Hi,
I'm aware of numerous Explainable AI (XAI) solutions for CNNs, however they ~all seem to be tailored for classification (or segmentation).
I also see that for a small home-made model (lets say 15 layers) in Keras or pythorch it's not a big deal to have these things like SHAP or LRP work. However i don't know if something easily applicable exists for detection task and complex CNN architectures like an FPN+backbone network+something_more_that_makes_things_complicated?
Article Object Detection With Deep Learning: A Review
Article A Deep Learning-Based End-to-End Composite System for Hand D...
https://arxiv.org/pdf/1807.05511https://arxiv.org/pdf/1901.06032
https://medium.com/alumnaiacademy/introduction-to-computer-vision-4fc2a2ba9dc
https://ieeexplore.ieee.org/abstract/document/7930521/http://openaccess.thecvf.com/content_ICCV_2017/papers/Selvaraju_Grad-CAM_Visual_Explanations_ICCV_2017_paper.pdf
https://ieeexplore.ieee.org/abstract/document/8944792/
Thank you for the answer, however i don't see any XAI in the papers you posted that would satisfy my question.
I repeat: given a DL model for detection of objects, is there any explainable AI method that can explain those predictions for the whole image AND is ready to use for complex architectures?
I'm not asking for XAI solutions that work with classification and e.g. ResNet. I'm asking for XAI solutions for detection and e.g. Faster R-CNN.
Maybe I know him. But to be sure, I need to see what the source data is. I also need to understand the issue better
- Badges
- Science topic
- Similar topics
- Design
- Architecture