Department of Image Processing

Neural networks are considered as black-box models, not specifying the reasoning behind their estimations. In this seminar, we will talk about model interpretability, i.e. a weakness in CNNs besides their superior performance in many visual tasks. We will take a close look at a method to modify traditional CNNs into interpretable CNNs. In short, the proposed method addresses the research question “how to regularize filters such that CNNs can have semantically meaningful filters when visualized?”. We will also discuss the method’s drawbacks and how to improve them.

[1] Zhang, Q., Wu, Y. N., & Zhu, S. C. (2018). Interpretable convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 8827-8836).