In this paper we investigate the performance of different types rectified activation functions in convolutional neural network: standard linear unit (ReLU), leaky (Leaky ReLU), parametric (PReLU) and a new randomized units (RReLU). We evaluate these function on image classification task. Our experiments suggest that incorporating non-zero slope for negative part could consistently improve results. Thus our findings are common belief sparsity is key good ReLU. Moreover, small scale dataset,...

MXNet is a multi-language machine learning (ML) library to ease the development of ML algorithms, especially for deep neural networks. Embedded in host language, it blends declarative symbolic expression with imperative tensor computation. It offers auto differentiation derive gradients. computation and memory efficient runs on various heterogeneous systems, ranging from mobile devices distributed GPU clusters. This paper describes both API design system implementation MXNet, explains how...

The robustness of neural networks to intended perturbations has recently attracted significant attention. In this paper, we propose a new method, \emph{learning with strong adversary}, that learns robust classifiers from supervised data. proposed method takes finding adversarial examples as an intermediate step. A and simple way is presented experimentally shown be efficient. Experimental results demonstrate resulting learning greatly improves the classification models produced.

This paper presents a method for aspect based sentiment classification tasks, named convolutional multi-head self-attention memory network (CMA-MemNet). is an improved model on networks, and makes it possible to extract more rich complex semantic information from sequences aspects. In order fix the network's inability capture context-related word-level, we propose utilizing convolution n-gram grammatical information. We use make up problem where ignores of sequence itself. Meanwhile, unlike...

Compared with unimodal data, multimodal data can provide more features to help the model analyze sentiment of data. Previous research works rarely consider token-level feature fusion, and few explore learning common related in fuse features. In this paper, we propose a Contrastive Learning Multi-Layer Fusion (CLMLF) method for detection. Specifically, first encode text image obtain hidden representations, then use multi-layer fusion module align image. addition analysis task, also designed...

Since the discovery of electrochemical coloration phenomenon, electrochromic devices capable monitoring transmittance, reflectance, and absorption at designated wavelengths have embraced great achievements. The marriage electrochemistry optical modulation has infused fascinating properties in devices, which find applications thermal management, display, smart windows, camouflage. Inspired by multipronged advancements incorporation multivalent metal ions having rich into is bloomed recent...

Reversible metal electrodeposition (RME)-based smart windows can realize large solar heat gain coefficient (SHGC) modulation, where Zn-RME emerges as an intriguing option with high theoretical coloration efficiency (CE), cost effectiveness, and color neutrality. Herein, Cu2+ was selected electrolyte additive to lower the activation energy homogenize electrical field distribution during plating, forming gradient CuZn alloy nanoparticles via reversible (RAE). Compared Zn-RME, CuZn-RAE...

This article deals with the non-line-of-sight (NLOS) reception issue in field of global navigation satellite system (GNSS). The NLOS has attracted a significant amount attention because it is one main factors that limit GNSS position accuracy urban areas. In this article, we dig into baseband signal processing level to explore new solution detection and correction by means vector tracking loop (VTL). effects on both conventional scalar loops (STLs) VTL are derived mathematically. Based this,...

Electrochromic technology has witnessed numerous achievements in recent years both research and commercialization. devices (ECD) based on different electrochemical mechanism have been developed for various applications, ranging from smart windows, thermal management, rear views, display, camouflage, etc. Compared to conventional ECDs monovalent charge carriers (e.g., H + , Li ), incorporating multivalent ions with rich electrochemistry, high density, small ionic radius opened new...

Abstract Electrochromic technology has recently made many achievements in research and commercialization. devices are being developed based on various coating printing methods for multipronged applications, have great potential next‐generation flexible electronics. Compared to other techniques, inkjet (IJP) enables non‐contact patterning a variety of substrates by programming the movement nozzle. IJP advantages smart electrochromic because its low cost, high resolution, material utilization...

Recent urbanization has posed challenges for the global navigation satellite system (GNSS) to provide accurate solutions. This is especially true in GNSS-denied environments, where clear line of sight (LOS) path between satellites and receiver lacking. For such fusion-based techniques relying on external sensors and/or other signals are widely used. However, may not be feasible cost-effective every time. To overcome these limitations, this work proposes a that makes explicit use past...

Abstract Forecasting river flow is important to water resources management and planning. In this study, an artificial neural network (ANN) model was successfully developed forecast in Apalachicola River. The used a feed‐forward, back‐propagation structure with optimized conjugated training algorithm. Using long‐term observations of rainfall during 1939–2000, the ANN satisfactorily trained verified. Model predictions match well observations. correlation coefficients between forecasting...

Representation learning, especially which by using deep has been widely applied in classification. However, how to use limited size of labeled data achieve good classification performance with neural network, and can the learned features further improve remain indefinite. In this paper, we propose Horizontal Voting Vertical Stacked Ensemble methods networks. ICML 2013 Black Box Challenge, via these independently, Bing Xu achieved 3rd public leaderboard, 7th private leaderboard; Jingjing Xie...

This paper proposes to use a correlator-level global positioning system (GPS) line-of-sight/multipath/non-line-of-sight (LOS/MP/NLOS) signal reception classifier improve performance in an urban environment. Conventional LOS/MP/NLOS classifiers, referred as national marine electronics association (NMEA)-level and receiver independent exchange format (RINEX)-level are usually performed using attributes extracted from basic observables or measurements such received strength, satellite elevation...

Performing precise positioning is still challenging for autonomous driving. Global navigation satellite system (GNSS) performance can be significantly degraded due to the non-line-of-sight (NLOS) reception. Recently, studies of 3D building model aided (3DMA) GNSS show promising improvements in urban canyons. In this study, benefits 3DMA are further extended GNSS/inertial (INS) integration system. Based on shadow matching solution and scoring information candidate positions, two methods...

The prevailing graph neural network models have achieved significant progress in representation learning. However, this paper, we uncover an ever-overlooked phenomenon: the pre-trained learning model tested with full graphs underperforms well-pruned graphs. This observation reveals that there exist confounders graphs, which may interfere semantic information, and current methods not eliminated their influence. To tackle issue, propose Robust Causal Graph Representation Learning (RCGRL) to...

Abstract Multipath is recognized as one of the major error sources for GNSS urban navigation. This study proposes a random forest (RF)-based multipath parameter estimator that uses regression estimation, thereby mitigating effect by removing estimated reflected signal components. The proposed evaluated and compared with estimation delay-lock loop (MEDLL) one-multipath three-multipath cases, respectively. Simulation results demonstrate RF-based less affected front-end bandwidth received...