Digital Library[ Search Result ]
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.
A Hybrid Deep Learning Model for Generating Time-series Fire Data in Underground Utility Tunnel based on Convolutional Attention TimeGAN
http://doi.org/10.5626/JOK.2024.51.6.490
Underground utility tunnels (UUTs) play a crucial role in urban operation and management. Fires are the most common disasters in the facilities, and there is a growing demand for fire management systems using artificial intelligence (AI). However, due to the difficulty of collecting fire data for AI training, utilizing data generation models reflecting the key characteristics of real fires can be an alternative. In this paper, we propose an approach for generating AI training data based on the fire data generation model CA-TimeGAN. To collect fire simulation data for training the proposed model, we constructed a UUT in Chungbuk Ochang within the fire dynamic simulator (FDS) virtual environment. In the experiments, we compared data generated by TimeGAN and CA-TimeGAN, verifying the data quality and effectiveness. Discriminative score converged to 0.5 for both CA-TimeGAN and TimeGAN. Predictive scores improved by 66.1% compared to models trained only on simulated data and by 22.9% compared to models incorporating TimeGAN-generated data. PCA and t-SNE analyses showed that the distribution of generated data was similar to that of simulated data.
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.
Copy-Paste Based Image Data Augmentation Method Using
http://doi.org/10.5626/JOK.2022.49.12.1056
In the field of computer vision, massive well-annotated image data are essential to achieve good performance of a convolutional neural network (CNN) model. However, in real world applications, gathering massive well-annotated data is a difficult and time-consuming job. Thus, image data augmentation has been continually studied. In this paper, we proposed an image data augmentation method that could generate more diverse image data by combining generative adversarial network (GAN) and copy-paste based augmentation. The proposed method generated not pixel-level or image-level augmentation, but object-level augmentation by cutting off segmentation boundaries(mask) instead of bounding boxes. It then applyied GAN to transform objects.
Facial Emotion Recognition Data Augmentation using Generative Adversarial Network
http://doi.org/10.5626/JOK.2021.48.4.398
The facial emotion recognition field of computer vision has recently been identified to demonstrate meaningful results through various neural networks. However, the major datasets of facial emotion recognition have the problem of “class imbalance,” which is a factor that degrades the accuracy of deep learning models. Therefore, numerous studies have been actively conducted to solve the problem of class imbalance. In this paper, we propose “RDGAN,” a facial emotion recognition data augmentation model that uses a GAN to solve the class imbalance of the FER2013 and RAF_single that are used as facial emotion recognition datasets. RDGAN is a network that generates images suitable for classes by adding expression discriminators based on the image-to-image translation model between the existing images as compared to the prevailing studies. The dataset augmented with RDGAN showed an average performance improvement of 4.805%p and 0.857%p in FER2013 and RAF_single, respectively, compared to the dataset without data augmentation.
Research on WGAN models with Rényi Differential Privacy
Sujin Lee, Cheolhee Park, Dowon Hong, Jae-kum Kim
http://doi.org/10.5626/JOK.2021.48.1.128
Personal data is collected through various services and managers extract values from the collected data and provide individually customized services by analyzing the results. However, data that contains sensitive information, such as medical data, must be protected from privacy breaches. Accordingly, to mitigate privacy invasion, Generative Adversarial Network(GAN) is widely used as a model for generating synthetic data. Still, privacy vulnerabilities exist because GAN models can learn not only the characteristics of the original data but also the sensitive information contained in the original data. Hence, many studies have been conducted to protect the privacy of GAN models. In particular, research has been actively conducted in the field of differential privacy, which is a strict privacy notion. But it is insufficient to apply it to real environments in terms of the usefulness of the data. In this paper, we studied GAN models with Rényi differential privacy, which preserve the utility of the original data while ensuring privacy protection. Specifically, we focused on WGAN and WGAN-GP models, compared synthetic data generated from non-private and differentially private models, and analyzed data utility in each scenario.
Detecting Mode Drop and Collapse in GANs Using Simplified Frèchet Distance
Chung-Il Kim, Seungwon Jung, Jihoon Moon, Eenjun Hwang
http://doi.org/10.5626/JOK.2019.46.10.1012
Even though generative adversarial network (GAN) is an excellent model for generating data based on the estimation of real data distribution by of two adversarial learning network, this model often suffers from mode drop that does not learn distribution during learning, or mode collapse that generates only one or very few distribution samples. Most studies to detect these problems have used well-balanced data or additional neural network models. In this paper, we propose a method to detect mode drop and collapse by using a simplified Frèchet distance, which does not require any additional model or well-balanced data. Through various experiments, we showed that our proposed distance metric detected mode drop and collapse more accurately than any other metrics used in GANs.
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