• IEMEK Journal of Embedded Systems and Applications
• /
• v.16 no.1
• /
• pp.1-8
• /
• 2021

Detecting Out-of-Distribution (OOD) data is fundamentally required when Deep Neural Network (DNN) is applied to real-world AI such as autonomous driving. However, modern DNNs are quite vulnerable to the over-confidence problem even if the test data are far away from the trained data distribution. To solve the problem, this paper proposes a novel Adversarial-Mixup training method to let the DNN model be more robust by detecting OOD data effectively. Experimental results show that the proposed Adversarial-Mixup method improves the overall performance of OOD detection by 78% comparing with the State-of-the-Art methods. Furthermore, we show that the proposed method can alleviate the over-confidence problem by reducing the confidence score of OOD data than the previous methods, resulting in more reliable and robust DNNs.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2021.11a
• /
• pp.694-696
• /
• 2021

본 논문에서는 CNN 모델에서 학습에 사용할 이미지 데이터를 늘리기 위해 사용되는 Mixup 알고리즘을 MLP 모델에 사용하는 데이터셋에 적용하여 data augmentation 효과를 얻을 수 있는 지에 대한 테스트를 수행했다. 테스트 결과 MLP 모델에 사용할 데이터셋에도 Mixup 알고리즘으로 data augmentation 효과를 기대할 수 있음을 보여준다.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2023.01a
• /
• pp.17-18
• /
• 2023

비지도형 도메인 적응(UDA)에 대한 최근 연구는 도메인 적응에 대한 설명 및 전이 가능한 특징을 풀어 내기 위해 적대적 학습에 의존한다. 그러나 기존 방법에는 대상 도메인의 클래스 인식(class-aware) 정보를 고려하지 않고는 잠재 공간의 구별 가능성을 완전히 보장할 수 없다는 것과 소스 및 대상 도메인의 샘플만으로는 잠재 공간에서 도메인 불변(domain- invariant) 특성을 추출하기에 부족하다는 두 가지 문제가 있다고 알려져 있다. 본 논문에서는 기존 알려진 UDA의 도메인 적응시 발생되는 문제를 해결하기 위해 Adversarial Discriminative Domain Adaptation(ADDA)에서 mixup을 활용해 신경망의 로버스트네스를 향상시키는 것을 확인하였다.

• Journal of information and communication convergence engineering
• /
• v.21 no.2
• /
• pp.139-144
• /
• 2023

Engineers prefer deep neural networks (DNNs) for solving computer vision problems. However, DNNs pose two major problems. First, neural networks require large amounts of well-labeled data for training. Second, the covariate shift problem is common in computer vision problems. Domain adaptation has been proposed to mitigate this problem. Recent work on adversarial-learning-based unsupervised domain adaptation (UDA) has explained transferability and enabled the model to learn robust features. Despite this advantage, current methods do not guarantee the distinguishability of the latent space unless they consider class-aware information of the target domain. Furthermore, source and target examples alone cannot efficiently extract domain-invariant features from the encoded spaces. To alleviate the problems of existing UDA methods, we propose the mixup regularization in adversarial discriminative domain adaptation (ADDA) method. We validated the effectiveness and generality of the proposed method by performing experiments under three adaptation scenarios: MNIST to USPS, SVHN to MNIST, and MNIST to MNIST-M.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.3
• /
• pp.355-363
• /
• 2020

In this paper, we propose a deep learning training method for accurately classifying Kuzushiji-MNIST and Kuzushiji-49 datasets for ancient and medieval Japanese characters. We analyze the latest convolutional neural network networks through experiments to select the most suitable network, and then use the networks to select the number of training to classify Kuzushiji-MNIST and Kuzushiji-49 datasets. In addition, the training is conducted with high accuracy by applying learning methods such as Mixup and Random Erase. As a result of the training, the accuracy of the proposed method can be shown to be high by 99.75% for MNIST, 99.07% for Kuzushiji-MNIST, and 97.56% for Kuzushiji-49. Through this deep learning-based technology, it is thought to provide a good research base for various researchers who study East Asian and Western history, literature, and culture.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.9
• /
• pp.47-54
• /
• 2023

