Digital Library[ Search Result ]
Effective Importance-Based Entity Grouping Method in Continual Graph Embedding
http://doi.org/10.5626/JOK.2025.52.7.627
This study proposed a novel approach to improving entity importance evaluation in continual graph embeddings by incorporating edge betweenness centrality as a weighting factor in a Weighted PageRank algorithm. By normalizing and integrating betweenness centrality, the proposed method effectively propagated entity importance while accounting for the significance of information flow through edges. Experimental results demonstrated significant performance improvements in MRR and Hit@N metrics across various datasets using the proposed method compared to existing methods. Notably, the proposed method showed enhanced learning performance after the initial snapshot in scenarios where new entities and relationships were continuously added. These findings highlight the effectiveness of leveraging edge centrality in promoting efficient and accurate learning in continual knowledge graph embeddings.
A VQG Framework for Accurate and Diverse Question Generation
http://doi.org/10.5626/JOK.2025.52.1.62
Visual Question Generation (VQG) aims to generate questions based on a given image, often utilizing additional information such as answers or answer types if necessary. A VQG system should be able to generate diverse questions for a single image, while maintaining relevance to the image alongside its additional information. However, models that highly focus on relevance to the image might overfit to the dataset, leading to limited diversity, while those that emphasize diversity might generate questions less related to the input. Therefore, balancing these two aspects is crucial in VQG. To address this challenge, we proposed BCVQG (BLIP-CVAE VQG), a system that could integrate a pre-trained vision-language model with a Conditional Variational AutoEncoder (CVAE). The effectiveness of the proposed method was validated through quantitative and qualitative evaluations on the VQA2.0 dataset.
Continual Learning using Memory-Efficient Parameter Generation
Hyung-Wook Lim, Han-Eol Kang, Dong-Wan Choi
http://doi.org/10.5626/JOK.2024.51.8.747
Continual Learning with Parameter Generation shows remarkable stability in retaining knowledge from previous tasks. However, it suffers from a gradual decline in parameter generation performance due to its lack of adaptability to new tasks. Furthermore, the difficulty in predetermining the optimal size of the parameter generation model (meta-model) can lead to memory efficiency issues. To address these limitations, this paper proposed two novel techniques. Firstly, the Chunk Save & Replay (CSR) technique selectively stored and replayed vulnerable parts of the generative neural network, maintaining diversity in the parameter generation model while efficiently utilizing memory. Secondly, the Automatically Growing GAN (AG-GAN) technique automatically expanded the memory of the parameter generation model based on learning tasks, enabling effective memory utilization in resource-constrained environments. Experimental results demonstrated that these proposed techniques significantly reduced memory usage while minimizing performance degradation. Moreover, their ability to recover from deteriorated network performance was observed. This research presents new approaches to overcoming limitations of parameter generation-based continual learning, facilitating the implementation of more effective and efficient continual learning systems.
Efficient Prompt Learning Method in Blurry Class Incremental Learning Environment
http://doi.org/10.5626/JOK.2024.51.7.655
Continual learning is the process of continuously integrating new knowledge to maintain performance across a sequence of tasks. While disjoint continual learning, which assumes no overlap between classes across tasks, blurry continual learning addresses more realistic scenarios where overlaps do exist. Traditionally, most related works have predominantly focused on disjoint scenarios and recent attention has shifted towards prompt-based continual learning. This approach uses prompt mechanism within a Vision Transformer (ViT) model to improve adaptability. In this study, we analyze the effectiveness of a similarity function designed for blurry class incremental learning, applied within a prompt-based continual learning framework. Our experiments demonstrate the success of this method, particularly in its superior ability to learn from and interpret blurry data.
A Token Selection Method for Effective Token Pruning in Vision Transformers
http://doi.org/10.5626/JOK.2024.51.6.567
The self-attention-based models, vision transformers, have recently been employed in the field of computer vision. While achieving excellent performance in a variety of tasks, the computation costs increase in proportion to the number of tokens during inference, which causes a degradation in inference speed. Especially when deploying the model in real-world scenarios, many limitations could be encountered. To address this issue, we propose a new token importance measurement, which can be obtained by modifying the structure of multi-head self-attention in vision transformers. By pruning less important tokens through our method during inference, we can improve inference speed while preserving performance. Furthermore, our proposed method, which requires no additional parameters, exhibits better robustness without fine-tuning and demonstrates that it can maximize performance when integrated with existing token pruning methods.
