• Journal of information and communication convergence engineering
• /
• 제19권4호
• /
• pp.241-247
• /
• 2021

Particulate matter has emerged as a serious global problem, necessitating highly reliable information on the matter. Therefore, various algorithms have been used in studies to predict particulate matter. In this study, we compared the prediction performance of neural network models that have been actively studied for particulate matter prediction. Among the neural network algorithms, a deep neural network (DNN), a recurrent neural network, and long short-term memory were used to design the optimal prediction model using a hyper-parameter search. In the comparative analysis of the prediction performance of each model, the DNN model showed a lower root mean square error (RMSE) than the other algorithms in the performance comparison using the RMSE and the level of accuracy as metrics for evaluation. The stability of the recurrent neural network was slightly lower than that of the other algorithms, although the accuracy was higher.

• 로봇학회논문지
• /
• 제17권2호
• /
• pp.133-141
• /
• 2022

This paper introduces model compression algorithms which make a deep neural network smaller and faster for embedded systems. The model compression algorithms can be largely categorized into pruning, quantization and knowledge distillation. In this study, gradual pruning, quantization aware training, and knowledge distillation which learns the activation boundary in the hidden layer of the teacher neural network are integrated. As a large deep neural network is compressed and accelerated by these algorithms, embedded computing boards can run the deep neural network much faster with less memory usage while preserving the reasonable accuracy. To evaluate the performance of the compressed neural networks, we evaluate the size, latency and accuracy of the deep neural network, DenseNet201, for image classification with CIFAR-10 dataset on the NVIDIA Jetson Xavier.

• 제어로봇시스템학회논문지
• /
• 제9권1호
• /
• pp.57-62
• /
• 2003

It is difficult to determine the feeding rate of coagulant in the water treatment plant, due to nonlinearity, multivariables and slow response characteristics etc. To deal with this difficulty, the genetic-fuzzy system genetic-equation system and the neural network system were used in determining the feeding rate of the coagulant. Fuzzy system and neural network system are excellently robust in multivariables and nonlinear problems. but fuzzy system is difficult to construct the fuzzy parameter such as the rule table and the membership function. Therefore we made the genetic-fuzzy system by the fusion of genetic algorithms and fuzzy system, and also made the feeding rate equation by genetic algorithms. To train fuzzy system, equation parameter and neural network system, the actual operation data of the water treatment plant was used. We determined optimized feeding rates of coagulant by the fuzzy system, the equation and the neural network and also compared them with the feeding rates of the actual operation data.

• 대한전기학회논문지
• /
• 제43권5호
• /
• pp.838-847
• /
• 1994

Although there is an analytical proof of modeling capability of the neural network, the convergency error in nonlinearity modeling is inevitable, since the steepest descent based practical larning algorithms do not guarantee the convergency of modeling error. Therefore, it is difficult to apply the neural network to control system in critical environments under an on-line learning scheme. Although the convergency of modeling error of a neural network is not guatranteed in the practical learning algorithms, the convergency, or boundedness of tracking error of the control system can be achieved if a proper feedback control law is combined with the neural network model to solve the problem of modeling error. In this paper, the neural network is introduced for compensating a system uncertainty to control a nonlinear dynamic system. And for suppressing inevitable modeling error of the neural network, an iterative neural network learning control algorithm is proposed as a virtual on-line realization of the Adaptive Variable Structure Controller. The efficiency of the proposed control scheme is verified from computer simulation on dynamics control of a 2 link robot manipulator.

• 대한전기학회논문지:시스템및제어부문D
• /
• 제53권12호
• /
• pp.804-810
• /
• 2004

There are two important things for the general purpose neural network processor. The first is a capability to build various structures of neural network, and the second is to be able to support suitable learning method for that neural network. Some way to process various learning algorithms is required for on-chip learning, because the more neural network types are to be handled, the more learning methods need to be built into. In this paper, an improved hardware structure is proposed to compute various kinds of learning algorithms flexibly. The hardware structure is based on the existing modular neural network structure. It doesn't need to add a new circuit or a new program for the learning process. It is shown that rearrangements of the existing processing elements can produce several neural network learning modules. The performance and utilization of this module are analyzed by comparing with other neural network chips.

• 전자통신동향분석
• /
• 제38권5호
• /
• pp.12-22
• /
• 2023

The development of neural networks is evolving towards the adoption of transformer structures with attention modules. Hence, active research focused on extending the concept of lightweight neural network algorithms and hardware acceleration is being conducted for the transition from conventional convolutional neural networks to transformer-based networks. We present a survey of state-of-the-art research on lightweight neural network algorithms and hardware architectures to reduce memory usage and accelerate both inference and training. To describe the corresponding trends, we review recent studies on token pruning, quantization, and architecture tuning for the vision transformer. In addition, we present a hardware architecture that incorporates lightweight algorithms into artificial intelligence processors to accelerate processing.

• 한국콘텐츠학회논문지
• /
• 제7권1호
• /
• pp.65-72
• /
• 2007

본 연구에서는 초음파 검사를 기반으로 하는 비파괴검사 방법을 사용하였으며, 용접결함의 패턴인식 알고리즘으로서 역전파 신경망과 확률 신경망을 비교하였다. 이러한 목적을 위한 과정에서 두 가지 알고리즘에 동일한 변수를 적용하였으며, 여기서 사용된 특징변수는 용접결함으로부터 반사된 시간영역 상의 전체 결함신호로부터 결함부분만을 분리한 신호파형을 사용하였다. 이상의 절차를 통하여 두 가지 알고리즘의 적용방안을 확인하였으며, 두 가지 알고리즘에 대하여 각각의 장단점을 비교하였다.

• 한국공작기계학회:학술대회논문집
• /
• 한국공작기계학회 2003년도 추계학술대회
• /
• pp.271-276
• /
• 2003

In this study, we compared backpropagation neural network(BPNN) with probabilistic neural network(PNN) as shape recognition algorithm of welding flaws. For this purpose, variables are applied the same to two algorithm. Here, feature variable is composed of time domain signal itself and frequency domain signal itself, Through this process, we comfirmed advantages/disadvantages of two algorithms and identified application methods of two algorithms.

• 한국품질경영학회:학술대회논문집
• /
• 한국품질경영학회 1998년도 The 12th Asia Quality Management Symposium* Total Quality Management for Restoring Competitiveness
• /
• pp.696-704
• /
• 1998

Artificial neural networks (ANN) have successfully applied into various areas. But, How to effectively established network is the one of the critical problem. This study will focus on this problem and try to extensively study. Firstly, four different learning algorithms ANNs were constructed. The learning algorithms include backpropagation, simulated annealing, genetic algorithm, and tabu search. The experimental results of the above four different learning algorithms were tested by statistical analysis. The training RMS, training time, and testing RMS were used as the comparison criteria.

• 한국공작기계학회:학술대회논문집
• /
• 한국공작기계학회 2004년도 춘계학술대회 논문집
• /
• pp.287-292
• /
• 2004

In this study, we made a comparative study of backpropagation neural network and probabilistic neural network and bayesian classifier and perceptron as shape recognition algorithm of welding flaws. For this purpose, variables are applied the same to four algorithms. Here, feature variable is composed of time domain signal itself and frequency domain signal itself, Through this process, we confirmed advantages/disadvantages of four algorithms and identified application methods of few algorithms.