Explaining Image Classification through Knowledge-aware Neuron Interpretation

Yong Guan, Freddy Lecue, Jiaoyan Chen, Ru Li, Jeff Z. Pan

Abstract: Although neural networks have achieved remarkable results, they still encounter doubts due to the intransparency. To this end, neural network prediction explanation is attracting more and more attentions. State of the art methods, however, rarely introduce human-understandable external knowledge, making the explanation hard to interpret by human beings. In this paper, we propose a knowledge-aware framework to explain neural network predictions for image scene classification. We introduce two notions of core concepts, with the help of knowledge graphs, to measure the association of concepts with respect to image scenes, and analyze solutions for prediction explanation and model manipulation. In our experiments on two popular scene classification datasets ADE20k and Opensurfaces, the proposed solutions produce better results than baseline and state of the art methods, e.g., our method produces over 25% IoU gain on compositional explanation for neuron behaviors. In addition, our core concepts and related explanation metrics can help effectively manipulate the model prediction, further leading to a new training method with 26.7% performance improvement.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)