• International journal of advanced smart convergence
• /
• v.13 no.2
• /
• pp.80-87
• /
• 2024

This paper focuses on improving accuracy in constrained computing settings by employing the ReLU (Rectified Linear Unit) activation function. The research conducted involves modifying parameters of the ReLU function and comparing performance in terms of accuracy and computational time. This paper specifically focuses on optimizing ReLU in the context of a Multilayer Perceptron (MLP) by determining the ideal values for features such as the dimensions of the linear layers and the learning rate (Ir). In order to optimize performance, the paper experiments with adjusting parameters like the size dimensions of linear layers and Ir values to induce the best performance outcomes. The experimental results show that using ReLU alone yielded the highest accuracy of 96.7% when the dimension sizes were 30 - 10 and the Ir value was 1. When combining ReLU with the Adam optimizer, the optimal model configuration had dimension sizes of 60 - 40 - 10, and an Ir value of 0.001, which resulted in the highest accuracy of 97.07%.

• Journal of IKEEE
• /
• v.27 no.4
• /
• pp.531-536
• /
• 2023

This paper proposes a sliding window method with frame feature insertion for immediate attack detection on in-vehicle networks. This method guarantees real-time attack detection by labeling based on the attack status of the current frame. Experiments show that the proposed method improves detection performance by giving more weight to the current frame in CNN computation. The proposed model was designed based on a lightweight LeNet-5 architecture and it achieves 100% detection for DoS attacks. Additionally, by comparing the complexity with conventional models, the proposed model has been proven to be more suitable for resource-constrained devices like ECUs.

• Journal of KIISE:Software and Applications
• /
• v.27 no.5
• /
• pp.463-472
• /
• 2000

In this paper, we consider the problem of realizing associative memories via cellular neural network(CNN). After introducing qualitative properties of the CNN model, we formulate the synthesis of CNN that can store given binary vectors with optimal performance as a constrained optimization problem. Next, we observe that this problem's constraints can be transformed into simple inequalities involving linear matrix inequalities(LMIs). Finally, we reformulate the synthesis problem as a generalized eigenvalue problem(GEVP), which can be efficiently solved by recently developed interior point methods. Proposed method can be applied to both space varying template CNNs and space-invariant template CNNs. The validity of the proposed approach is illustrated by design examples.

• Journal of the Korea Concrete Institute
• /
• v.20 no.5
• /
• pp.627-634
• /
• 2008

This paper presents an optimum mixture design method for proportioning a concrete. In the proposed method, the search space is constrained as the domain defined by the minimal convex region of a database, instead of the available range of each component and the ratio composed of several components. The model for defining the search space which is expressed by the effective region is proposed. The effective region model evaluates whether a mix-proportion is effective on processing for optimization, yielding highly reliable results. Three concepts are adopted to realize the proposed methodology: A genetic algorithm for the optimization; an artificial neural network for predicting material properties; and a convex hull for evaluating the effective region. And then, it was applied to an optimization problem wherein the minimum cost should be obtained under a given strength requirement. Experimental test results show that the mix-proportion obtained from the proposed methodology using convex hulls is found to be more accurate and feasible than that obtained from a general optimum technique that does not consider this aspect.

• Smart Media Journal
• /
• v.11 no.3
• /
• pp.18-25
• /
• 2022

Neuromorphic computing generally shows significantly better power, area, and speed performance than neural network computation using CPU and GPU. These characteristics are suitable for resource-constrained IoT environments where energy consumption is important. However, there is a problem in that it is necessary to modify the source code for environment setting and application operation according to heterogeneous IoT devices that support neuromorphic computing. To solve these problems, NAAL was proposed and implemented in this paper. NAAL provides functions necessary for IoT device control and neuromorphic architecture abstraction and inference model operation in various heterogeneous IoT device environments based on common APIs of NAAL. NAAL has the advantage of enabling additional support for new heterogeneous IoT devices and neuromorphic architectures and computing devices in the future.

