What are the best techniques for geospatial datasets? Also, are there some techniques that are better suited for stacking of models than using a single model.?
No, I was curious about the various feature visualization techniques and where do they work best?
Hi, You can use SHAP for this purpose. There is a DeepShap algorithm for deep learning models.
https://github.com/slundberg/shap
Thanks, Asgar Hussain, It looks interesting.
So far I have only heard of activation maximization and Network Dissection.
https://arxiv.org/pdf/1804.11191.pdf
This was also a good paper to follow but truly saying, only got half of it. Are there some more good resources( preferably practical ones ) for the same.
Jason Brownlee has a nice tutorial for visualizing filters and feature maps in convolutional neural networks:
https://machinelearningmastery.com/how-to-visualize-filters-and-feature-maps-in-convolutional-neural-networks/
Here are some example by using MATLAB: (Click on image and see its coding example) https://au.mathworks.com/help/deeplearning/examples.html?category=deep-learning-tuning-and-visualization
For model stacking, you can check AutoGluon ( Preprint AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data
). Many AutoML methods use stacking to improve the final results. In terms of feature visualization, if it's in terms of interpretability of computer vision tasks, you can check this blog: https://bair.berkeley.edu/blog/2020/04/23/decisions/
Check this out
https://github.com/yosinski/deep-visualization-toolbox
I dig myself into CNN explainability in the past months. Most of the projects are suitable only one purpose or damn slow per image. I found an implementation of GradCAM, GradCAM++ and ScoreCAM from a Japanese developer on Github. It is awesome, no dependencies and very easy to integrate into your Python stack. And these methods can be run on a larger scale on many images in a reasonable amount of time:
https://github.com/tabayashi0117/Score-CAM/blob/master/gradcamutils.py
But it is for Python-Tensorflow (Keras). I can't help in other frameworks-programming languages.
Other notes:
- SHAP APIs are not straightforward to use and that method is damn slow. Not very useful to do quick iterations of your work.
- Backprop-based techniques got criticism for being not really working.
- LIME got similar criticism.
- For saliency maps, I did not find them useful at all.
Any other methods are very much WIP, no public software or you have to implement yourself. I found the above mentioned CAM-based methods, easy to integrate, more or less working and they are relative quick per image.
Thanks, Csaba Kertész for the explanation part. I was also not sure which one of them would run well on a large dataset ( i.e. performance).
Which one did you find most useful from an interpretation point of view ( neglecting the slow performance).?
Can you also share some resources that you found useful during your study about them?
- Badges
- Science topic