namely Spatially Modulated Co-Attention (SMCA) mechanism. The core idea of SMCA is to conduct location-aware co-attention in DETR by constraining co-attention responses to be high near initially estimated bounding box locations. Our proposed SMCA increases DETRs convergence speed by replacing the original co-attention mechanism in the decoder while keeping other operations in DETR unchanged. Furthermore, by integrating multi-head and scale-selection attention designs into SMCA, our fully-fledged SMCA can achieve better performance compared to DETR with a dilated convolution-based backbone (45.6 mAP at 108 epochs vs. 43.3 mAP at 500 epochs). We perform extensive ablation studies on COCO dataset to validate SMCA. Code is released at https://github.com/gaopengcuhk/SMCA-DETR .
Multi-hop machine reading comprehension is a challenging task in natural language processing, which requires more reasoning ability and explainability. Spectral models based on graph convolutional networks grant the inferring abilities and lead to co
mpetitive results, however, part of them still face the challenge of analyzing the reasoning in a human-understandable way. Inspired by the concept of the Grandmother Cells in cognitive neuroscience, a spatial graph attention framework named crname, imitating the procedure was proposed. This model is designed to assemble the semantic features in multi-angle representations and automatically concentrate or alleviate the information for reasoning. The name crname is a metaphor for the pattern of the model: regard the subjects of queries as the start points of clues, take the reasoning entities as bridge points, and consider the latent candidate entities as the grandmother cells, and the clues end up in candidate entities. The proposed model allows us to visualize the reasoning graph and analyze the importance of edges connecting two entities and the selectivity in the mention and candidate nodes, which can be easier to be comprehended empirically. The official evaluations in open-domain multi-hop reading dataset WikiHop and Drug-drug Interactions dataset MedHop prove the validity of our approach and show the probability of the application of the model in the molecular biology domain.
We study the $2k$-th discrete moment of the derivative of the Riemann zeta-function at nontrivial zeros to establish sharp lower bounds for all real $k geq 0$ under the Riemann hypothesis (RH).
Convolutional neural networks (CNNs) are ubiquitous in computer vision, with a myriad of effective and efficient variations. Recently, Transformers -- originally introduced in natural language processing -- have been increasingly adopted in computer
vision. While early adopters continue to employ CNN backbones, the latest networks are end-to-end CNN-free Transformer solutions. A recent surprising finding shows that a simple MLP based solution without any traditional convolutional or Transformer components can produce effective visual representations. While CNNs, Transformers and MLP-Mixers may be considered as completely disparate architectures, we provide a unified view showing that they are in fact special cases of a more general method to aggregate spatial context in a neural network stack. We present the model (CONText AggregatIon NEtwoRk), a general-purpose building block for multi-head context aggregation that can exploit long-range interactions emph{a la} Transformers while still exploiting the inductive bias of the local convolution operation leading to faster convergence speeds, often seen in CNNs. In contrast to Transformer-based methods that do not scale well to downstream tasks that rely on larger input image resolutions, our efficient network, named modellight, can be employed in object detection and instance segmentation networks such as DETR, RetinaNet and Mask-RCNN to obtain an impressive detection mAP of 38.9, 43.8, 45.1 and mask mAP of 41.3, providing large improvements of 6.6, 7.3, 6.9 and 6.6 pts respectively, compared to a ResNet-50 backbone with a comparable compute and parameter size. Our method also achieves promising results on self-supervised learning compared to DeiT on the DINO framework.
With the proliferation of edge smart devices and the Internet of Vehicles (IoV) technologies, intelligent fatigue detection has become one of the most-used methods in our daily driving. To improve the performance of the detection model, a series of t
echniques have been developed. However, existing work still leaves much to be desired, such as privacy disclosure and communication cost. To address these issues, we propose FedSup, a client-edge-cloud framework for privacy and efficient fatigue detection. Inspired by the federated learning technique, FedSup intelligently utilizes the collaboration between client, edge, and cloud server to realizing dynamic model optimization while protecting edge data privacy. Moreover, to reduce the unnecessary system communication overhead, we further propose a Bayesian convolutional neural network (BCNN) approximation strategy on the clients and an uncertainty weighted aggregation algorithm on the cloud to enhance the central model training efficiency. Extensive experiments demonstrate that the FedSup framework is suitable for IoV scenarios and outperforms other mainstream methods.
We numerically study the breathing dynamics induced by collision between bright solitons in the one-dimensional Bose-Einstein condensates with strong dipole-dipole interaction. This breathing phenomenon is closely related to the after-collision short
-lived attraction of solitons induced by the dipolar effect. The initial phase difference of solitons leads to the asymmetric dynamics after collision, which is manifested on their different breathing amplitude, breathing frequency, and atom number. We clarify that the asymmetry of breathing frequency is directly induced by the asymmetric atom number, rather than initial phase difference. Moreover, the collision between breathing solitons can produce new after-two-collision breathing solitons, whose breathing amplitude can be adjusted and reach the maximum (or minimum) when the peak-peak (or dip-dip) collision happens.
BERT has achieved superior performances on Natural Language Understanding (NLU) tasks. However, BERT possesses a large number of parameters and demands certain resources to deploy. For acceleration, Dynamic Early Exiting for BERT (DeeBERT) has been p
roposed recently, which incorporates multiple exits and adopts a dynamic early-exit mechanism to ensure efficient inference. While obtaining an efficiency-performance tradeoff, the performances of early exits in multi-exit BERT are significantly worse than late exits. In this paper, we leverage gradient regularized self-distillation for RObust training of Multi-Exit BERT (RomeBERT), which can effectively solve the performance imbalance problem between early and late exits. Moreover, the proposed RomeBERT adopts a one-stage joint training strategy for multi-exits and the BERT backbone while DeeBERT needs two stages that require more training time. Extensive experiments on GLUE datasets are performed to demonstrate the superiority of our approach. Our code is available at https://github.com/romebert/RomeBERT.
