Mr. PARAJULI (2nd-year master's student in Racharak Lab, Computing Science Research Area) received Best Poster Award in the 8th International Conference on Information Technology (InCIT 2024).

InCIT 2024 is a premier forum for sharing research in areas related to information and communication technologies, aimed at fostering research on intelligent technology and innovation for the future of society. InCIT 2024 was held in Chonburi, Thailand and Kanazawa, Ishikawa, Japan on November 14 - 15, 2024.

*Reference: InCIT 2024

■Date Awarded
November 14, 2024

■Title
Towards Identification of Underspecification Problem in Image Classification

■Authors
Prabhat Parajuli, Teeradaj Racharak

■Abstract
In machine learning pipelines, models with different architectures and inductive biases can achieve similar performance on a given dataset if their design is complex enough to fit the training data. However, these models often fail to generalize well to real-world scenarios. This failure mode is known as underspecification, which poses a major threat to the reliability of ML/AI systems. We propose a framework that systematically identifies this issue using eXplainable AI (XAI) techniques, specifically Local Interpretable Model-Agnostic Explanations (LIME). We assume that models with comparable performance should exhibit similar feature learning; thus, inconsistencies in LIME explanations for the same input data indicate divergent feature reliance. Such differences suggest potential reliance on spurious features, underscoring the risk of underspecification. We validate our approach on the MNIST and Imagenette datasets, demonstrating that even simple tasks can reveal underspecification when using complex models.

■Comment
Attending InCIT was an incredible experience, made even more memorable by receiving the Best Poster Award for our work, "Towards Identification of the Underspecification Problem in Image Classification. "Sharing insights from our research and receiving valuable feedback was both inspiring and motivating, encouraging me to pursue even higher-quality research. The engaging discussions with fellow researchers and experts not only reinforced the significance of our work but also showcased the collaborative and innovative spirit of the community. I am deeply grateful for this platform to contribute, learn, and grow.

Mr. Parajuli (Left), Award Presenter (Center), and Senior Lecturer Racharak (Right)

　　

January 15, 2025