• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.18 no.3
• /
• pp.670-684
• /
• 2024

3D Cross-Modal Retrieval (3DCMR) is a task that retrieves 3D objects regardless of modalities, such as images, meshes, and point clouds. One of the most prominent methods used for 3DCMR is the Cross-Modal Center Loss Function (CLF) which applies the conventional center loss strategy for 3D cross-modal search and retrieval. Since CLF is based on center loss, the center features in CLF are also susceptible to subtle changes in hyperparameters and external inferences. For instance, performance degradation is observed when the batch size is too small. Furthermore, the Mean Squared Error (MSE) used in CLF is unable to adapt to changes in batch size and is vulnerable to data variations that occur during actual inference due to the use of simple Euclidean distance between multi-modal features. To address the problems that arise from small batch training, we propose a Noisy Center Loss (NCL) method to estimate the optimal center features. In addition, we apply the simple Siamese representation learning method (SimSiam) during optimal center feature estimation to compare projected features, making the proposed method robust to changes in batch size and variations in data. As a result, the proposed approach demonstrates improved performance in ModelNet40 dataset compared to the conventional methods.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.19 no.1
• /
• pp.137-142
• /
• 2024

Recently, deep learning technology has become those methods as de facto standards in the area of medical data representation. But, deep learning inherently requires a large amount of training data, which poses a challenge for its direct application in the medical field where acquiring large-scale data is not straightforward. Additionally, brain signal modalities also suffer from these problems owing to the high variability. Research has focused on designing deep neural network structures capable of effectively extracting spectro-spatio-temporal characteristics of brain signals, or employing self-supervised learning methods to pre-learn the neurophysiological features of brain signals. This paper analyzes methodologies used to handle small-scale data in emerging fields such as brain-computer interfaces and brain signal-based state prediction, presenting future directions for these technologies. At first, this paper examines deep neural network structures for representing brain signals, then analyzes self-supervised learning methodologies aimed at efficiently learning the characteristics of brain signals. Finally, the paper discusses key insights and future directions for deep learning-based brain signal analysis.

• Journal of the Korea Society of Computer and Information
• /
• v.25 no.5
• /
• pp.187-197
• /
• 2020

Recently, various methods of text embedding using deep learning algorithms have been proposed. Especially, the way of using pre-trained language model which uses tremendous amount of text data in training is mainly applied for embedding new text data. However, traditional pre-trained language model has some limitations that it is hard to understand unique context of new text data when the text has too many tokens. In this paper, we propose self-supervised learning-based fine tuning method for pre-trained language model to infer vectors of long-text. Also, we applied our method to news articles and classified them into categories and compared classification accuracy with traditional models. As a result, it was confirmed that the vector generated by the proposed model more accurately expresses the inherent characteristics of the document than the vectors generated by the traditional models.

• The Journal of Korea Robotics Society
• /
• v.18 no.3
• /
• pp.346-351
• /
• 2023

With the release of numerous open driving datasets, the demand for domain adaptation in perception tasks has increased, particularly when transferring knowledge from rich datasets to novel domains. However, it is difficult to solve the change 1) in the sensor domain caused by heterogeneous LiDAR sensors and 2) in the environmental domain caused by different environmental factors. We overcome domain differences in the semi-supervised setting with 3-stage model parameter training. First, we pre-train the model with the source dataset with object scaling based on statistics of the object size. Then we fine-tine the partially frozen model weights with copy-and-paste augmentation. The 3D points in the box labels are copied from one scene and pasted to the other scenes. Finally, we use the knowledge distillation method to update the student network with a moving average from the teacher network along with a self-training method with pseudo labels. Test-Time Augmentation with varying z values is employed to predict the final results. Our method achieved 3rd place in ECCV 2022 workshop on the 3D Perception for Autonomous Driving challenge.

• Proceedings of the Korean Operations and Management Science Society Conference
• /
• 2005.05a
• /
• pp.599-606
• /
• 2005

Novelty detection involves identifying novel patterns. They are not usually available during training. Even if they are, the data quantity imbalance leads to a low classification accuracy when a supervised learning scheme is employed. Thus, an unsupervised learning scheme is often employed ignoring those few novel patterns. In this paper, we propose two ways to make use of the few available novel patterns. First, a scheme to determine local thresholds for the Self Organizing Map boundary is proposed. Second, a modification of the Learning Vector Quantization learning rule is proposed so that allows one to keep codebook vectors as far from novel patterns as possible. Experimental results are quite promising.

