Search

Open domain response generation is the task of creating a response given a user query in any topics/domain. Limited by context and reference information, responses generated by current systems are often “bland” or generic. In this paper, we combine a response generation model with a retrieval system that searches for relevant utterances and responses. The generation model has two main components: a keyword extraction module and a two-stage transformer. The keyword extraction module aims to extract two types of keywords in an unsupervised fashion from the retrieved results: (1) keywords in the query not found in the retrieved utterances (DiffKey), and (2) overlapping keywords among the retrieved responses (SimKey). Given these keywords, the two-stage transformer first decides where to insert the keywords in the response, and the second generates the full response given the location of the keywords. The keyword extraction module and the two-stage transformer are connected in a single network, and so our system is trained end-to-end. Experimental results on Cornell Movie-Dialog corpus, Douban and Weibo demonstrate that our model outperforms state-of-the-art systems in terms of ROUGE, relevance scores and human evaluation.

Under the explosive growth of “low-slow-small” unmanned aerial vehicles (UAVs), harassment incidents with “low-slow-small” UAVs, such as illegal surveying and mapping, no-fly zone flights, terrorist attacks, have frequently been occurring due to the low entry barrier to the industry and the lack of market norms and control standards, and severely impair the safety of public life and property. The premise of developing anti-UAV technology is efficient and accurate UAV target detection, which plays a vital role in the follow-up UAV target tracking, attack, and interception. This paper proposes a new DEAX algorithm named enhanced adaptive feature pyramid networks-based Target Detection Algorithm for infrared small UAV target detection. Our proposed algorithm improves the original feature pyramid networks in four aspects for the small UAV target detection task. 1) Use channel separation to keep more channels when adjusting the number of channels in convolutional to avoid the loss of helpful feature information; 2) Design a feature enhancement module to enhance “target feature” and suppress “non-target feature”; 3) Alleviate the differences of receptive fields and semantic information among different layers by shared convolution; 4) Introduce adaptive feature fusion method into feature fusion and complete the construction of enhanced adaptive feature pyramid networks (EAFPN) to solve the problem of weakening feature expression in cross-scale feature fusion. EAFPN is applied to the single-stage target detection networks. We find the detection accuracy and speed of the algorithm outperform those based on feature pyramid networks (FPN), where the improvement on mAP is 8.9%.

Most of the existing deception jamming techniques focus on designing effective jamming signals for static targets or scenes. This paper proposes a multi-scene deception jamming against synthetic aperture radar (SAR) imaging with frequency diverse array (FDA) antenna. Different from traditional single-element and phased-array antennas, FDA antenna employs a small frequency offset across the array elements to produce angle-, range- and even time-dependent transmit beampattern, which brings bright prospect for the application of new jamming techniques against SAR imaging. Such a new method will generate many false targets or multiple deception scenes for the jammed SAR, which helps jam better on the static scene and moving target. Moreover, the location and number of false targets can be manipulated. This paper will illustrate the FDA-based deception jamming against SAR imaging, specially designed waveform and the relationship between the number of false targets or scenes and FDA parameters. The relationship between the location of false targets and the position of jammer will also be described in detail. All proposed methods are validated by extensive simulation results.

With the continuous development of wireless network technology, more and more mobile devices are connected to the network, and the location information of these devices is becoming one of the important basis for analyzing other geographic data in the net. Location is an important problem in spatial information sensing, and the theoretical basis of location is the free space fading feature of wireless signals. In this paper, we proposes WiLocWare, which is a middle ware to understand and modeling the performance of the state-of-the-art and to be proposed algorithms. We evaluated and characterized the correlation between localization accuracy and networks parameters such as signal propagation model, the coverage of wireless radios, the distributions of wireless devices and the density of the anchor nodes. The experimental results show the localization accuracies under different wireless signatures, WiLocWare is also scalable for the performance evaluation for the to be proposed algorithms.

