PAIR-code/saliency
Framework-agnostic library implementing state-of-the-art saliency methods (XRAI, BlurIG, SmoothGrad, Integrated Gradients, Grad-CAM) for explaining CNN/DNN predictions.
Not currently ranked — collecting fresh signals.
star history
This repository from Google PAIR provides implementations of multiple saliency and attribution techniques for deep learning models, particularly convolutional neural networks. It enables visualization of which input features most influence model predictions, serving the Explainable AI domain. The library includes methods like Guided Integrated Gradients, XRAI, SmoothGrad, and Blur IG, along with the PIC metric for evaluating saliency quality without human annotation.
Frequently asked
- What is PAIR-code/saliency?
- Framework-agnostic library implementing state-of-the-art saliency methods (XRAI, BlurIG, SmoothGrad, Integrated Gradients, Grad-CAM) for explaining CNN/DNN predictions.
- Is saliency open source?
- Yes — PAIR-code/saliency is open source, released under the Apache-2.0 license.
- What language is saliency written in?
- PAIR-code/saliency is primarily written in Jupyter Notebook.
- How popular is saliency?
- PAIR-code/saliency has 997 stars on GitHub.
- Where can I find saliency?
- PAIR-code/saliency is on GitHub at https://github.com/PAIR-code/saliency.