[Feature Request] map_to_interpret_space function for Lime and Kernel Shap

[dejeanph]

The methods Lime and KernelShap need a map_to_interpret_space function. By default, the function is the quickshift segmentation algorithm. For better results, watershed function on images have to be used. This algorithm, in the skimage, package needs a parameter (makers) which value is dedicated to the image treated and not for all images of the dataset.
The current API of Lime and KernelShap cannot allowed such configuration (function + dedicated value).

A solution could be to have a parameter for the image maps computed in advance instead of compute them every time in the explicability method function.

[lucashervier]

You want to change the default map_to_interpret_space function for LIME & KernelSHAP for images inputs to watershed. But from my understanding this new function need to be parametrized for EACH images which I do not believe make the utilisation of the library easier for someone beginning with XAI. Therefore, I do not believe that changing the default function is a good solution.

However, what we can do is to propose a pool of image segmentation algorithms including watershed. Since I am not familiar with watershed can you provide some of your workaround (in a colab notebook) in order to know exactly what you expect and how you are currently doing ?

Thanks!

[dejeanph]

As I said, the watershed algorithm in the skimage package needs a parameter based on the processed image. This is difficult to include and transmit to the map_to_interpret_space function.
A solution seems to add a parameter indicating the macropixel map directly, calculated outside the XAI process.
Thus we no longer transmit the calculation function of the map but the map itself.

[RuoyuChen10]

You want to change the default map_to_interpret_space function for LIME & KernelSHAP for images inputs to watershed. But from my understanding this new function need to be parametrized for EACH images which I do not believe make the utilisation of the library easier for someone beginning with XAI. Therefore, I do not believe that changing the default function is a good solution.

However, what we can do is to propose a pool of image segmentation algorithms including watershed. Since I am not familiar with watershed can you provide some of your workaround (in a colab notebook) in order to know exactly what you expect and how you are currently doing ?

Thanks!

How to write a function? Can you give an example, please?

[AntoninPoche]

There is a tutorial on Lime and KernelShap with Xplique: https://colab.research.google.com/drive/1frholXRE4XQQ3W5yZuPQ2-xqc-LTczfT

Does it help in the use of map_to_interpret_space?

[dejeanph]

Je pense que le problème de la fonction de segmentation est un des points les plus importants de la méthode, car avec l’expérience de Confiance.AI et les nombreux cas d’usage traités, c’est le point le plus complexe. En fait, « image processing » parlant, il est directement lié aux couches de convolutions du modèle en réduisant l’espace de la donnée présentée au modèle. Cette réduction par segmentation doit être en cohérence avec les convolutions effectuées lors de l’apprentissage. S’il n’y a pas cohérence, les paramètres de convolutions ne sont pas efficients avec l’espace segmenté. Ceci est mon simple avis. Philippe De : Antonin Poché ***@***.***> Envoyé : lundi 25 septembre 2023 17:11 À : deel-ai/xplique ***@***.***> Cc : DEJEAN Philippe ***@***.***>; Author ***@***.***> Objet : Re: [deel-ai/xplique] [Feature Request] map_to_interpret_space function for Lime and Kernel Shap (Issue #92) There is a tutorial on Lime and KernelShap with Xplique: https://colab.research.google.com/drive/1frholXRE4XQQ3W5yZuPQ2-xqc-LTczfT Does it help in the use of map_to_interpret_space? — Reply to this email directly, view it on GitHub<#92 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AVFVKJSI5ZBK7BCKJ5OAKFTX4GNHVANCNFSM5NF4VJRQ>. You are receiving this because you authored the thread.Message ID: ***@***.******@***.***>>

[AntoninPoche]

I agree that those methods' results are highly dependent on the segmentation. However, I did not see in the literature how to include the model in the segmentation, if you saw something and think it is necessary for the method, feel free to contribute.

However, we should keep in mind that Lime and KernelShap are black box methods, hence, the model's convolution layers cannot be taken into account, otherwise, it becomes a white box method.