Medical text classification assigns medical related text into different categories such as topics or disease types. Machine learning based techniques have been widely used to perform such tasks despite the obvious drawback in such “black box” approach, leaving no easy way to fine-tune the resultant model for better performance. We propose a novel constructive heuristic approach to generate a set of regular expressions that can be used as effective text classifiers. The main innovation of our approach is that we develop a novel regular expression based text classifier with both satisfactory classification performance and excellent interpretability. We evaluate our framework on real-world medical data provided by our collaborator, one of the largest online healthcare providers in the market, and observe the high performance and consistency of this approach. Experimental results show that the machine-generated regular expressions can be effectively used in conjunction with machine learning techniques to perform medical text classification tasks. The proposed methodology improves the performance of baseline methods (Naive Bayes and Support Vector Machines) by 9% in precision and 4.5% in recall. We also evaluate the performance of modified regular expressions by human experts and demonstrate the potential of practical applications using the proposed method.

Cell shapes provide crucial biological information on complex tissues. Different cell types often have distinct cell shapes, and collective shape changes usually indicate morphogenetic events and mechanisms. The identification and detection of collective cell shape changes in an extensive collection of 3D time-lapse images of complex tissues is an important step in assaying such mechanisms but is a tedious and time-consuming task. Machine learning provides new opportunities to automatically detect cell shape changes. However, it is challenging to generate sufficient training samples for pattern identification through deep learning because of a limited amount of images and annotations. We present a deep learning approach with minimal well-annotated training samples and apply it to identify multicellular rosettes from 3D live images of the Caenorhabditis elegans embryo with fluorescently labeled cell membranes. Our strategy is to combine two approaches, namely, feature transfer and generative adversarial networks (GANs), to boost image classification with small training samples. Specifically, we use a GAN framework and conduct an unsupervised training to capture the general characteristics of cell membrane images with 11,250 unlabelled images. We then transfer the structure of the GAN discriminator into a new Alex-style neural network for further learning with several dozen labeled samples. Our experiments showed that with 10-15 well-labeled rosette images and 30-40 randomly selected nonrosette images our approach can identify rosettes with more than 80% accuracy and capture more than 90% of the model accuracy achieved with a training data et that is five times larger. We also established a public benchmark dataset for rosette detection. This GAN-based transfer approach can be applied to the study of other cellular structures with minimal training samples.

RF-based gesture sensing and recognition has increasingly attracted intense academic and industrial interest due to its various device-free applications in daily life, such as elder monitoring, mobile games. State-of-the-art approaches achieved accurate gesture sensing by using fine-grained RF signatures (such as CSI, Doppler effect) while could not achieve the same accuracy with coarse-grained RF signatures such as received signal strength (RSS). This paper presents rRuler, a novel feature extraction method which aims to get fine-grained human gesture features with coarse-grained RSS readings, which means rought ruler could measure fine things. In order to further verify the performance of rRuler, we further propose rRuler-HMM, a hidden Markov model (HMM) based human gesture sensing and prediction algorithm which utilizes the features extracted by rRuler as input. We implemented rRuler and rRuler-HMM using TI Sensortag platforms and off-the-shelf (CTOS) laptops in an indoor environment, extensively performance evaluations show that rRuler and rRuler-HMM stand out for their low cost and high practicability, and the average gesture sensing accuracy of rRuler-HMM can achieve 95.71% in NLoS scenario and 97.14% in LoS scenario, respectively, which is similar to the performance that fine-grained RF signatures based approaches could achieve.

