• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2020.07a
• /
• pp.22-25
• /
• 2020

본 논문에서는 큰 크기의 심층 신경망을 압축하기위해 네트워크 수준의 가중치 공유방법인 Global Weight 패러다임을 최초로 제시한다. 기존의 가중치 공유방법은 계층별로 가중치를 공유하는 것이 대부분이었다. Global Weight 는 기존 방법과 달리 전체 네트워크에서 가중치를 공유하는 효율적인 방법이다. 우리는 Global Weight 를 사용하여 학습되는 새로운 컨볼루션 연산인 Global Weight Convolution(GWConv)연산과 GWConv를 적용한 Global Weight Networks(GWNet)을 제안한다. CIFAR10 데이터셋에서 실험한 결과 2.18 배 압축에서 85.64%, 3.41 배 압축에서 85.46%의 정확도를 보였다. Global Weight 패러다임은 가중치 공유가 궁극적으로 풀고자 했던 중복되는 가중치를 최소화하는 획기적인 방법이며, 추후 심도 있는 연구가 수행될 수 있음을 시사한다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.580-581
• /
• 2024

고령화 사회에 접어들면서 황반 변성과 당뇨 망막 병증 등 시야결손을 동반하는 안구 질환의 발병률은 증가하지만 이러한 질환의 조기 발견에 인공지능을 접목시킨 연구는 부족한 실정이다. 본 논문은 안구 질환 자가 검사용 인공 신경망을 학습시키기 위한 데이터 베이스 구축 방법을 제안한다. MNIST와 CIFAR-10을 합성하여 중첩 이미지 데이터셋인 G-Dataset을 생성하였고, 7개의 인공신경망에 학습시켜 최종적으로 90% 이상의 정확도를 얻음으로 그 유효성을 입증하였다. G-Dataset을 안구 질환 자가 검사용 딥러닝 모델에 학습시켜 모바일 어플에 적용하면 사용자가 주기적인 검사를 통해 안구 질환을 조기에 진단하고 치료할 수 있을 것으로 기대된다.

• The Korean Journal of Applied Statistics
• /
• v.32 no.5
• /
• pp.693-702
• /
• 2019

Deep learning has gained popularity for the classification and prediction task. Neural network layers become deeper as more data becomes available. Saturation is the phenomenon that the gradient of an activation function gets closer to 0 and can happen when the value of weight is too big. Increased importance has been placed on the issue of saturation which limits the ability of weight to learn. To resolve this problem, Glorot and Bengio (Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, 249-256, 2010) claimed that efficient neural network training is possible when data flows variously between layers. They argued that variance over the output of each layer and variance over input of each layer are equal. They proposed a method of initialization that the variance of the output of each layer and the variance of the input should be the same. In this paper, we propose a new method of establishing initialization by adopting truncated normal distribution and truncated cauchy distribution. We decide where to truncate the distribution while adapting the initialization method by Glorot and Bengio (2010). Variances are made over output and input equal that are then accomplished by setting variances equal to the variance of truncated distribution. It manipulates the distribution so that the initial values of weights would not grow so large and with values that simultaneously get close to zero. To compare the performance of our proposed method with existing methods, we conducted experiments on MNIST and CIFAR-10 data using DNN and CNN. Our proposed method outperformed existing methods in terms of accuracy.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.31 no.6
• /
• pp.1215-1225
• /
• 2021

Although Artificial Intelligence (AI) techniques have shown impressive performance in various fields, they are vulnerable to adversarial examples which induce misclassification by adding human-imperceptible perturbations to the input. Previous studies to defend the adversarial examples can be classified into three categories: (1) model retraining methods; (2) input transformation methods; and (3) adversarial examples detection methods. However, even though the defense methods against adversarial examples have constantly been proposed, there is no research to classify the type of adversarial attack. In this paper, we proposed an adversarial attack family classification method based on dimensionality reduction and clustering. Specifically, after extracting adversarial perturbation from adversarial example, we performed Linear Discriminant Analysis (LDA) to reduce the dimensionality of adversarial perturbation and performed K-means algorithm to classify the type of adversarial attack family. From the experimental results using MNIST dataset and CIFAR-10 dataset, we show that the proposed method can efficiently classify five tyeps of adversarial attack(FGSM, BIM, PGD, DeepFool, C&W). We also show that the proposed method provides good classification performance even in a situation where the legitimate input to the adversarial example is unknown.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.12
• /
• pp.519-524
• /
• 2023

In hashing-based image retrieval, the hash code of a manipulated image is different from the original image, making it difficult to search for the same image. This paper proposes and evaluates a self-supervised deephashing model that generates perceptual hash codes from feature information such as texture, shape, and color of images. The comparison models are autoencoder-based variational inference models, but the encoder is designed with a fully connected layer, convolutional neural network, and transformer modules. The proposed model is a variational inference model that includes a SimAM module of extracting geometric patterns and positional relationships within images. The SimAM module can learn latent vectors highlighting objects or local regions through an energy function using the activation values of neurons and surrounding neurons. The proposed method is a representation learning model that can generate low-dimensional latent vectors from high-dimensional input images, and the latent vectors are binarized into distinguishable hash code. From the experimental results on public datasets such as CIFAR-10, ImageNet, and NUS-WIDE, the proposed model is superior to the comparative model and analyzed to have equivalent performance to the supervised learning-based deephashing model. The proposed model can be used in application systems that require low-dimensional representation of images, such as image search or copyright image determination.

