[4I2-GS-11] AI and Society:

[4I2-GS-11-04] Proposal of an image decomposition method using CNN latent representation

〇Shungo Nukui¹, Noboru Murata¹, Shotaro Akaho² (1. waseda university, 2. National Institute of Advanced Industrial Science and Technology)

Keywords:xai, machine learning, computer vision

In this paper, we propose an explainable AI (XAI) method that decomposes the original image to correspond to arbitrary decompositions of latent representations. With the accelerated societal integration of AI in recent years, numerous XAI methods have been proposed to enhance the interpretability and explainability of AI. However, it remains challenging to explicitly confirm which features of the input image are extracted in the latent representations. To address this, we visualized the correspondence between arbitrarily decomposed components of the latent representations and the features or regions of the original image through image decomposition.
We applied the proposed method to latent representations obtained from an age estimation regression task using facial images. When the latent representations were decomposed into channels that contributed significantly to the prediction and those that did not, the corresponding decomposed images revealed intuitive and significant facial features. For younger individuals, low-frequency stripe patterns indicating smooth skin were associated with channels contributing significantly to the prediction. For older individuals, high-frequency stripe patterns indicating wrinkles were linked to the channels that contributed greatly to the predictions.