Spaceborne azimuth multichannel bistatic synthetic aperture radar (AMC-BiSAR) system employed inclined geosynchronous (GEO) transmitter and low-Earth-orbit (LEO) receiver is capable of providing a vast area of surveillance and fine spatial resolution, which presents great potentials for Earth observation. Nevertheless, owing to complex geometry and big differences of orbital characteristics, high-resolution wide-swath (HRWS) imaging for GEO-LEO AMC-BiSARs remains three major issues: 1) “Stop-and-Go” approximation is invalid because of long round-trip delay and high receiver velocity; 2) Due to the special bistatic configuration, the conventional effective phase center (EPC) operation for unambiguous signal reconstruction cannot be directly used; 3) Two-dimension spatial-variant characteristics of the echo signal caused by nonlinear and nonparallel trajectories of transmitter and receiver raise challenges for BiSAR focusing. To address these problems, starting from modeling geometry of inclined GEO-LEO AMC-BiSAR, an accurate slant range model without “Stop-and-Go” approximation is derived in the paper. Then the new effective positions of the receiving channels in GEO-LEO AMC-BiSARs are obtained based on the geometry, and the unambiguous Doppler spectrum can be recovered effectively. Finally, a nonlinear chirp scaling imaging algorithm based on the time-domain perturbation (TP-NLCS) is derived and presented, which can correct the variation of range cell migration and azimuth frequency modulation rate. Simulation results show the validity of the proposed method, which can implement GEO-LEO AMC-BiSAR HRWS focusing and has fine phase-preserved capability.

Inspired by recent advances in frequency diverse array (FDA) and multiple-input-multiple-output (MIMO) radars, this paper investigates the target detection performance of FDA-MIMO radar, which jointly utilizes the advantages of FDA radar and MIMO radar, for single- and multiple- coherent pulses, respectively. Firstly, the signal models in target detection are formulated for FDA-MIMO, standard FDA, phased-array and MIMO radars, respectively. We present a unified framework detector to comparatively analyze their target detection performance in the Neyman-Pearson sense. Furthermore, the deflection coefficient and signal-to-interference-plus-noise ratio (SINR) are also adopted as a performance metric to compare their performance. Numerical results show that FDA-MIMO radar achieves better target detection performance as compared to other conventional radar systems, specially when multiple coherent pulses are utilized.

This paper focuses on the long-overlooked phenomenon that individual travel patterns are not always stable over the long term and may change due to seasonal changes, moving, and work schedule changes. Unlike previous studies that identified sudden peak points, this paper treats travel pattern change as a change point detection problem in a time series and defines change as “sudden, substantial, and continuous”. Considering the complexity of travel behavior, this paper measures changes in travel patterns in three dimensions: time, space, and frequency, and establishes a Bayesian change point detection model. A nine-month period of private car GPS data from Aichi, Japan, is used for an example analysis. The results show that the Bayesian approach can effectively identify travel pattern changes. Compared with the traditional GLR, the proposed method in this paper has higher recognition accuracy with lower model complexity. Meanwhile, the experimental results show that individual travel patterns may change in only one dimension or in multiple dimensions at the same time. Based on this, the correlation analysis of travel patterns in the temporal and spatial dimensions is carried out, and it is verified that there is a certain positive correlation between the two. The Bayesian change-point detection model proposed is robust and generally applicable to other fields besides travel patterns.

As a novel power steering technology, Differential Drive Assisted Steering (DDAS) technology for the independent-wheel-drive electric vehicle has gradually appealed to researcher's attention. However, the previous experimental results show that its assistance quality cannot be fully accepted due to its caused sensitive steering wheel torque fluctuation in actual work environment. According to the working principle of the DDAS system, it is founded that the road roughness, the front wheel alignment parameters and sensor noise are the main factors that influence the quality of assisted steering and driver's road feel. Hence the three factors are added as interference into the ideal vehicle model. The simulation results and its comparison with the previous real vehicle tests confirm this causality between these factors considered and the steering wheel torque fluctuation of the DDAS system. Then a robust H∞ loop-shaping controller is designed to solve the issue caused by these inner interferences and outer noises. Simulations results validate the proposed controller and show better steering wheel torque performance than the traditional anti-windup PID controller published in the literature earlier. The proposed robust controller is further verified via real vehicle tests and the results are similar to the simulation results which can effectively suppress the steering wheel torque ripple, improve the anti-interference ability of DDAS system and greatly improve the driver's road feel.