My primary research interest lies in applying the machine learning model to real-world applications in various domains. In particular, I am motivated by the challenging tasks such as reducing the gap between benchmark datasets and real-world datasets, and designing efficient models for the real world. In the end, I hope that my contribution will help improve the applicability of the machine learning models to real-world applications. Details of my previous and current research interests are as follows:
Addressing Dataset Discrepancies.
1. In time-series synthesis, previous works focus on either regular or irregular time series synthesis. However, time-series data is not standardized into regular and irregular in real-world applications. Therefore, I proposed a general purpose model capable of synthesizing regular and irregular time series data using a neural controlled differential equation that can handle both time series datasets.
2. Visual recognition models often encounter images different from the training sets in resolution and/or subject to natural variations such as weather changes. The images significantly degrade the performance of the models. Therefore, I proposed a Fourier neural operator-based model that operates in the frequency domain, which can significantly alleviate performance degradation by processing images of different resolutions and natural variations within the single model.
Designing Efficient Models.
1. In time series anomaly detection, early detection of anomalies is very important. I proposed a proactive approach to predict and detect anomalies, moving beyond reactive methods prevalent in previous studies which detect anomalies after they occur. By integrating a time-series forecasting model, I aimed to enhance early predicting anomalies capabilities.
2. Diffusion models show good performance in terms of sampling quality and sampling diversity, but suffer from long sampling times which is one of the generation task trilemmas. To resolve the trilemma of generative models, I proposed a GAN-based method that approximates the straight-path interpolation. This method significantly improved sampling efficiency while maintaining balanced performance across generative task trilemma criteria.
Looking ahead, I am eager to explore the potential of foundation models, such as large language models, in addressing machine learning tasks for real-world applications. While these models have demonstrated exceptional performance in certain domains like natural language processing and image recognition, their broader applicability across diverse domains remains largely untapped. I aspire to contribute to this emerging field by leveraging foundation models to solve challenges in various real-world applications.
In summary, my research endeavors are driven by a commitment to advancing machine learning techniques that not only excel in benchmark settings but also translate seamlessly to real-world scenarios, ultimately delivering tangible benefits to anyone who uses machine learning models.