Explainable AI for Human-Metaverse

Deep learning is a powerful tool for modeling behavior in science and medicine. However,  it is difficult to analyze the model because it has a black-box nature. We will study explainable AI models for various topics in the human-metaverse project for visualizing and easy to analyze the models.

Research Outline

We have been studying an explainable AI model for visualizing which part of the image is important for the model decision in classification problems. We proposed SCOUTER model which visualizes the positive and negative locations of features in the input image. The figure shows the positive “why 7” and the negative “why not 1 or 2” visualizing results in handwriting classification. The heatmaps show which part is important for “7” or the differences between “7” and “1” or “2”. We will apply this model to the bio-medical models in the project to analyze the relation between the input and prediction.

References

LiangzhiLi, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, Hajime Nagahara, “SCOUTER: Slot attention-based classifier for explainable image recognition”, International Conference on Computer Vision, Oct., 2021.