• Journal of KIISE:Software and Applications
• /
• v.26 no.4
• /
• pp.546-556
• /
• 1999

다층 퍼셉트론은 다양한 응용 분야에 성공적으로 적용되고 있는 대표적인 신경회로망 모델이다. 그러나 다층 퍼셉트론의 학습에 사용되는 오류역전파 알고리즘으로 알려진 기울기 강하 학습법은 느린 수렴속도로 인해 실시간 처리가 요구되거나 시간에 따라 환경이 변하는 문제에의 적용이 불가능하다. 이러한 느린 수렴속도는 기울기 강하법을 사용한 학습과정에서의 오차함수의 기울기 변화가 극히 적어 오차의 감소가 거의 일어나지 않는 부분인 플라토에 기인하는 것으로 알려져있다. 본 논문에서는 정보기하이론의 관점에서 기존의 학습법에 사용되는 기울기의 이론적 문제를 지적하고, 그로부터 플라토 문제의 원인을 밝힌다. 또한 이를 바탕으로 정보기하이론에 의해 새롭게 정의되는 자연 기울기를 이용한 학습법을 제시하고, 이를 이용한 플라토 문제가 문제해결의 가능성을 분석적으로 고찰하고 실험을 통해 확인한다.

• Proceedings of the Korean Society for Information Management Conference
• /
• 2003.08a
• /
• pp.65-72
• /
• 2003

kNN 기법은 높은 자동분류 성능을 보여주지만 처리 속도가 느리다는 단점이 있다. 이를 극복하기 위해 입력문서의 대표용어 w개를 선정하고 이를 포함한 학습문서만으로 학습집단을 축소함으로써 자동분류 속도를 향상시키는 kw_kNN을 제안하였다. 실험 결과 대표 용어를 5개 사용할 경우에는 kNN 대비 문서간 비교횟수를 평균 18.4%로 축소할 수 있었다. 그러면서도 성능저하를 최소화하여 매크로 평균 F1 척도면에서는 차이가 없고 마이크로 평균정확률 면에서는 약 l∼2% 포인트 이내로 kNN 기법의 성능에 근접한 결과를 얻었다.

• The Journal of the Acoustical Society of Korea
• /
• v.20 no.4
• /
• pp.48-53
• /
• 2001

In this paper, selectively attentive learning method has been proposed to improve the learning speed of multilayer Perceptron based on the error backpropagation algorithm. Three attention criterions are introduced to effectively determine which set of input patterns is or which portion of network is attended to for effective learning. Such criterions are based on the mean square error function of the output layer and class-selective relevance of the hidden nodes. The acceleration of learning time is achieved by lowering the computational cost per iteration. Effectiveness of the proposed method is demonstrated in a speaker adaptation task of isolated word recognition system. The experimental results show that the proposed selective attention technique can reduce the learning time more than 60% in an average sense.

• Journal of the Korea Society of Computer and Information
• /
• v.14 no.8
• /
• pp.11-18
• /
• 2009

This paper proposes the way of improving learning speed in Levenberg-Marquardt algorithm using the principal submatrix of Jacobian matrix. The Levenberg-Marquardt learning uses Jacobian matrix for Hessian matrix to get the second derivative of an error function. To make the Jacobian matrix an invertible matrix. the Levenberg-Marquardt learning must increase or decrease ${\mu}$ and recalculate the inverse matrix of the Jacobian matrix due to these changes of ${\mu}$. Therefore, to have the proper ${\mu}$, we create the principal submatrix of Jacobian matrix and set the ${\mu}$ as the eigenvalues sum of the principal submatrix. which can make learning speed improve without calculating an additional inverse matrix. We also showed that our method was able to improve learning speed in both a generalized XOR problem and a handwritten digit recognition problem.

• Journal of the Korean Geotechnical Society
• /
• v.17 no.1
• /
• pp.47-56
• /
• 2001

본 논문에서는 단일 및 군말뚝의 수평변위를 예측하기 위하여 신경망 학습속도의 향상과 지역 최소점 수렴을 방지하는 Readjusting 기법을 적용한 인공신경망을 도입하였다. 이 인공신경망을 M-EBPNN 이라고 한다. M-EBPNN에 의한 결과는 낙동강 모래지반에서 단일 및 군말뚝에 대하여 수행한 일련의 모형실험결과와 비교하였으며, 그리고 신경망의 학습속도와 지역 최소점의 수렴성을 평가하기 위하여 오류 역전파 신경망(EBPNN)의 결과와도 비교 분석하였다. M-EBPNN의 적용성 검증을 위하여 200개의 모형실험결과들을 이용하였으며, 신경망의 구조는 EBPNN의 구조와 동일한 한 개의 입력층과 두 개의 은닉층 그리고 한 개의 출력층으로 구성되었다. 전체 데이터의 25%, 50% 그리고 75% 결과는 각각 신경망의 학습에 이용되었으며 학습에 이용하지 않은 데이터들은 예측에 이용되었다. 그리고, 신경망의 최적학습을 위하여 적합한 은닉층의 뉴런 수와 학습률은 EBPNN에서 결정한 값들을 본 신경망에 이용하였다. 해석결과들에 의하면, 동일한 학습패턴에서의 M-EBPNN이 학습 반복횟수는 EBPNN 보다 최고 88% 감소하였으며 지역 최소점에 수렴하는 현상은 거의 나타나지 않았다. 따라서, 인공신경망 모델이 수평하중을 받는 말뚝의 수평변위 예측에 적용될 수 있는 가능성을 보여 주었다.