PGB: Permutation and Grouping for BERT Pruning
http://doi.org/10.5626/JOK.2023.50.6.503
Recently, pre-trained Transformer-based models have been actively used for various artificial intelligence tasks, such as natural language processing and image recognition. However, these models have billions of parameters, which require significant computation for inference, and may be subject to many limitations for use in resource-limited environments. To address this problem, we propose PGB(Permutation Grouped BERT pruning), a new group-based structured pruning method for Transformer models. PGB effectively finds a way to change the optimal attention order according to resource constraints, and prunes unnecessary heads based on the importance of the heads to minimize the information loss in the model. Through various comparison experiments, PGB shows better performance in terms of inference speed and accuracy loss than the other existing structured pruning methods for the pre-trained BERT model.
Rehearsal with Stored Latent Vectors for Incremental Learning Over GANs
http://doi.org/10.5626/JOK.2023.50.4.351
Unlike humans, sequential learning of multiple tasks is a difficult problem in a deep learning model. This problem is not only for discriminative models, but also for generative models, such as GAN. The Generative Replay method, which is frequently used in GAN continual learning, uses images generated by GAN provided in the previous task together for learning new tasks, but does not generate good images for CIFAR10, which is a relatively challenging task. Therefore, we can consider a rehearsal-based method that stores a portion of the real data, which cannot store a huge amount of images in limited memory because of large dimension of the real image. In this paper, we propose LactoGAN and LactoGAN+, continual learning methods that store latent vectors that are the inputs of GANs rather than storing real images, as the existing rehearsal-based approaches. As a result, more image knowledge can be stored in the same memory; thus, showing better results than the existing GAN continual learning methods.
Re-Generation of Models via Generative Adversarial Networks and Bayesian Neural Networks for Task-Incremental Learning
http://doi.org/10.5626/JOK.2022.49.12.1115
In contrast to the human ability of continual learning, deep learning models have considerable difficulty maintaining their original performance when the model learns a series of incrementally arriving tasks. In this paper, we propose ParameterGAN, a novel task-incremental learning approach based on model synthesis. The proposed method leverages adversarial generative learning to regenerate neural networks themselves which have a parameter distribution similar to that of a pre-trained Bayesian network. Also, using pseudo-rehearsal methods, ParameterGAN enables continual learning by regenerating the networks of all previous tasks without catastrophic forgetting. Our experiment showed that the accuracy of the synthetic model composed of regenerated parameters was comparable to that of the pre-trained model, and the proposed method outperformed the other SOTA methods in the comparative experiments using the popular task-incremental learning benchmarks Split-MNIST and Permuted-MNIST.
FedGC: Global Consistency Regularization for Federated Semi-supervised Learning
http://doi.org/10.5626/JOK.2022.49.12.1108
Recently, in the field of artificial intelligence, methods of learning neural network models in distributed environments that use sufficient data and hardware have been actively studied. Among them, federated learning, which guarantees privacy preservation without sharing data, has been a dominant scheme. However, existing federated learning methods assume supervised learning using only labeled data. Since labeling costs are incurred for supervised learning, the assumption that only label data exists in the clients is unrealistic. Therefore, this study proposes a federated semi-supervised learning method using both labeled data and unlabeled data, considering a more realistic situation where only labeled data exists on the server and unlabeled data on the client. We designed a loss function considering consistency regularization between the output distributions of the server and client models and analyzed how to adjust the influence of consistency regularization. The proposed method improved the performance of existing semi-supervised learning methods in federated learning settings, and through additional experiments, we analyzed the influence of the loss term and verified the validity of the proposed method.
A Perimeter-Based IoU Loss for Efficient Bounding Box Regression in Object Detection
http://doi.org/10.5626/JOK.2021.48.8.913
In object detection, neural networks are generally trained by minimizing two types of losses simultaneously, namely classification loss and regression loss for bounding boxes. However, the regression loss often fails to achieve its ultimate goal, that is, it often obtains a predicted bounding box that maximally intersects with its target box. This is due to the fact that the regression loss is not highly correlated with the IoU, which actually measures how much the bounding box and its target box overlap with each other. Although several penalty terms have been invented and added to the IoU loss in order to address the problem of regression losses, they still show some inefficiency particularly when penalty terms become zero by enclosing another box or overlapping with the center point before the bounding box and its target box are perfectly the same. In this paper, we propose a perimeter based IoU (PIoU) loss exploiting the perimeter differences of the minimum bounding rectangle of both a predicted box and its target box from those of two boxes themselves. In our experiments using the state-of-the-art object detection models (e.g., YOLO v3, SSD, and FCOS), we show that our PIoU loss consistently achieves better accuracy than all the other existing IoU losses.
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