Detecting small objects is a challenging task. We focus on a special case: the detection and classification of traffic signals in street views. We present a novel framework that utilizes a visual attention model to make detection more efficient, without loss of accuracy, and which generalizes. The attention model is designed to generate a small set of candidate regions at a suitable scale so that small targets can be better located and classified. In order to evaluate our method in the context of traffic signal detection, we have built a traffic light benchmark with over 15,000 traffic light instances, based on Tencent street view panoramas. We have tested our method both on the dataset we have built and the Tsinghua-Tencent 100K (TT100K) traffic sign benchmark. Experiments show that our method has superior detection performance and is quicker than the general faster RCNN object detection framework on both datasets. It is competitive with state-of-theart specialist traffic sign detectors on TT100K, but is an order of magnitude faster. To show generality, we tested it on the LISA dataset without tuning, and obtained an average precision in excess of 90%.

No abstract

The fully memristive neural network is emerging as a game‐changer in the artificial intelligence competition. Artificial synapses and neurons, as two fundamental elements for hardware neural networks, have been substantially implemented by different devices with memory and threshold switching (TS) behaviors, respectively. However, obtaining controllable memory and TS behaviors in the same memristive material system is still a considerable challenge that holds great potential for realizing compatible artificial neurons and synapses. Here, a heterogeneous bilayer conductive filamentary memristor comprising two different electrolytes with distinct copper ion mobility is reported: Cu/GeTe/Al2O3/Pt, which can demonstrate the governance of switching types. Experimentally, when the thickness of the Al2O3 layer is 3 nm, stable nonvolatile multilevel memory switching (MS) is observed and employed to mimic the synaptic plasticity. With increasing oxide thickness, the switching behavior under the same compliance current alters from MS to volatile TS and is used to emulate the integrate‐and‐fire neuron function. The controllable switching stems from the change in the metal filament morphology within the Al2O3 layer, which is supported by ab initio calculation results. This method enables a new pathway for constructing functionally reconfigurable neuromorphic devices for intelligence neuromorphic systems.

The enhanced preservation potential of continental material during continental convergence is thought to be responsible for the episodic continental growth process. However, the mechanism of preservation potential variation is unclear. In this study, we use a novel high‐density passive‐source seismic approach to image the whole‐crust architecture of the juvenile continent in southern Altaids. Two arcuate crust fragments are found between Paleozoic island‐arc belts, which indicate the relicts of inter‐arc oceanic basins. The results show that the trapped oceanic basins make up a large proportion of the juvenile continental crust and that the reduction in subduction erosion due to its incomplete subduction can explain a period of rapid continental growth revealed by previous zircon studies. We suggest that a large number of ocean basins may be trapped during supercontinent formation, and they play a critical role in continental material preservation and continental episodic growth.

No abstract

True random number generators (TRNGs) that can generate unpredictable data by exploiting physical entropy sources are the main security primitive. Despite impressive demonstrations of TRNGs with various memories by exploiting their spatiotemporal variability, realizing efficient and reliable TRNGs without calibration remains a significant challenge. Diffusive memristors with rich stochastic switching behaviors offer an attractive alternative to designing efficient TRNGs, but they are still plagued by difficulties in improving device performance and inefficient circuit calibration. Here, the authors report a Ag/TiN/HfOx/HfOy/HfOx/Pt diffusive memristor with superior threshold switching characteristics, enabled by the good control of switching dynamics by introducing a TiN barrier and trilayer‐hafnia with different ion mobility and migrate barriers. The random integrate‐and‐fire behaviors of the memristor are exploited as excellent intrinsic entropy sources for random number generation with robustness to device variation and without extra calibration. Experimental measurements validate the highly stochastic, compact, self‐clocking TRNG design with a high throughput of ≈108 kb s‐1. The generated random bits have successfully passed all 15 National Institute of Standards and Technology randomness test benchmarks. This work demonstrates that high‐performance diffusive memristors can play an important role in hardware security electronics deployed in edge computing systems.

Figure 1: We propose an adversarial approach for denoising Monte Carlo Renderings (DMCR-GAN). Our network uses several convolution dense blocks to extract rich information of auxiliary buffers and then use these different hierarchical features to modulate the noisy features in the residual blocks. Moreover, we introduce the channel and spatial attention mechanism to exploit the dependencies of inter-channel and inter-spatial features. We use the public dataset [Bitterli 2016] rendered by Tungsten renderer and compare our results with works of NFOR [Bitterli et al. 2016], KPCN[Bako et al. 2017] and ACFM [Xu et al. 2019] methods. The noisy input image and the ground truth image are rendered with 32 spp and 32k spp respectively. The values above images indicate the RMSE, PSNR and SSIM metrics. The comparison of denoised images generated by our proposed and other methods is shown in the supplementary material.

No abstract