• International Journal of Internet, Broadcasting and Communication
• /
• v.11 no.4
• /
• pp.31-36
• /
• 2019

Unsupervised neural networks have not caught enough attention until Generative Adversarial Network (GAN) was proposed. By using both the generator and discriminator networks, GAN can extract the main characteristic of the original dataset and produce new data with similarlatent statistics. However, researchers understand fully that training GAN is not easy because of its unstable condition. The discriminator usually performs too good when helping the generator to learn statistics of the training datasets. Thus, the generated data is not compelling. Various research have focused on how to improve the stability and classification accuracy of GAN. However, few studies delve into how to improve the training efficiency and to save training time. In this paper, we propose a novel optimizer, named FAST-ADAM, which integrates the Lookahead to ADAM optimizer to train the generator of a semi-supervised generative adversarial network (SSGAN). We experiment to assess the feasibility and performance of our optimizer using Canadian Institute For Advanced Research - 10 (CIFAR-10) benchmark dataset. From the experiment results, we show that FAST-ADAM can help the generator to reach convergence faster than the original ADAM while maintaining comparable training accuracy results.

• Journal of Multimedia Information System
• /
• v.9 no.1
• /
• pp.1-10
• /
• 2022

China is a big country in animal fur industry. The total production and consumption of fur are increasing year by year. However, the recognition of fur in the fur production process still mainly relies on the visual identification of skilled workers, and the stability and consistency of products cannot be guaranteed. In response to this problem, this paper proposes a feature fusion-based animal fur recognition network on the basis of typical convolutional neural network structure, relying on rapidly developing deep learning techniques. This network superimposes texture feature - the most prominent feature of fur image - into the channel dimension of input image. The output feature map of the first layer convolution is inverted to obtain the inverted feature map and concat it into the original output feature map, then Leaky ReLU is used for activation, which makes full use of the texture information of fur image and the inverted feature information. Experimental results show that the algorithm improves the recognition accuracy by 9.08% on Fur_Recognition dataset and 6.41% on CIFAR-10 dataset. The algorithm in this paper can change the current situation that fur recognition relies on manual visual method to classify, and can lay foundation for improving the efficiency of fur production technology.

• 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.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.7
• /
• pp.11-20
• /
• 2024

In this paper, we propose the BAG (Bit-width Aware Generator) and the Intermediate Layer Knowledge Distillation using Channel-wise Attention to reduce the knowledge gap between a quantized network, a full-precision network, and a generator in GDFQ (Generative Data-Free Quantization). Since the generator in GDFQ is only trained by the feedback from the full-precision network, the gap resulting in decreased capability due to low bit-width of the quantized network has no effect on training the generator. To alleviate this problem, BAG is quantized with same bit-width of the quantized network, and it can generate synthetic images, which are effectively used for training the quantized network. Typically, the knowledge gap between the quantized network and the full-precision network is also important. To resolve this, we compute channel-wise attention of outputs of convolutional layers, and minimize the loss function as the distance of them. As the result, the quantized network can learn which channels to focus on more from mimicking the full-precision network. To prove the efficiency of proposed methods, we quantize the network trained on CIFAR-100 with 3 bit-width weights and activations, and train it and the generator with our method. As the result, we achieve 56.14% Top-1 Accuracy and increase 3.4% higher accuracy compared to our baseline AdaDFQ.

• Journal of Korea Multimedia Society
• /
• v.23 no.7
• /
• pp.803-811
• /
• 2020

This paper proposes an image classification algorithm that transforms the number of convolution layers in the residual block of ResNet, CNN's representative method. The proposed method modified the structure of 34/50 layer of ResNet structure. First, we analyzed the performance of small and many convolution layers for the structure consisting of only shortcut and 3 × 3 convolution layers for 34 and 50 layers. And then the performance was analyzed in the case of small and many cases of convolutional layers for the bottleneck structure of 50 layers. By applying the results, the best classification method in the residual block was applied to construct a 34-layer simple structure and a 50-layer bottleneck image classification model. To evaluate the performance of the proposed image classification model, the results were analyzed by applying to the cifar10 dataset. The proposed 34-layer simple structure and 50-layer bottleneck showed improved performance over the ResNet-110 and Densnet-40 models.