Infants utilize crying as a non-verbal means of communication [1]. However, deciphering infant cries presents challenges. Extensive research has been conducted to interpret infant cry audios [2,3]. This paper proposes the classification of infant cries using 3D feature vectors augmented with various audio data techniques. A total of 5 classes (belly pain, burping, discomfort, hungry, tired) are employed in the study dataset. The data is augmented using 5 techniques (Pitch, Tempo, Shift, Mixup-noise, CutMix). Tempo, Shift, and CutMix augmentation techniques demonstrated improved performance. Ultimately, applying effective data augmentation techniques simultaneously resulted in a 17.75% performance enhancement compared to models using single feature vectors and original data.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.479-487
• /
• 2021

In this paper, we propose an Sound Event Detection (SED) model using self-training based on a noisy student model. The proposed SED model consists of two stages. In the first stage, a mean-teacher model based on an Residual Convolutional Recurrent Neural Network (RCRNN) is constructed to provide target labels regarding weakly labeled or unlabeled data. In the second stage, a self-training-based noisy student model is constructed by applying different noise types. That is, feature noises, such as time-frequency shift, mixup, SpecAugment, and dropout-based model noise are used here. In addition, a semi-supervised loss function is applied to train the noisy student model, which acts as label noise injection. The performance of the proposed SED model is evaluated on the validation set of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2020 Challenge Task 4. The experiments show that the single model and ensemble model of the proposed SED based on the noisy student model improve F1-score by 4.6 % and 3.4 % compared to the top-ranked model in DCASE 2020 challenge Task 4, respectively.

• The Korean Journal of Applied Statistics
• /
• v.36 no.2
• /
• pp.101-114
• /
• 2023

The data augmentation technique is effectively used to solve the problem of overfitting the model by allowing the training dataset to be viewed from various perspectives. In addition to image augmentation techniques such as rotation, cropping, horizontal flip, and vertical flip, occlusion-based data augmentation methods such as Cutmix and Cutout have been proposed. For models based on speech data, it is possible to use an occlusion-based data-based augmentation technique after converting a 1D speech signal into a 2D spectrogram. In particular, SpecAugment is an occlusion-based augmentation technique for speech spectrograms. In this study, we intend to compare and study data augmentation techniques that can be used in the problem of false-voice detection. Using data from the ASVspoof2017 and ASVspoof2019 competitions held to detect fake audio, a dataset applied with Cutout, Cutmix, and SpecAugment, an occlusion-based data augmentation method, was trained through an LCNN model. All three augmentation techniques, Cutout, Cutmix, and SpecAugment, generally improved the performance of the model. In ASVspoof2017, Cutmix, in ASVspoof2019 LA, Mixup, and in ASVspoof2019 PA, SpecAugment showed the best performance. In addition, increasing the number of masks for SpecAugment helps to improve performance. In conclusion, it is understood that the appropriate augmentation technique differs depending on the situation and data.

• International Journal of Computer Science & Network Security
• /
• v.22 no.9
• /
• pp.175-182
• /
• 2022

In our daily life, we come across different types of information, for example in the format of multimedia and text. We all need different types of information for our common routines as watching/reading the news, listening to the radio, and watching different types of videos. However, sometimes we could run into problems when a certain type of information is required. For example, someone is listening to the radio and wants to listen to jazz, and unfortunately, all the radio channels play pop music mixed with advertisements. The listener gets stuck with pop music and gives up searching for jazz. So, the above example can be solved with an automatic audio classification system. Deep Learning (DL) models could make human life easy by using audio classifications, but it is expensive and difficult to deploy such models at edge devices like nano BLE sense raspberry pi, because these models require huge computational power like graphics processing unit (G.P.U), to solve the problem, we proposed DL model. In our proposed work, we had gone for a low complexity model for Audio Event Detection (AED), we extracted Mel-spectrograms of dimension 128×431×1 from audio signals and applied normalization. A total of 3 data augmentation methods were applied as follows: frequency masking, time masking, and mixup. In addition, we designed Convolutional Neural Network (CNN) with spatial dropout, batch normalization, and separable 2D inspired by VGGnet [1]. In addition, we reduced the model size by using model quantization of float16 to the trained model. Experiments were conducted on the updated dataset provided by the Detection and Classification of Acoustic Events and Scenes (DCASE) 2020 challenge. We confirm that our model achieved a val_loss of 0.33 and an accuracy of 90.34% within the 132.50KB model size.