The convolution neural network (CNN) is vulnerable to the adversarial attack, because the attack can generate adversarial images to force the CNN to misclassify the original label of the clean image. To defend against...
详细信息
This paper proposes an improved semisupervised K-means clustering algorithm to deal with the data set which has a small number of labeled *** with the external indexes,this algorithm determines the optimal cluster num...
详细信息
ISBN:
(纸本)9781467391955
This paper proposes an improved semisupervised K-means clustering algorithm to deal with the data set which has a small number of labeled *** with the external indexes,this algorithm determines the optimal cluster number and the initial clustering *** cluster effect is *** to the experience and the external information offered by the labeled data,this algorithm selects the maximum and minimum values of the cluster *** each cluster number,it determines the initial clustering centers according to the labeled data and measure the clustering *** the optimal clustering result is *** simulation experiment shows that the algorithm in this paper has improved the cluster *** also has the high veracity and stability.
Recent advances in self-supervised speech models have shown significant improvement in many downstream tasks. However, these models predominantly centered on frame-level training objectives, which can fall short in sp...
详细信息
ISBN:
(数字)9798350374513
ISBN:
(纸本)9798350374520
Recent advances in self-supervised speech models have shown significant improvement in many downstream tasks. However, these models predominantly centered on frame-level training objectives, which can fall short in spoken language understanding tasks that require semantic comprehension. Existing works often rely on additional speech-text data as intermediate targets, which is costly in the real-world setting. To address this challenge, we propose Pseudo-Word HuBERT (PW-HuBERT), a framework that integrates pseudo word-level targets into the training process, where the targets are derived from a visually-ground speech model, notably eliminating the need for speech-text paired data. Our experimental results on four spoken language understanding (SLU) benchmarks suggest the superiority of our model in capturing semantic information.
High resolution direction-of-arrival (DOA) estimation is one of the most challenging problems in array signalprocessing. In this paper, a variation of the improved polynomial rooting (IPR) method is proposed for DOA ...
详细信息
High resolution direction-of-arrival (DOA) estimation is one of the most challenging problems in array signalprocessing. In this paper, a variation of the improved polynomial rooting (IPR) method is proposed for DOA estimation of multiple targets by a sensor array. The variation, unitary IPR (UIPR), transforms the complex-valued covariance matrix of the sensor signals to a real-valued matrix using unitary transformations. Then the IPR method is applied to determine the DOA of the targets. Simulation results indicate the potential improvement provided by our approach compared with MUSIC, Root-MUSIC, ESPRIT, and IPR.
Providing efficient machine learning (ML) analytics over relational data is a mainstream requirement for data management systems [2,5]. Several projects are building tools to closely integrate ML and dataprocessing [...
详细信息
ISBN:
(纸本)9781450341998
Providing efficient machine learning (ML) analytics over relational data is a mainstream requirement for data management systems [2,5]. Several projects are building tools to closely integrate ML and dataprocessing [3, 4, 9, 10]. Thus, there is a lot of interest in accelerating ML workloads using data management ideas. Recently, there are quite a few studies on accelerating ML workload by exploiting data redundancy introduced by joins [11-13]. For example, [12] shows how to train generalized linear models (GLMs) over primary key-foreign key (PK-FK) joined tables. [13] designs and analyzes a mechanism for speeding up linear regression over factorized joins. However, there are several limitations of the existing results. First, they are all tied closely with the underlying data systems, which incurs a large development overhead when a developer needs to implement the accelerated algorithms in his/her own systems. Secondly, to the best of our knowledge, all existing systems only work for a small subset of ML algorithms. [12] aims at GLMs, while [13] targets exclusively on linear regression. Consequently, it poses challenges for data scientists to design new ML algorithms over mutli-table data. This raises an important question: how to exploit redundancy in data generated by joins such that it i) helps developers avoid a large development overhead and ii) enables data scientists to design new algorithms without bothering about joins? In this abstract, we answer this question by pushing linear algebra (LA) operators over joins;we call this factorized linear algebra. The key insight is that a number of ML algorithms share common operators. Accelerating these operators can automatically speed up a large set of ML algorithms. Note that, by analogy to the database world where relational algebra is the main language, LA is the dominant language in the ML world. Thus, we focus on optimizing LA over normalized data, show how to speed up LA operators (e.g., matrix multiplicatio
This paper presents a new neural solution for solving the data association problem. This problem, also known as the multidimensional assignment problem, arises in data fusion systems like radar and sonar targets track...
详细信息
ISBN:
(纸本)0780350413
This paper presents a new neural solution for solving the data association problem. This problem, also known as the multidimensional assignment problem, arises in data fusion systems like radar and sonar targets tracking, robotic vision... Since it leads to an NP-complete combinatorial optimization, the optimal solution can not be reached in an acceptable calculation time, and the use of approximation methods like the Lagrangian relaxation is necessary. In this paper, we propose an alternative approach based on a Hopfield neural model. We show that it converges to an interesting solution that respects the constraints of the association problem. Some simulation results are presented to illustrate the behaviour of the proposed neural solution for an artificial association problem.
Wireless sensor technologies can provide the leverage needed to enhance patient-caregivers collaboration through ubiquitous access and direct communication, which promotes smart and scalable vital sign monitoring of t...
详细信息
ISBN:
(纸本)9781631900259
Wireless sensor technologies can provide the leverage needed to enhance patient-caregivers collaboration through ubiquitous access and direct communication, which promotes smart and scalable vital sign monitoring of the chronically ill and elderly people live an independent life. However, the design and operation of BASNs are challenging, because of the limited power and small form factor of biomedical sensors. In this paper, an adaptive compression technique that aims at achieving low-complexity energy-efficient compression subject to time delay and distortion constraints is proposed. In particular, we analyze the processing energy consumption, then an energy consumption optimization model with constraints of distortion and time delay is proposed. Using this model, the Personal data Aggregator (PDA) dynamically chooses the optimal compression parameters according to real-time measurements of the packet delivery ratio (PDR) or individual users. To evaluate and verify our optimization model, we develop an experimental testbed, where the EEG data is sent to the PDA that compresses the gathered data and forwards it to the server which decompresses and reconstructs the original signal. Experimental testbed and simulation results show that our adaptive compression technique can offer significant savings in the delivery time with low complexity and without affecting application accuracies.
In order to use the spaceborne optical imaging and electronic reconnaissance data to monitor the ships formation,this paper proposes a target association matching algorithm based on the structure of ships formation an...
详细信息
In order to use the spaceborne optical imaging and electronic reconnaissance data to monitor the ships formation,this paper proposes a target association matching algorithm based on the structure of ships formation and targets'*** results that this algorithm delivers better performance compared to the traditional method,in the situation that because of the sensors are sparse sampling and the targets location precision are greatly different so that can not be established to creation the targets'motion model.
Neural network training targets for speech recognition are estimated using a novel method. Rather than use zero and one, continuous targets are generated using forward-backward probabilities. Each training pattern has...
详细信息
Neural network training targets for speech recognition are estimated using a novel method. Rather than use zero and one, continuous targets are generated using forward-backward probabilities. Each training pattern has more than one class active. Experiments showed that the new method effectively decreased the error rate by 15% in a continuous digits recognition task.
暂无评论