Federated Learning
Federated learning enables model training using edge devices while protecting client data privacy. However, there remain several limitations for practical deployment.
This research focuses on various challenges in federated learning, including:
- Data Heterogeneity
- System Heterogeneity
- Backdoor vulnerability
- Model Privacy
Graph Neural Networks
Graph Neural Networks (GNNs) often suffer from limitations such as over-smoothing and over-squashing, which arise from the iterative message-passing process.
To mitigate these problems, this research aims to optimize graph structures by applying rewiring techniques that appropriately modify the adjacency matrix before training.
Shortcut Learning
Artificial neural networks tend to learn biased features when the training data itself contains bias. This research aims to explain why neural networks learn such biases and to develop algorithms that mitigate this issue.
In particular, this study proposes the Neural Tangent Kernel (NTK) as an analysis framework and reveals why biased features are learned faster than other features.
AI4Code
AI4Code focuses on advancing language models for code-related tasks, with the goal of achieving more effective and safer code generation.
This research explores how different types of language models can be adapted to the code domain, particularly by improving their ability to generate correct, reliable, and secure code.
A major focus of this research is on Diffusion Language Models, which provide an alternative generation paradigm to conventional autoregressive models.
Through this direction, the study aims to improve code generation performance and enhance the reliability of language models in programming-related applications.