Search

The modern conversational agent requires high-quality datasets, which are often the bottlenecks when building models. This paper introduces MatDC, an entirely human-produced dialogue dataset with full semantic annotations in Chinese. The dataset features linguistic variations given users’ intents and fully annotated semantic slots. MatDC dataset was completely human-edited, and the curation comprises two stages. At first, templates design stage, domain editors first construct schemas and compose ten dialogues between the agents and the users based on the back- end database. Secondly, in the dialogue rewrite stage, rewriters generate sentential variations for each template, under the constraints that the normalized slot values are kept unchanged. The underlying methodology of the MatDC is more open to extension and more adaptable to different domains. To demonstrate the applicability of the dataset, we build a dialogue agent with conventional pipeline architecture. We expect the MatDC dataset to provide additional training data and testing ground for dialogue agent studies.

In this paper, we study the problem of blind spectrum sensing by exploring signal sampling at each cognitive radio (CR) in a distributed cognitive radio network. Specifically, a new cooperative sampling scheme is proposed to deal with the challenge of unknown signal-to-noise ratio and level of noise power at each CR, which is a more realistic assumption in distributed spectrum sensing. We prove that the observations acquired by the proposed collaborative sampling are informative from an information theoretic perspective. Simulation results also demonstrate better receiver operating characteristics (ROC) of the proposed sampling scheme than the traditional uniform sampling scheme.

We design an array of transverse slots cut out from the broad wall of a groove-gap waveguide, with periodic spacing between slots of a guided wavelength at 28 GHz that is centered about an important millimeter-wave band. External horns are brought upon to address the problem of grating lobes. The simulations have been carried out for operation in the designed frequency band from 27.676 to 28.652 GHz while preserving the conditions for port matching. This antenna solution lends itself to many applications within this millimeter-wave band.

We report on the strain relaxation properties of an InAs 0.43 P 0.57 / In 0.73 Ga 0.27 As/InAs 0.43 P 0.57 triple layer photodetector structure with InAs y P 1−y metamorphic step-graded buffer layers grown on InP substrate. Reciprocal space mapping (RSM) of (004) reflection along [110] direction show that the InAs 0.3 P 0.57 and In 0.73 Ga 0.27 As slightly tilt toward opposite direction and separates to two lattice points. However, in the RSM of (115) reflection, only one lattice point is observed. Our results indicate the layers fully relaxed in the coordinate system of the substrate, not in that of the tilted layers. TEM analysis indicates that there is no threading dislocation within the InGaAs absorption layer. I-V measurement on the diffusion junction of the detector shows a dark current value of 2.10 × 10−8 A at a bias voltage of −0.5 V, which is among the best of the current state of art. The energy gap obtained from room temperature PL is 0.573 eV, corresponding to a cut-off wavelength of ∼2.16 μm.

To interact with humans in Human Social Environments (HSEs), robots are expected to possess the ability of situational context perception and behave appropriately. In this paper, we propose two deep learning models, as situational context perception of robot, to learn from observations of human-robot interaction. Based on these models, we endow robot the capability of perceiving human's mentation. Thus, the appropriate social behaviors can be performed by the robot with respect to human's mental state. The experimental results demonstrate that robot can significantly improve the accuracy of predicting a person's mentation through the proposed deep learning models by comparison to conventional classifiers and possess potential of providing agreeable serving.

In this paper, a 13.56 MHz regulated dual-output active rectifier is proposed for implantable medical devices (IMDs). By adopting the pulse-width modulation (PWM) to control the conduction time of the rectifying switches, the proposed active rectifier is able to realize the rectification and regulation of dual-output voltages without adding DC-DC converters. The proposed automatic digital offset compensation (ADOC) can automatically compensate both turn-on and turn-off delay to maximize the conversion power and minimize the reverse current. Dual-output voltages 1.8 V and 3.3 V are generated to supply the internal circuit in IMDs. The simulated maximum power conversion efficiency (PCE) is 78.5%, and the input operating voltage ranges from 3.6 V to 5 V. Comparing to the conventional constant offset calibration, the efficiency improvement of 8% is obtained.

An eccentric annular slot patch antenna is initially studied for UHF (ultrahigh frequency) RFID (radio frequency identification) reader antenna application that can operate within the North America RFID band (902–928 MHz). Circular polarization (CP) radiation is achieved in this band by loading two narrow open-ended slots (slits) into the slot patch. To further enhance the CP bandwidth, a parasitic element is loaded beside the patch antenna that forms the proposed antenna. From the experimental results, the proposed antenna can yield an impedance bandwidth (10-dB return loss) between 618 to 1105 MHz, while good CP bandwidth (3-dB axial ratio, AR) from 886 to 942 MHz is also attained. In addition to that, gain level and efficiency of more than 7.6 dBic and 91%, respectively, were also measured.

