Earticle

최근 인공지능 딥러닝 분야는 컴퓨팅 자원의 높은 연산량과 가격문제로 인해 상용화에 어려움이 존재했다. 본 논문은 더블 프루닝 기법을 적용하여 심층신경망 모델들과 다수의 데이터셋에서의 성능을 평가하고자 한다. 더 블 프루닝은 기본의 네트워크 간소화(Network-Slimming)과 파라미터 프루닝(Parameter-Pruning)을 결합한다. 이는 기존의 학습에 중요하지 않는 매개변수를 절감하여 학습 정확도를 저해하지 않고 속도를 향상시킬 수 있다는 장점이 있다. 다양한 데이터셋 학습 이후에 프루닝 비율을 증가시켜, 모델의 사이즈를 감소시켰다. NetScore 성능 분석 결과 MobileNet-V3가 가장 성능이 높게 나타났다. 프루닝 이후의 성능은 Cifar 10 데이터셋에서 깊이 우선 합성곱 신경망으로 구성된 MobileNet-V3이 가장 성능이 높았고, 전통적인 합성곱 신경망으로 이루어진 VGGNet, ResNet또한 높은 폭으로 성능이 증가함을 확인하였다.

Recently, the artificial intelligence deep learning field has been hard to commercialize due to the high computing power and the price problem of computing resources. In this paper, we apply a double pruning techniques to evaluate the performance of the in-depth neural network and various datasets. Double pruning combines basic Network-slimming and Parameter-prunning. Our proposed technique has the advantage of reducing the parameters that are not important to the existing learning and improving the speed without compromising the learning accuracy. After training various datasets, the pruning ratio was increased to reduce the size of the model.We confirmed that MobileNet-V3 showed the highest performance as a result of NetScore performance analysis. We confirmed that the performance after pruning was the highest in MobileNet-V3 consisting of depthwise seperable convolution neural networks in the Cifar 10 dataset, and VGGNet and ResNet in traditional convolutional neural networks also increased significantly.

Information spread maximization is to find a small subset of nodes in social network such that they can maximize the expected spread of information. In this paper, we attempt harnessing historical information cascades data to learn how information propagates in social networks and how to maximize its spread. In particular, we proposed a voting algorithm to learn diffusion probabilities of edges from cascades data. Then a pruning method is developed to remove trivial edges whose weights are smaller than a threshold. Moreover, motivated by the social influence locality, we propose a Local Influence Model to evaluate node's influence within a local area instead of the whole network, which can effectively reduce the computational complexity. Based on Local Influence Model, we use greedy algorithm to find an approximate optimal solution. Experimental results show that our method significantly outperforms state-of-the-art models both in terms of information spread and algorithm runtime.

This paper presents a SVM classification method based on cluster boundary sampling and sample pruning. We actively explore an effective solution to solve the difficult problem of imbalanced data set classification from data re-sampling and algorithm improving. Firstly, we creatively propose the method of cluster boundary sampling, using the clustering density threshold and the boundary density threshold to determine the cluster boundaries, in order to guide the process of re-sampling more scientifically and accurately. Secondly, we put forward a new sample pruning algorithm based on dynamic threshold KNN to deal with the complexity and overlapping problem of imbalanced data set. The phenomenon of data complexity and overlapping will reduce the classification performance and generalization ability of SVM classifier. Experiments show that our method acquires obviously promotion effect in various different imbalanced data sets and it can prove the validity and st

It has been known that GMM (Gaussian Mixture Model) based speaker identification systems using ML (Maximum Likelihood) and WMR (Weighting Model Rank) demonstrate very high performances. However, such systems are not so effective under practical environments, in terms of real time processing, because of their high calculation costs. In this paper, we propose a new speaker-pruning algorithm that effectively reduces the calculation cost. In this algorithm, we select 20% of speaker models having higher likelihood with a part of input speech and apply MWMR (Modified Weighted Model Rank) to these selected speaker models to find out identified speaker. To verify the effectiveness of the proposed algorithm, we performed speaker identification experiments using TIMIT database. The proposed method shows more than 60% improvement of reduced processing time than the conventional GMM based system with no pruning, while maintaining the recognition accuracy.

Least squares support vector machine(LS-SVM) is a well known and useful machine learning ways for statistical classification and regression analysis. LS-SVM can be a good substitute for traditional statistical method but computational difficulties are still remained to operate the inversion of matrix of huge data set. In modern information society, we can easily obtain a large data sets by on-line or batch mode. For the analysis of these kind of huge data sets, we suggest an on-line pruning regression method based on LS-SVM. With relatively small number of pruned support vectors, we can have almost same performance as regression with full data set.

4단계 스카이라인 영역 결정 기법[2]은 영역 결정 시간이 객체의 개수에 비례해서 현저히 증가하기 때문에 다수의 객체를 포함하는 도메인들에 적용하기 어렵다. 이러한 문제점은 스카이라인 영역이 지배 객체 집합의 부분 집합으로 이루어지는 특성을 고려하지 않았기 때문에 발생한다. 이 논문에서는 스카이라인 영역 결정에 불필요한 객체들을 제거할 수 있는 거리 기반 가지치기 기법과 영역 결정 선분의 범위 축소 기법을 제안한다. 제안한 기법들을 R*-트리와 INN(Incremental Nearest Neighbor) 알고리즘에 적용함으로써 점진적으로 스카이라인 영역을 결정할 수 있으며 영역 결정 시간을 현저하게 감소시킬 수 있다. 제안한 기법의 성능 향상을 증명하기 위해 4단계 영역 결정 기법과의 비교 실험을 수행한다.

