Different linguistic perspectives causes many diverse segmentation criteria for Chinese word (CWS). Most existing methods focus on improve the performance each single criterion. However, it is interesting to exploit these different and mining their common underlying knowledge. In this paper, we propose adversarial multi-criteria learning CWS by integrating shared knowledge from multiple heterogeneous criteria. Experiments eight corpora with show that of corpus obtains a significant...

Despite its success in many areas, deep learning is a poor fit for use hardware predictors because these models are impractically large and slow, but this paper shows how we can to help design new cache replacement policy. We first show that replacement, powerful LSTM model an offline setting provide better accuracy than current predictors. then perform analysis interpret model, deriving key insight allows us simple online matches the model's with orders of magnitude lower cost.

This paper presents Voyager, a novel neural network for data prefetching. Unlike previous models prefetching, which are limited to learning delta correlations, our model can also learn address important prefetching irregular sequences of memory accesses. The key solution is its hierarchical structure that separates addresses into pages and offsets introduces mechanism relations among offsets. Voyager provides significant prediction benefits over current prefetchers. For set programs from the...

Abstract Motivation Predicting pathological complete response (pCR) to neoadjuvant chemotherapy (NAC) in triple-negative breast cancer (TNBC) patients accurately is direly needed for clinical decision making. pCR also regarded as a strong predictor of overall survival. In this work, we propose deep learning system predict NAC based on serial pathology images stained with hematoxylin and eosin two immunohistochemical biomarkers (Ki67 PHH3). To support human prior domain knowledge-based...

In the background of cloud, self-destructing data mainly aims at protecting privacy. All and its copies will become destructed or unreadable after a user-specified period, without any user intervention. Besides, anyone cannot get decryption key timeout, neither sender nor receiver. The Washington's Vanish system is for under cloud computing, it vulnerable to "hopping attack" "sniffer attack". We propose new scheme in this paper, called Safe Vanish, prevent hopping attacks by way extending...

It is desirable but challenging to simultaneously support latency SLO at a pre-defined percentile, i.e., the Xth percentile SLO, and throughput for consolidated VM storage. Ensuring contributes accurately differentiating service levels in metric of application-level compliance, especially application built on multiple VMs. However, enforcement are opposite sides same coin due conflicting requirements level IO concurrency. To address this challenge, paper proposes PSLO, framework supporting...

Dialogue disentanglement aims to separate intermingled messages into detached sessions. The existing research focuses on two-step architectures, in which a model first retrieves the relationships between two and then divides message stream clusters. Almost all work puts significant efforts selecting features for message-pair classification clustering, while ignoring semantic coherence within each session. In this paper, we introduce end-to- end transition-based online dialogue...

Abstract The k-Nearest Neighbors (k-NN) algorithm is a classic non-parametric method that has wide applications in data classification and prediction. Like many other machine learning schemes, the performance of k-NN classifiers will be significantly impacted by imbalanced class distributions data. That is, instances majority tend to dominate prediction test instances. In this paper, we look into pre-processing techniques can used rebalance training enhance sets. We conduct extensive...

Conversation disentanglement aims to separate intermingled messages into detached sessions, which is a fundamental task in understanding multi-party conversations. Existing work on conversation relies heavily upon human-annotated datasets, expensive obtain practice. In this work, we explore training model without referencing any human annotations. Our method built the deep co-training algorithm, consists of two neural networks: message-pair classifier and session classifier. The former...

Fraud detection on multi-relation graphs aims to identify fraudsters in graphs. Graph Neural Network (GNN) models leverage graph structures pass messages from neighbors the target nodes, thereby enriching representations of those nodes. However, feature and structural inconsistency graph, owing fraudsters' camouflage behaviors, diminish suspiciousness fraud nodes which hinders effectiveness GNN-based models. In this work, we propose DiG-In-GNN, Discriminative Feature Guided GNN against...

The address allocation policy in SSD aims to translate the logical of I/O requests into a physical address, and static is widely used modern SSD. Through extensive experiments, we find that there are significant differences utilization parallelism among different policies. We also observe fixed design prevents SSDs from continuing meet challenges posed by cloud workloads misses possibility further optimization. These situations stem our excessive reliance on over time. In this paper, propose...

Winograd Schema Challenge (WSC) was proposed as an AI-hard problem in testing computers' intelligence on common sense representation and reasoning. This paper presents the new state-of-theart WSC, achieving accuracy of 71.1%. We demonstrate that leading performance benefits from jointly modelling sentence structures, utilizing knowledge learned cutting-edge pretraining models, performing fine-tuning. conduct detailed analyses, showing fine-tuning is critical for performance, but it helps...

As the performance of computer systems stagnates due to end Moore's Law, there is a need for new models that can understand and optimize execution general purpose code. While growing body work on using Graph Neural Networks (GNNs) learn representations source code, these do not how code dynamically executes. In this work, we propose approach use GNNs fused its execution. Our defines multi-task GNN over low-level program state (i.e., assembly dynamic memory states), converting complex...

Determining the number of clusters is important but challenging, especially for data high dimension. In this paper, we propose Deep Embedding Determination (DED), a method that can solve jointly unknown and feature extraction. DED first combines virtues convolutional autoencoder t-SNE technique to extract low dimensional embedded features. Then it determines using an improved density-based clustering algorithm. Our experimental evaluation on image datasets shows significant improvement over...

Nowadays, artificial intelligence (AI) has a big impact in the field of painting. In contrast to hand-painting and challenging personal creativity, AI applications practices add noise, remove restore image conserve present process after converting data. The various generator models, Diffusion model is latest application consist two main word-image mapping diffusion algorithm. This article introduced mechanism how images are generated makes arguments about several ethical issues generation....

Effective feature selection methods are essential for improving the accuracy and efficiency of text categorization. Motivated by document frequency, we proposed a new filter-based approach, called categorical frequency. The frequency displays distribution term over each category. Mathematically, variance reflects contribution to Finally, experiments carried out on Reuters-21578 standard corpus. results showed that categorization performance approach is similar or better than information gain...

As a fundamental cloud service for modern Web applications, the object storage system stores and retrieves millions or even billions of read-heavy data objects. Serving massive amount requests each day makes response latency be vital component user experiences. Due to lack suitable understanding on distribution, current practice is use overprovision resources meet Service Level Agreement (SLA). Hence we build performance model predict percentiles meeting SLA (response requirement), in...