• Journal of Multimedia Information System
• /
• v.9 no.3
• /
• pp.177-182
• /
• 2022

A deep neural network (DNN) includes variables whose values keep on changing with the training process until it reaches the final point of convergence. These variables are the co-efficient of a polynomial expression to relate to the feature extraction process. In general, DNNs work in multiple 'dimensions' depending upon the number of channels and batches accounted for training. However, after the execution of feature extraction and before entering the SoftMax or other classifier, there is a conversion of features from multiple N-dimensions to a single vector form, where 'N' represents the number of activation channels. This usually happens in a Fully connected layer (FCL) or a dense layer. This reduced 2D feature is the subject of study for our analysis. For this, we have used the FCL, so the trained weights of this FCL will be used for the weight-class correlation analysis. The popular DNN models selected for our study are ResNet-101, VGG-19, and GoogleNet. These models' weights are directly used for fine-tuning (with all trained weights initially transferred) and scratch trained (with no weights transferred). Then the comparison is done by plotting the graph of feature distribution and the final FCL weights.

• Journal of Internet of Things and Convergence
• /
• v.9 no.3
• /
• pp.61-67
• /
• 2023

In order to optimize the pulse electroforming copper process, a double hidden layer BP (Back Propagation) neural network is constructed. Through sample training, the mapping relationship between electroforming copper process conditions and target properties is accurately established, and the prediction of microhardness and tensile strength of the electroforming layer in the pulse electroforming copper process is realized. The predicted results are verified by electrodeposition copper test in copper pyrophosphate solution system with pulse power supply. The results show that the microhardness and tensile strength of copper layer predicted by "3-4-3-2" structure double hidden layer neural network are very close to the experimental values, and the relative error is less than 2.32%. In the parameter range, the microhardness of copper layer is between 100.3~205.6MPa and the tensile strength is between 112~485MPa.When the microhardness and tensile strength are optimal,the corresponding process conditions are as follows: current density is 2A-dm-2, pulse frequency is 2KHz and pulse duty cycle is 10%.

• Computers and Concrete
• /
• v.15 no.5
• /
• pp.759-770
• /
• 2015

In this study, a multi-layer perceptron neural network (MLPNN) prediction model for compressive strength of the cement mortars has been developed. For purpose of constructing this model, 8 different mixes with 240 specimens of the 2, 7, 28, 56 and 90 days compressive strength experimental results of cement mortars containing fly ash (FA), silica fume (SF) and FA+SF used in training and testing for MLPNN system was gathered from the standard cement tests. The data used in the MLPNN model are arranged in a format of four input parameters that cover the FA, SF, FA+SF and age of samples and an output parameter which is compressive strength of cement mortars. In the model, the training and testing results have shown that MLPNN system has strong potential as a feasible tool for predicting 2, 7, 28, 56 and 90 days compressive strength of cement mortars.

• The Korean Journal of Applied Statistics
• /
• v.26 no.5
• /
• pp.847-856
• /
• 2013

We propose a modified learning process for generalized neural network using a learning algorithm by Liu et al. (2001). We consider the effect of initial weights, training results and learning errors using a modified learning process. We employ an incremental training procedure where training patterns are learned systematically. Our algorithm starts with a single training pattern and a single hidden layer neuron. During the course of neural network training, we try to escape from the local minimum by using a weight scaling technique. We allow the network to grow by adding a hidden layer neuron only after several consecutive failed attempts to escape from a local minimum. Our optimization procedure tends to make the network reach the error tolerance with no or little training after the addition of a hidden layer neuron. Simulation results with suitable initial weights indicate that the present constructive algorithm can obtain neural networks very close to minimal structures and that convergence to a solution in neural network training can be guaranteed. We tested these algorithms extensively with small training sets.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.9
• /
• pp.1158-1165
• /
• 2021

This paper proposes a CNN accelerator which is optimized Pooling layer operation incorporated in Multiplication And Accumulation(MAC) to reduce the memory size. For optimizing memory and data path circuit, the quantized 8bit integer weights are used instead of 32bit floating-point weights for pre-training of MNIST data set. To reduce chip area, the proposed CNN model is reduced by a convolutional layer, a 4*4 Max Pooling, and two fully connected layers. And all the operations use specific MAC with approximation adders and multipliers. 94% of internal memory size reduction is achieved by simultaneously performing the convolution and the pooling operation in the proposed architecture. The proposed accelerator chip is designed by using TSMC65nmGP CMOS process. That has about half size of our previous paper, 0.8*0.9 = 0.72mm². The presented CNN accelerator chip achieves 94% accuracy and 77us inference time per an MNIST image.

