In recent years, deep learning has spread beyond both academia and industry with many exciting real-world applications. The development of presented obvious privacy issues. However, there been lack scientific study about preservation in learning. this paper, we concentrate on the auto-encoder, a fundamental component learning, propose private auto-encoder (dPA). Our main idea is to enforce ε-differential by perturbing objective functions traditional rather than its results. We apply dPA...

Universal style transfer retains styles from reference images in content images. While existing methods have achieved state-of-the-art performance, they are not aware of the leak phenomenon that image may corrupt after several rounds stylization process. In this paper, we propose ArtFlow to prevent during universal transfer. consists reversible neural flows and an unbiased feature module. It supports both forward backward inferences operates a projection-transfer-reversion scheme. The...

Drug discovery often relies on the successful prediction of protein-ligand binding affinity. Recent advances have shown great promise in applying graph neural networks (GNNs) for better affinity by learning representations complexes. However, existing solutions usually treat complexes as topological data, thus biomolecular structural information is not fully utilized. The essential long-range interactions among atoms are also neglected GNN models. To this end, we propose a structure-aware...

Denoising is a crucial step for hyperspectral image (HSI) applications. Though witnessing the great power of deep learning, existing HSI denoising methods suffer from limitations in capturing nonlocal self-similarity. Trans-formers have shown potential longrange de-pendencies, but few attempts been made with specifically designed Transformer to model spatial and spec-tral correlation HSIs. In this paper, we address these issues by proposing spectral enhanced rectangle Trans-former, driving...

Evaluating on adversarial examples has become a standard procedure to measure robustness of deep learning models. Due the difficulty creating white-box for discrete text input, most analyses NLP models have been done through black-box examples. We investigate character-level neural machine translation (NMT), and contrast adversaries with novel adversary, which employs differentiable string-edit operations rank changes. propose two types attacks aim remove or change word in translation,...

While recent studies on semi-supervised learning have shown remarkable progress in leveraging both labeled and unlabeled data, most of them presume a basic setting the model is randomly initialized. In this work, we consider transfer jointly, leading to more practical competitive paradigm that can utilize powerful pre-trained models from source domain as well labeled/unlabeled data target domain. To better exploit value weights examples, introduce adaptive consistency regularization consists...

Recently many efforts have been devoted to applying graph neural networks (GNNs) molecular property prediction which is a fundamental task for computational drug and material discovery. One of major obstacles hinder the successful by GNNs scarcity labeled data. Though contrastive learning (GCL) methods achieved extraordinary performance with insufficient data, most focused on designing data augmentation schemes general graphs. However, molecule could be altered method (like random...

Electric Vehicle (EV) has become a preferable choice in the modern transportation system due to its environmental and energy sustainability. However, many large cities, EV drivers often fail find proper spots for charging, because of limited charging infrastructures spatiotemporally unbalanced demands. Indeed, recent emergence deep reinforcement learning provides great potential improve experience from various aspects over long-term horizon. In this paper, we propose framework, named...

To understand human behaviors, action recognition based on videos is a common approach. Compared with image-based recognition, provide much more information, reducing the ambiguity of actions. In last decade, many works focus datasets, novel models and learning approaches have improved video to higher level. However, there are challenges unsolved problems, in particular sports analytics where data collection labeling sophisticated, requiring people domain knowledge even sport professionals...

Over the last decade, machine learning (ML) and deep (DL) algorithms have significantly evolved been employed in diverse applications, such as computer vision, natural language processing, automated speech recognition, etc. Real-time safety-critical embedded Internet of Things (IoT) systems, autonomous driving UAVs, drones, security robots, etc., heavily rely on ML/DL-based technologies, accelerated with improvement hardware technologies. The cost a deadline (required time constraint) missed...

Abstract The rising mental health difficulties of the urban population in developing countries may be attributed to high levels air pollution. However, nationwide large-scale empirical works that examine this claim are rare. In study, we construct a daily metric using volume mental-health-related queries on largest search engine China, Baidu, test hypothesis. We find pollution causally undermines people’s and impact becomes stronger as duration exposure increases. Heterogeneity analyses...

Time series forecasting has been a widely explored task of great importance in many applications. However, it is common that real-world time data are recorded short period, which results big gap between the deep model and limited noisy series. In this work, we propose to address problem with generative modeling bidirectional variational auto-encoder (BVAE) equipped diffusion, denoise, disentanglement, namely D3VAE. Specifically, coupled diffusion probabilistic proposed augment without...

Molecular property prediction plays a fundamental role in AI-aided drug discovery to identify candidate molecules, which is also essentially few-shot problem due lack of labeled data. In this paper, we propose Property-Aware Relation networks (PAR) handle problem. We first introduce property-aware molecular encoder transform the generic embeddings ones. Then, design query-dependent relation graph learning module estimate and refine w.r.t. target property. Thus, facts that both...

We present a novel approach for relation classification, using recursive neural network (RNN), based on the shortest path between two entities in dependency graph.Previous works RNN are constituencybased parsing because phrasal nodes parse tree can capture compositionality sentence.Compared with constituency-based trees, graphs represent relations more compactly.This is particularly important sentences distant entities, where spans words that not relevant to relation.In such cases cannot be...

While deep learning succeeds in a wide range of tasks, it highly depends on the massive collection annotated data which is expensive and time-consuming. To lower cost annotation, active has been proposed to interactively query an oracle annotate small proportion informative samples unlabeled dataset. Inspired by fact that with higher loss are usually more model than loss, this paper we present novel approach queries for annotation when sample believed incorporate high loss. The core our...

Recent advances in wireless sensing techniques have made it possible to recognize hand gestures using channel state information (CSI) commodity WiFi devices. Existing WiFi-based gesture recognition systems mainly use learning-based pattern methods different gestures, however, these fail work well when the locations of transceivers, relative location and orientation with respect and/or gesturing size change, leading inconsistent signal patterns caused by those factors. Although some recent...

Mobile Sensing Apps have been widely used as a practical approach to collect behavioral and health-related information from individuals provide timely intervention promote health well-beings, such mental chronic cares. As the objectives of mobile sensing could be either \emph{(a) personalized medicine for individuals} or \emph{(b) public populations}, in this work we review design these apps, propose categorize apps/systems two paradigms -- \emph{(i) Personal Sensing} \emph{(ii) Crowd...

Despite achieving remarkable performance, Federated Learning (FL) suffers from two critical challenges, i.e., limited computational resources and low training efficiency. In this paper, we propose a novel FL framework, FedDUAP, with original contributions, to exploit the insensitive data on server decentralized in edge devices further improve First, dynamic update algorithm is designed server, order dynamically determine optimal steps of for improving convergence accuracy global model....

As a promising approach to deal with distributed data, Federated Learning (FL) achieves major advancements in recent years. FL enables collaborative model training by exploiting the raw data dispersed multiple edge devices. However, is generally non-independent and identically distributed, i.e., statistical heterogeneity, devices significantly differ terms of both computation communication capacity, system heterogeneity. The heterogeneity leads severe accuracy degradation while prolongs...

Summary While data is distributed in multiple edge devices, federated learning (FL) attracting more and attention to collaboratively train a machine model without transferring raw data. FL generally exploits parameter server large number of devices during the whole process training, while several are selected each round. However, straggler may slow down training or even make system crash training. Meanwhile, other idle remain unused. As bandwidth between relatively low, communication...