• Journal of Internet Computing and Services
• /
• v.11 no.1
• /
• pp.77-84
• /
• 2010

Children, who have a language disorder, have difficulty in expressing their reaction about stimulus of sound and vision. So it is very hard to grasp that they recognize external stimulus or not. For solving these problem, we can check response time and make them to choose stimulus by giving stimulus of sound and vision to them through Audio and Visual Stimulus and Reaction Meter System. Additionally, We can help them by improving response time by repeated study based on the results and making them to recognize and choose stimulus faster without aversion about external stimulus. It would make them not to feel uncomfortable and isolated because they are unfamiliar with external stimulus.

• Journal of KIISE:Software and Applications
• /
• v.34 no.12
• /
• pp.1045-1055
• /
• 2007

Latent topic models are statistical models which automatically captures salient patterns or correlation among features underlying a data collection in a probabilistic way. They are gaining an increased popularity as an effective tool in the application of automatic semantic feature extraction from text corpus, multimedia data analysis including image data, and bioinformatics. Among the important issues for the effectiveness in the application of latent topic models to the massive data set is the efficient learning of the model. The paper proposes an accelerated learning technique for PLSA model, one of the popular latent topic models, by an incremental EM algorithm instead of conventional EM algorithm. The incremental EM algorithm can be characterized by the employment of a series of partial E-steps that are performed on the corresponding subsets of the entire data collection, unlike in the conventional EM algorithm where one batch E-step is done for the whole data set. By the replacement of a single batch E-M step with a series of partial E-steps and M-steps, the inference result for the previous data subset can be directly reflected to the next inference process, which can enhance the learning speed for the entire data set. The algorithm is advantageous also in that it is guaranteed to converge to a local maximum solution and can be easily implemented just with slight modification of the existing algorithm based on the conventional EM. We present the basic application of the incremental EM algorithm to the learning of PLSA and empirically evaluate the acceleration performance with several possible data partitioning methods for the practical application. The experimental results on a real-world news data set show that the proposed approach can accomplish a meaningful enhancement of the convergence rate in the learning of latent topic model. Additionally, we present an interesting result which supports a possible synergistic effect of the combination of incremental EM algorithm with parallel computing.

• Journal of the Korea Society of Computer and Information
• /
• v.7 no.4
• /
• pp.204-210
• /
• 2002

The self-organizing feature map of Kohonen has disadvantage that needs too much input patterns in order to converge into the equilibrium state when it trains. In this paper we proposed the method of improving the convergence speed and rate of self-organizing feature map converting the interaction set into Dynamic Gaussian function. The proposed method Provides us with dynamic Properties that the deviation and width of Gaussian function used as an interaction function are narrowed in proportion to learning times and learning rates that varies according to topological position from the winner neuron. In this Paper. we proposed the method of improving the convergence rate and the degree of self-organizing feature map.

• Journal of Korean Society of Transportation
• /
• v.20 no.1
• /
• pp.131-139
• /
• 2002

This study introduces a developed artificial neural networks(ANN) model as a more efficient and reliable prediction model in operating speed Prediction with the 85th percentile horizontal curve of two-way rural highway in the aspect of evaluating highway design consistency. On the assumption that the speed is decided by highway geometry features, total 30 survey sites were selected. Data include currie radius, curve length, intersection angle, sight distance, lane width, and lane of those sites and were used as input layer data of the ANN. The optimized model structure was drawn by number of unit of hidden layer, learning coefficient, momentum coefficient, and change in learning frequency in multi-layer a ANN model. To verify learning Performance of ANN, 30 survey sites were selected while data in obtained from the 20 cites were used as learning data and those from the remaining 10 sites were used as predictive data. As a result of statistical verification, the model D of 4 types of ANN was evaluated as the most similar model to the actual operating speed value: R2 was 85% and %RMSE was 0.0204.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.10 no.5
• /
• pp.466-477
• /
• 2000

For developing a CMAC-based adaptive critic learning system to control the cart-pole system, various papers including neural network based learning control schemes as well as an adaptive critic learning algorithm with Adaptive Search Element are reviewed and the adaptive critic learning algorithm for the ASE is integrated into a CMAC controller. Also, quantization problems involved in integrating CMAC into ASE system are studied. By comparing the learning speed of the CMAC system with that of the ASE system and by considering the learning genemlization of the CMAC system with the adaptive critic learning, the applicability of the adaptive critic learning algorithm to CMAC is discussed.