• The KIPS Transactions:PartB
• /
• v.13B no.4 s.107
• /
• pp.429-438
• /
• 2006

Neural Networks(NNs) are applied for solving a wide variety of nonlinear problems in several areas, such as image processing, pattern recognition etc. Although NN can be simulated by using software, many potential NN applications required real-time processing. Thus they need to be implemented as hardware. The hardware implementation of multi-layer perceptrons(MLPs) in several kind of NNs usually uses a fixed-point arithmetic due to a simple logic operation and a shorter processing time compared to the floating-point arithmetic. However, the fixed-point arithmetic-based MLP has a drawback which is not able to apply the MLP software that use floating-point arithmetic. We propose a design method for MLPs which has the floating-point arithmetic-based fully-pipelining architecture. It has a processing speed that is proportional to the number of the hidden nodes. The number of input and output nodes of MLPs are generally constrained by given problems, but the number of hidden nodes can be optimized by user experiences. Thus our design method is using optimized number of hidden nodes in order to improve the processing speed, especially in field of a repeated processing such as image processing, pattern recognition, etc.

• KIPS Transactions on Software and Data Engineering
• /
• v.9 no.5
• /
• pp.169-176
• /
• 2020

Recently, deep neural networks (DNNs) have brought revolutions to many mobile and embedded devices by providing human-level machine intelligence for various applications. However, high inference accuracy of such DNNs comes at high computational costs, and, hence, there have been significant efforts to reduce computational overheads of DNNs either by compressing off-the-shelf models or by designing a new small footprint DNN architecture tailored to resource constrained devices. One notable recent paradigm in designing small footprint DNN models is sharing parameters in several layers. However, in previous approaches, the parameter-sharing techniques have been applied to large deep networks, such as ResNet, that are known to have high redundancy. In this paper, we propose a parameter-sharing method for already parameter-efficient small networks such as ShuffleNetV2. In our approach, small templates are combined with small layer-specific parameters to generate weights. Our experiment results on ImageNet and CIFAR100 datasets show that our approach can reduce the size of parameters by 15%-35% of ShuffleNetV2 while achieving smaller drops in accuracies compared to previous parameter-sharing and pruning approaches. We further show that the proposed approach is efficient in terms of latency and energy consumption on modern embedded devices.

• The Sea:JOURNAL OF THE KOREAN SOCIETY OF OCEANOGRAPHY
• /
• v.21 no.1
• /
• pp.1-10
• /
• 2016

Currently available surface seawater partial pressure carbon dioxide ($pCO_2$) data sets in the East Sea are not enough to quantify statistically the carbon dioxide flux through the air-sea interface. To complement the scarcity of the $pCO_2$ measurements, we construct a neural network (NN) model based on satellite data to map $pCO_2$ for the areas, which were not observed. The NN model is constructed for the Ulleung Basin, where $pCO_2$ data are best available, to map and estimate the variability of $pCO_2$ based on in situ $pCO_2$ for the years from 2003 to 2012, and the sea surface temperature (SST) and chlorophyll data from the MODIS (Moderate-resolution Imaging Spectroradiometer) sensor of the Aqua satellite along with geographic information. The NN model was trained to achieve higher than 95% of a correlation between in situ and predicted $pCO_2$ values. The RMSE (root mean square error) of the NN model output was $19.2{\mu}atm$ and much less than the variability of in situ $pCO_2$. The variability of $pCO_2$ with respect to SST and chlorophyll shows a strong negative correlation with SST than chlorophyll. As SST decreases the variability of $pCO_2$ increases. When SST is lower than $15^{\circ}C$, $pCO_2$ variability is clearly affected by both SST and chlorophyll. In contrast when SST is higher than $15^{\circ}C$, the variability of $pCO_2$ is less sensitive to changes in SST and chlorophyll. The mean rate of the annual $pCO_2$ increase estimated by the NN model output in the Ulleung Basin is $0.8{\mu}atm\;yr^{-1}$ from 2003 to 2014. As NN model can successfully map $pCO_2$ data for the whole study area with a higher resolution and less RMSE compared to the previous studies, the NN model can be a potentially useful tool for the understanding of the carbon cycle in the East Sea, where accessibility is limited by the international affairs.