Digital Library[ Search Result ]
Adversarial Training with Contrastive Learning in NLP
Daniela N. Rim, DongNyeong Heo, Heeyoul Choi
http://doi.org/10.5626/JOK.2025.52.1.52
Adversarial training has been extensively studied in natural language processing (NLP) settings to make models robust so that similar inputs derive similar outcomes semantically. However, since language has no objective measure of semantic similarity, previous works use an external pre-trained NLP model to ensure this similarity, introducing an extra training stage with huge memory consumption. This work proposes adversarial training with contrastive learning (ATCL) to train a language processing model adversarially using the benefits of contrastive learning. The core idea is to make linear perturbations in the embedding space of the input via fast gradient methods (FGM) and train the model to keep the original and perturbed representations close via contrastive learning. We apply ATCL to language modeling and neural machine translation tasks showing an improvement in the quantitative (perplexity and BLEU) scores. Furthermore, ATCL achieves good qualitative results in the semantic level for both tasks without using a pre-trained model through simulation.
Model Contrastive Federated Learning on Re-Identification
Seongyoon Kim, Woojin Chung, Sungwoo Cho, Yongjin Yang, Shinhyeok Hwang, Se-Young Yun
http://doi.org/10.5626/JOK.2024.51.9.841
Advances in data collection and computing power have dramatically increased the integration of AI technology into various services. Traditional centralized cloud data processing raises concerns over the exposure of sensitive user data. To address these issues, federated learning (FL) has emerged as a decentralized training method where clients train models locally on their data and send locally updated models to a central server. The central server aggregates these locally updated models to improve a global model without directly accessing local data, thereby enhancing data privacy. This paper presents FedCON, a novel FL framework specifically designed for re-identification (Re-ID) tasks across various domains. FedCON integrates contrastive learning with FL to enhance feature representation, which is crucial for Re-ID tasks that emphasize similarity between feature vectors to match identities across different images. By focusing on feature similarity, FedCON can effectively addresses data heterogeneity challenges and improve the global model's performance in Re-ID applications. Empirical studies on person and vehicle Re-ID datasets demonstrated that FedCON outperformed existing FL methods for Re-ID. Our experiments with FedCON on various CCTV datasets for person Re-ID showed superior performance to several baselines. Additionally, FedCON significantly enhanced vehicle Re-ID performance on real-world datasets such as VeRi-776 and VRIC, demonstrating its practical applicability.
Improving the Quality of Generating Imbalance Data in GANs through an Exhaustive Contrastive Learning Method
Hyeonjun Shin, Sangbaek Lee, Kyuchul Lee
http://doi.org/10.5626/JOK.2023.50.4.295
As the performance of deep learning algorithms has improved, they are being used as a way to solve various problems in the real world. In the case of data that reflect the real world, imbalance data may occur depending on the frequency of occurrence of events or the difficulty of collection. Data with an inconsistent number of classes that make up the data are called imbalance data, and in particular, it is difficult to learn the minority classes with relatively little data through Deep Learning algorithms. Recently, Generative Adversarial Nets (GANs) have been applied as a method for data augmentation, and self-supervised learning-based pre-learning has been proposed for minority class learning. However, because class information of imbalance data is utilized in the process of learning the Generative Model, the quality of generated data is poor due to poor learning of minority classes. To solve this problem, this paper proposes a similarity-based exhaustive contrast learning method. The proposed method is quantitatively evaluated through the Frechet Inception Distance (FID) and Inception Score (IS). The method proposed in this paper confirmed the performance improvement of the Frechet Inception Distance of 16.32 and the Inception Score of 0.38, as compared to the existing method.
CoEM: Contrastive Embedding Mapper for Audio-visual Latents
Gihun Lee, Kyungchae Lee, Minchan Jeong, Myungjin Lee, Se-young Yun, Chan-hyun Yun
http://doi.org/10.5626/JOK.2023.50.1.80
Human perception can link audio-visual information to each other, making it possible to recall visual information from audio information and vice versa. Such ability is naturally acquired by experiencing situations where these two kinds of information are combined. However, it is hard to obtain video datasets that are richly combined with both types of information, and at the same time, labeled for the semantics of each scene. This paper proposes a Contrastive Embedding Mapper (CoEM), which maps embedding from one type of information to the another, corresponding to its categorical modality. Paired data is not required, CoEM learns to contrast the mapped embedding by its categories. We validated the efficacy of CoEM on the embeddings for audio and visual datasets which were trained to classify 20 shared categories. In the experiment, the embedding mapped by CoEM showed that it was capable of retrieving and generating data on its mapped domain.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr