Zero-shot learning (ZSL) utilizes semantic information that is auxiliary information to transfer knowledge from seen classes to unseen classes, thereby realizing the recognition of unseen classes. ...The generative methods train the model to generate pseudo image features for the unseen classes, thus transforming ZSL into a supervised learning problem. However, the classification performances of these methods still suffer from weak generalization. To address this issue, the paper proposes a novel Generative Adversarial Network based on Generative Feature Constraint (GFCGAN). First, based on the LsrGAN 1, a semantic constraint module is added to restrict the generated image features by keeping the semantic information consistent. Then, in the case of keeping the semantic relationships between classes consistent, the generated image features of unseen classes are farther away from the feature prototypes of their most similar classes in both semantic and feature space so as to train the generative model and generate high-quality pseudo image features. Finally, the trained classifier is used to classify images of unseen classes. Experiments on standard datasets verify the effectiveness of our model and achieve good results.
In this study, a high-speed and self-calibratable true random number generator (TRNG) is demonstrated based on a unified selector-RRAM device with dual switching modes. In fast selector mode (∼35 ...ns), the V th fluctuation is the entropy for the high-speed TRNG. However, V th has large device-to-device (D2D) variation, which leads to unstable switching probability. By adjusting the resistance of RRAM, the D2D variation of V th can be substantially regulated through self-calibration ( 10\times \sigma/\mu improvement), enabling the reliable deployment of TRNG.
Generative Adversarial Imitation learning (GAIL) enables the agent to learn a desired policy from the demonstrations provided by human beings. For complex continuous control tasks, such as robotic ...tasks, demonstrations are usually be obtained from multiple demonstrators to save time. However, the differences in the abilities of multiple demonstrators lead to the drift of demonstration distribution, which results in learning a suboptimal policy in the adversarial training process. In this paper, we proposes a new imitation learning (IL) algorithm, IM-GAIL, to enable the agent to learn effectively from imperfect demonstrations of multiple demonstrators. The proposed method rebuild the objective function in the adversarial training, making the policy match the optimal policy. Meanwhile, to improve the sample efficiency, off-policy training framework is adopted and a curious experience replay mechanism is designed for the training. The experiments are conducted on robotic simulation tasks, and results show that our method can efficiently learn from the given demonstrations and it achieves better performance than other state-of-the-art methods.
In this paper, we investigated the influence of reading conditions on foundry BEOL RRAM characteristics for core transistor integration. The resistance states are statistically analyzed by varying ...read voltages from 0.05V to 0.4 V with a dedicated test flow to focus on cycle to cycle variation, nonlinearity and switching window. It indicates that optimization of reading voltage should be reviewed by leveraging switching window, read noise, and nonlinearity to obtain better read performance with a limited voltage swing range. This work aims to shed light on the co-impact of several common parameters in RRAM for advanced technology nodes.
We have successfully demonstrated, for the first time, the STI-less dynamic-gate (DG) technique with self-passivation sidewall (SPS) enhanced RRAM cells on a commercial 40nm CMOS production platform. ...This achievement resulted in a record-density of 15.43 Mb/mm 2 and a high retention of 10years@150°C. Through a comprehensive design-technology co-optimization (DTCO) process, we obtain significant improvements in various key characteristics, as evidenced by experimental results at both the wafer (12-inch) and chip (4K&1M) level. These improvements include improved memory window (>20μA), enhanced uniformity, extended retention (10years@150°C), and multilevel cell (MLC>3bit). This work indicates the potential of RRAM as embedded non-volatile memory (eNVM) in advanced technology node for consumer and industrial applications.
For the first time, a novel hybrid-domain (time, analog, and digital) in-memory polynomial computing is proposed and experimentally demonstrated using a RRAM-based expandable ternary multiplier (ETM) ...to accelerate polynomial transformation (PT) algorithms. The ETM is constructed based on 1T1R cell, which subtly implements multiplication of two multibit signed operands by orchestrating the inputs on the word lines (time-domain sequence) and bit lines (analog voltage-domain) to enable ternary multiplication (TM). The subsequent accumulation is completed through peripheral shift-adder (digital-domain). A 40nm RRAM chip with 30 integrated four-quadrant TM arrays are designed and implemented to realize high-order parallel matrix-vector computation and PT acceleration through synthesizing the hybrid-domain in-memory polynomial computation, demonstrating software-comparable calibration of lens distortion in machine vision applications with high calibration throughput (158Mpixels/s) and energy-efficiency (3.81Gpixels/W).
For the purpose of obtaining desirable dielectric properties, a variety of inorganic particles fillers were doped into epoxy resin composites to change electrical behavior of epoxy resin composites. ...We researched electric properties of epoxy resin composites doped by nonlinear fillers, including zinc oxide(ZnO) with the size of 40nm and barium strontium titanate (BST) with the size about 100-200nm, and measured the dielectric constant (ϵ r ) and conductivity (σ). Different mass fractions (0%, 5%, 10% and 20%) of nano-sized nonlinear powder (BST and ZnO) were doped separately into epoxy composites to obtain multi-phase dielectric composites. As the electric field changed from 10 3 V/m to 10 7 V/m, electrical properties changed accordingly. The results showed that BST and ZnO fillers tended to exhibit enhanced dielectric constant and lower dielectric loss with the increase of effective mass fraction. The influence mechanism was discussed in accordance with mass fractions, electric intensity and polarization mode of each kind of nonlinear fillers.
The dielectric loss is a small value, is also an important indicator to measure the performance of the electrical equipment insulation 1,11 . But outside interference can easily cause a big impact to ...the dielectric loss measurement, and real-time response is required on the dielectric loss online monitoring system. Therefore, high accuracy, rapid digitization of the dielectric loss measurement method is a key part of the dielectric loss online monitoring system. This article describes a practical harmonic analysis method to measure the dielectric loss. Combined with the IEC61850 and IEEE1588, according to the special circumstances of the measurement site, the digital method of measuring dielectric loss is realized through the joint work of a voltage acquisition unit and another entirely different current acquisition unit. Finally the source of calculating errors is verified and the error is measured less than 0.00025 (radian) in test, which is less than 1μs when measured in sampling-time.
To study the effect of xanthnotoxol (XT) on contractility of rabbit thoracic aorta strips and its relationship with Ca2+.
Routine experimental methods for isolated thoracic aorta strips were adopted.
...XT and verapamil (Ver) can make the dose-response curve of rabbit thoracic aorta strips induced by NE, KCl, CaCl2 shifed right (the pD'2 of KCl and CaCl2 value was 3.58 +/- 0.07 and 4.12 +/- 0.12) and depressed the maximal responses. Like verapamil (Ver), it is accomplished by calcium antagonism. XT could selectively block the voltage dependence calcium channel (VDC), and have no effect on the receptor operated calcium channel (ROC).
It suggested XT had selectly blocking effect on voltage dependence calcium channel.