• ETRI Journal
• /
• v.17 no.1
• /
• pp.11-22
• /
• 1995

This paper proposes a modified error function to improve the error back-propagation (EBP) algorithm for multi-Layer perceptrons (MLPs) which suffers from slow learning speed. It can also suppress over-specialization for training patterns that occurs in an algorithm based on a cross-entropy cost function which markedly reduces learning time. In the similar way as the cross-entropy function, our new function accelerates the learning speed of the EBP algorithm by allowing the output node of the MLP to generate a strong error signal when the output node is far from the desired value. Moreover, it prevents the overspecialization of learning for training patterns by letting the output node, whose value is close to the desired value, generate a weak error signal. In a simulation study to classify handwritten digits in the CEDAR [1] database, the proposed method attained 100% correct classification for the training patterns after only 50 sweeps of learning, while the original EBP attained only 98.8% after 500 sweeps. Also, our method shows mean-squared error of 0.627 for the test patterns, which is superior to the error 0.667 in the cross-entropy method. These results demonstrate that our new method excels others in learning speed as well as in generalization.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.9
• /
• pp.57-64
• /
• 2021

In this paper, we propose a device that can control the primary control surface of an aircraft by recognizing speech commands. The speech command consists of 19 commands, and a learning model is constructed based on a total of 2,500 datasets. The training model is composed of a CNN model using the Sequential library of the TensorFlow-based Keras model, and the speech file used for training uses the MFCC algorithm to extract features. The learning model consists of two convolution layers for feature recognition and Fully Connected Layer for classification consists of two dense layers. The accuracy of the validation dataset was 98.4%, and the performance evaluation of the test dataset showed an accuracy of 97.6%. In addition, it was confirmed that the operation was performed normally by designing and implementing a Raspberry Pi-based control device. In the future, it can be used as a virtual training environment in the field of voice recognition automatic flight and aviation maintenance.

• International journal of advanced smart convergence
• /
• v.7 no.4
• /
• pp.57-65
• /
• 2018

Artificial neural network is inspired by the biological neural network. For simplicity, in computer science, it is represented as a set of layers. Many research has been made in evaluating the number of neurons in the hidden layer but still, none was accurate. Several methods are used until now which do not provide the exact formula for calculating the number of thehidden layer as well as the number of neurons in each hidden layer. In this paper model selection approach was presented. Proposed model is based on geographical analysis of decision boundary. Proposed model selection method is useful when we know the distribution of the training data set. To evaluate the performance of the proposed method we compare it to the traditional architecture on IRIS classification problem. According to the experimental result on Iris data proposed method is turned out to be a powerful one.

• Journal of Korea Multimedia Society
• /
• v.25 no.6
• /
• pp.783-792
• /
• 2022

Versatile Video Coding (VVC) is the latest video coding standard, which had been developed by the Joint Video Experts Team (JVET) of ITU-T Video Coding Experts Group (VCEG) and ISO/IEC Moving Picture Experts Group (MPEG) in 2020. Although VVC can provide powerful coding performance, it requires tremendous computational complexity to determine the optimal block structures during the encoding process. In this paper, we propose a fast ternary tree decision method using two neural networks with 7 nodes as input vector based on the multi-layer perceptron structure, names STH-NN and STV-NN. As a training result of neural network, the STH-NN and STV-NN achieved accuracies of 85% and 91%, respectively. Experimental results show that the proposed method reduces the encoding complexity up to 25% with unnoticeable coding loss compared to the VVC test model (VTM).

• Magazine of the Korean Society of Agricultural Engineers
• /
• v.36 no.2
• /
• pp.88-95
• /
• 1994

This paper is to evaluate the applicability of neural networks to the estimation of evapotranspiration. Two neural networks were developed to forecast daily evapotranspiration of the rice crop with back-propagation and counter-propagation algorithm. The neural network trained by back-propagation algorithm with delta learning rule is a three-layer network with input, hidden, and output layers. The other network with counter-propagation algorithm is a four-layer network with input, normalizing, competitive, and output layers. Training neural networks was conducted using daily actual evapotranspiration of rice crop and daily climatic data such as mean temperature, sunshine hours, solar radiation, relative humidity, and pan evaporation. During the training, neural network parameters were calibrated. The trained networks were applied to a set of field data not used in the training. The created response of the back-propagation network was in good agreement with desired values and showed better performances than the counter-propagation network did. Evaluating the neural network performance indicates that the back-propagation neural network may be applied to the estimation of evapotranspiration of the rice crop. This study does not provide with a conclusive statement as to the ability of a neural network to evapotranspiration estimating. More detailed study is required for better understanding and evaluating the behavior of neural networks.