Feature selection is essential to rule learning in the context of functional verification. In practice today, features are selected manually and the selection requires domain knowledge. In contrast, this work proposes using automatic feature extraction from design documents as a viable approach to support rule learning. To demonstrate its effectiveness, document-extracted features are employed to learn the rules for covering a set of assertions based on a commercial SoC. Experiments show that 100%-accurate rules can be obtained for more than 70% of the assertions.

This paper describes an algorithm for localization of a robot which can efficiently estimate robot in 6 degrees-of-freedom (DoF) pose which consist of position and orientation with large scale point cloud data without giving the initial pose. We introduce the Fast Scene Recognition and Alignment algorithm to reduce the computation time needed for the point cloud alignment by matching robot's scene only with the retrieved Sub-Map in database. Our developed algorithm is to extract Sub-Maps descriptor by cascading several features, and learn a Distance-Metric to increase the precision of place recognition due to the environmental changes. We then align the robot's scene with Sub-Map to estimate robot pose. Our technique has been implemented and tested extensively in different buildings. The experimental results show that our Fast Scene Recognition and Alignment system can localize mobile robot in a variety of large scale 3D point cloud dataset efficiently.

In this paper, a 13.56 MHz power-efficient CMOS active rectifier is proposed for implantable medical devices. The digital offset compensation technique automatically compensates the turn-off delay to minimize the reverse leakage current to improve the conversion efficiency. The simulation results show that the proposed rectifier achieves 92.9% PCE with wide input voltage range. The input operating voltage ranges from 1.5V to 4.5V and the efficiency improvement of 6.5% is also obtained.

Given the increasing number of mobile platforms, a key technical challenge is how to provide an optimal photo browsing experience given the limited screen size available on mobile devices. This paper proposes a novel technique for intelligent mobile image categorization on mobile platform to reduce computation complexity based on cloud computing. In this technique, captured images are analyzed to detect visual salient area, which is then classified in real-time using sparse representation. Mathematically, the derived algorithm regards the salient regions as the dictionary in sparse representation, and selects the salient regions that minimize the residual output error iteratively, thus the resulting regions have a direct correspondence to the performance requirements of the given problem. Experimental results obtained using extensive datasets captured under uncontrolled conditions show the proposed system effectively manages mobile images using sparse representation on cloud computing.

Recently, NAND flash memory has become a widely-used data storage media. However, it has endurance problem that each NAND flash block has limited erase cycles. The first time erase failure usually occurs when a block has over 10K–100K erase cycles. Wear-leveling is a technique to solve this problem and extend the lifetime of a flash memory. In this paper, a conditional threshold wear-leveling algorithm fitting multi-channel architecture is proposed. The overall wear-leveling operation is separated and distributed to different channels. This reduces the runtime of the wear-leveling operation. The conditional threshold can reduce extra erase operations caused by wear-leveling when the flash memory is young. As the flash memory becomes older, the smaller threshold makes wear-leveling work more frequently to get more even erase cycle distribution.

A K-band low-power, low phase noise modified Colpitts VCO was implemented in 0.13 µm CMOS process. The timed current switches of the conventional Colpitts VCO are replaced by a differential transformer, and thus the supply voltage of the proposed VCO can be reduced. The VCO has −100 and −109.5 dBc/Hz phase noises at 1 MHz offset under 4 and 10 mW dc power consumptions, respectively. The chip size is 570 × 570 µm2.

An ultra-low-power VCO based on the complementary cross-coupled structure with the three-coil transformer feedback is proposed and realized in 0.18 µm CMOS. This VCO has a 1 MHz offset phase noise of −118.5 dBc/Hz and consumes only 0.66 mW dc consumption. The FOM of this VCO is better than −195 dBc/Hz form 8.09 to 8.42 GHz, and the highest FOM of −198.6 dBc/Hz occurs at 8.2 GHz. The chip size is 530 × 430 µm2.

In this paper, we present an interactive system to transfer and concatenate motions in different articulated figures, such as human and dog. This system first constructs a union skeleton that contains both bone structures with bone correspondence manually assigned by users, and then aligns the initial poses of these two skeletons. With the initial poses aligned, motion data can be transferred between them. In addition, by blending similar poses of different skeletons, we can produce a seamless transition sequence which can be used to drive a meta-mesh and generate an animated morphing result.

In this paper, we present an efficient and effective digital watermarking method for image tamper protection and recovery. Our method is efficient as it only uses simple operations such as parity check and the comparison between the average intensities. It is effective because the detection is hierarchically structured such that the accuracy of tamper localization can be ensured. That is, if a tampered block, as small as 2times2 pixels, is not detected in level-1 inspection, it will be detected in level-2 or level-3 inspection with probability nearly 1. Our method is also very storage effective, as it only requires a secret key K and a public chaotic mixing algorithm to recover a tampered image. Compared with the method in Celik et al (IEEE Trans. Image Proc., vol. 11, no. 4, pp. 585-594, 2002), our method not only is as simple and as effective in tamper detection and localization, it also provides the capability of tamper recovery