Lithium-ion batteries are actively used in various industrial sites such as field robots, drones, and electric vehicles due to their high energy efficiency, light weight, long life span, and low self-discharge rate. When using a lithium-ion battery in a field, it is important to accurately estimate the SoC (State of Charge) of batteries to prevent damage. In recent years, SoC estimation using data-based artificial neural networks has been in the spotlight, but it has been difficult to deploy in the embedded board environment at the actual site because the computation is heavy and complex. To solve this problem, neural network lightening technologies such as network pruning have recently attracted attention. When pruning a neural network, the performance varies depending on which layer and how much pruning is performed. In this paper, we introduce an optimized pruning technique by improving the existing pruning method, and perform a comparative experiment to analyze the results.

대부분의 스카이라인 질의에 대한 이전 연구들은 대상 객체의 정적 속성만을 고려하였다. 최근에는 모바일 응용 환경의 발전에 따라 이동 객체에 대한 연속적인 스카이라인 질의에 대한 필요성이 증대되고 있다. 연속적인 스카이라인 질의를 처리하기 위해 최근에 몇 가지 기법들이 제안되었지만, 이 기법들은 사용자가 관심을 가지는 일부 속성을 임의로 선택하는 부분차원 스카이라인 질의에 대해서는 고려하지 않았다. 이로 인하여 이동 객체와 부분차원을 동시에 고려해야 하는 모바일 응용에 있어서는 이전 연구들을 적용할 수 없다. 이 논문에서는 질의 시점에 이동 객체의 부분차원 스카이라인을 효율적으로 계산하기 위한 지배 객체 기반 가지치기 기법을 제안한다. 그리고 제안한 기법의 효율성을 증명하기 위해 모의실험을 통한 성능 평가를 수행한다.

Most of previous works for skyline queries have focused only on static attributes of target objects. With the advance in mobile applications, however, the need of continuous skyline queries for moving objects has been increasing. Even though several techniques to process continuous skyline queries have been proposed recently, they cannot process subspace queries, which use only the subset of attribute dimensions. Therefore it is not feasible to utilize those methods for mobile applications which must consider moving objects and subspaces simultaneously. In this paper, we propose a dominant object-based pruning method to compute subspace skyline of moving objects efficiently at query time and present the experimental results to show the effectiveness of the proposed method.

본 논문은 HEVC (high efficiency video coding)의 복잡도 감소를 위한 고속 부호화 방법을 제안한다. 제안 방법은 마지막 0이 아닌 DCT 계수의 위치를 기준으로 블록의 평활성을 판단하고, 이에 따라 Transform Unit의 분할 여부를 빠르게 결정한다. 실험결과를 통해 미미한 0.5% BD-RATE 증가만으로 인코딩 타임을 약 20% 감소시킬 수 있음을 보인다.

In this paper, we show how to design based on solving Sudoku problem that is one of the NP-complete problems like Go. We show how to use greedy method which can minimize depth based on tree expansion and how to apply heuristic algorithm for pruning unnecessary branches. As a result of measuring the performance of the proposed method for solving of Sudoku problems, this method can reduce the number of function call required for solving compared with the method of heuristic algorithm or recursive method, also this method is able to reduce the 46~64 depth rather than simply expanding the tree and is able to pruning unnecessary branches. Therefore, we could see that it can reduce the number of leaf nodes required for the calculation to 6 to 34.

주어진 문제에 대하여 최적의 성능을 가지는 신경회로망을 얻기 위해서는 학습을 통한 매개변수의 최적화 (parameter optimization)와 모델 선택을 통한 구조 최적화(structure optimization )의 통합적인 과정이 필요하다. 본 논문에서는, 각 세부 방법들의 특성을 고려하여, 공통의 특성을 갖는 방법들을 결합함으로써 효율적이면서도 일반화 성능을 높이는 총체적인 신경회로망 최적화 방법을 제안한다. 먼저 다양한 오차 함수를 사용할 수 있는 자연 기울기 강하 학습에 적응적 정규화 방법을 도입함으로써 가중치 매개변수(weight parameter)들을 최적화한다. 그리고 이렇게 최적화된 매개변수(parameter)들에 자연 프루닝(natural pruning)을 적용하여 불필요한 요소들을 제저하여 최적화 된 구조를 생성한다. 반복적인 과정에 의하여 후보 모델들을 구성하고 베이시안 정보 기준(Bayesian Information Criterion: BIC )을 이 용하여 최적의 모델을 평가하여 선택하는 방법을 제안하였다. 벤치마크 데이터에 대한 실험을 통하여 제안하는 방법의 구조 최적화 능력과 일반화 성능의 우수성을 보였다.

To achieve an optimal performance for a given problem, we need an integrative process of the parameter optimization via learning and the structure optimization via model selection. In this paper, we propose an efficient optimization method for improving generalization performance by considering the property of each sub-method and by combining them with common theoretical properties. First, weight parameters are optimized by natural gradient teaming with adaptive regularization, which uses a diverse error function. Second, the network structure is optimized by eliminating unnecessary parameters with natural pruning. Through iterating these processes, candidate models are constructed and evaluated based on the Bayesian Information Criterion so that an optimal one is finally selected. Through computational experiments on benchmark problems, we confirm the weight parameter and structure optimization performance of the proposed method.