Artificial Intelligence, Machine Learning, Deep Learning, Evolutionary Computation
The following accounts share information about our lab, such as student activities and educational research.
Artificial intelligence (AI) has demonstrated superior performance in various fields, surpassing human capabilities and finding applications in the real world. However, it has become evident that AI is susceptible to vulnerabilities. For instance, the introduction of imperceptible noise into training samples (known as adversarial samples) can lead to misclassification. Additionally, there have been cases where AI gradually learns discriminatory biases through user interactions, resulting in issues such as "AI bias" or " AI that spews hate speech".
To address these vulnerabilities, our research focuses on automatically detecting and improving the weaknesses of AI models to construct models that are safe, stable, and reliable in AI systems.
Tuning the parameters of deep learning models to fully demonstrate their performance often requires domain expertise and a significant amount of time for fine-tuning. Additionally, optimizing models with a large number of structural parameters presents even greater challenges.
To enable users without prior machine learning expertise to effectively apply machine learning techniques, we are actively researching automatically designing models, essentially allowing AI to learn AI, to address the aforementioned problem.
Furthermore, with the rapid advancements in smart wearable devices in recent years, it has become crucial to be able to execute models efficiently on mobile devices as well. To overcome this challenge, our research focuses on designing lightweight models that simplify the structure and reduce computational complexity, allowing the execution of models on mobile devices without sacrificing performance.
Evolutionary computation (EC) is an approximation solution method that uses heuristic search strategies to iteratively optimize within acceptable computational costs, aiming to find feasible and optimal solutions with the required accuracy. EC has been applied to various tasks and has found practical applications in the real world. However, the search efficiency is still not satisfactory, and trial-and-error approaches remain the focus of research.
Therefore, we are addressing this remaining challenge with the goal of achieving improved search efficiency based on a mathematical foundation. Our objective is to develop a new framework for optimization that realizes enhanced search efficiency by leveraging the principles of mathematics. Furthermore, we aim to obtain analytical insights into the operations and strategies of EC that can be linked to this mathematical foundation.
We are deeply interested in various fundamental research and applications related to computational intelligence, not limited to the three topics mentioned above. We also focus on research in machine learning and computer vision, and we welcome individuals who share an interest and enthusiasm in these fields to join us.