• ETRI Journal
• /
• v.46 no.3
• /
• pp.513-525
• /
• 2024

Deep neural networks trained on labeled medical data face major challenges owing to the economic costs of data acquisition through expensive medical imaging devices, expert labor for data annotation, and large datasets to achieve optimal model performance. The heterogeneity of diseases, such as Alzheimer's disease, further complicates deep learning because the test cases may substantially differ from the training data, possibly increasing the rate of false positives. We propose a reconstruction-based self-supervised anomaly detection model to overcome these challenges. It has a dual-subnetwork encoder that enhances feature encoding augmented by skip connections to the decoder for improving the gradient flow. The novel encoder captures local and global features to improve image reconstruction. In addition, we introduce an entropy-based image conversion method. Extensive evaluations show that the proposed model outperforms benchmark models in anomaly detection and classification using an encoder. The supervised and unsupervised models show improved performances when trained with data preprocessed using the proposed image conversion method.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.24 no.5
• /
• pp.482-488
• /
• 2014

This paper aims to analyze user's emotion automatically by analyzing Twitter, a representative social network service (SNS). In order to create sentiment analysis models by using machine learning techniques, sentiment labels that represent positive/negative emotions are required. However it is very expensive to obtain sentiment labels of tweets. So, in this paper, we propose a sentiment analysis model by using self-training technique in order to utilize "data without sentiment labels" as well as "data with sentiment labels". Self-training technique is that labels of "data without sentiment labels" is determined by utilizing "data with sentiment labels", and then updates models using together with "data with sentiment labels" and newly labeled data. This technique improves the sentiment analysis performance gradually. However, it has a problem that misclassifications of unlabeled data in an early stage affect the model updating through the whole learning process because labels of unlabeled data never changes once those are determined. Thus, labels of "data without sentiment labels" needs to be carefully determined. In this paper, in order to get high performance using self-training technique, we propose 3 policies for updating "data with sentiment labels" and conduct a comparative analysis. The first policy is to select data of which confidence is higher than a given threshold among newly labeled data. The second policy is to choose the same number of the positive and negative data in the newly labeled data in order to avoid the imbalanced class learning problem. The third policy is to choose newly labeled data less than a given maximum number in order to avoid the updates of large amount of data at a time for gradual model updates. Experiments are conducted using Stanford data set and the data set is classified into positive and negative. As a result, the learned model has a high performance than the learned models by using "data with sentiment labels" only and the self-training with a regular model update policy.

• Korean Journal of Radiology
• /
• v.24 no.11
• /
• pp.1061-1080
• /
• 2023

Artificial intelligence (AI) in radiology is a rapidly developing field with several prospective clinical studies demonstrating its benefits in clinical practice. In 2022, the Korean Society of Radiology held a forum to discuss the challenges and drawbacks in AI development and implementation. Various barriers hinder the successful application and widespread adoption of AI in radiology, such as limited annotated data, data privacy and security, data heterogeneity, imbalanced data, model interpretability, overfitting, and integration with clinical workflows. In this review, some of the various possible solutions to these challenges are presented and discussed; these include training with longitudinal and multimodal datasets, dense training with multitask learning and multimodal learning, self-supervised contrastive learning, various image modifications and syntheses using generative models, explainable AI, causal learning, federated learning with large data models, and digital twins.

• Journal of the Korea Convergence Society
• /
• v.12 no.10
• /
• pp.27-35
• /
• 2021

Automatic post editing is a research field that aims to automatically correct errors in machine translation results. This research is mainly being focus on high resource language pairs, such as English-German. Recent APE studies are mainly adopting transfer learning based research, where pre-training language models, or translation models generated through self-supervised learning methodologies are utilized. While translation based APE model shows superior performance in recent researches, as such researches are conducted on the high resource languages, the same perspective cannot be directly applied to the low resource languages. In this work, we apply two transfer learning strategies to Korean-English APE studies and show that transfer learning with translation model can significantly improves APE performance.

• Journal of Electrical Engineering and Technology
• /
• v.10 no.4
• /
• pp.1910-1914
• /
• 2015

Human posture recognition is an attractive and challenging topic in computer vision due to its promising applications in the areas of personal health care, environmental awareness, human-computer-interaction and surveillance systems. Human posture recognition in video sequences consists of two stages: the first stage is training and evaluation and the second is deployment. In the first stage, the system is trained and evaluated using datasets of human postures to ‘teach’ the system to classify human postures for any future inputs. When the training and evaluation process is deemed satisfactory as measured by recognition rates, the trained system is then deployed to recognize human postures in any input video sequence. Different classifiers were used in the training such as Multilayer Perceptron Feedforward Neural networks, Self-Organizing Maps, Fuzzy C Means and K Means. Results show that supervised learning classifiers tend to perform better than unsupervised classifiers for the case of human posture recognition.