วีดีโอ

That is correct! LIME is model agnostic, which means you can use any image classification model and it would work the same.

@@cristianarteaga and this sentence "Xi = (Xi - 0.5)*2" in the image, is it independent of the model? i really appreciate your answer.

@@RiuHayabusa15 Good point, that part is actually specific for the Inception model, as this model applies that specific transformation for pre-processing. For other models you do not need that transformation.

@@cristianarteaga Okey i get it, thanks a lot for your help

Thank you for the feedback. I will have it in mind for future videos. :)

You can also adjust the speed of the playback.

@@datakristen8500 that's true and how I did it but then some words are stretched out and the speech doesn't feel natural and it can be hard to listen to it.

Sure! Here it is: nbviewer.jupyter.org/url/arteagac.github.io/blog/lime_image.ipynb

@@cristianarteaga thank you and happy new year

@@cristianarteaga not available anymore

@@atejap I think the nbviewer website is temporarily down but you can see the original notebook in the following link: github.com/arteagac/arteagac.github.io/blob/master/blog/lime_image.ipynb

​@@cristianarteaga Tks so much.

Sure! github.com/arteagac/arteagac.github.io/blob/master/blog/lime.ipynb

Hola @melalastro. No estoy seguro si entiendo bien tu pregunta pero si te refieres al nuevo dataset de perturbaciones que se genera, entonces si necesitas predecir la clase para cada una de las perturbaciones. Recuerda que la idea general detras de LIME es ver como se comportan las predicciones de tu modelo de machine learning en el vecindario de la instancia que quieres explicar. Por tanto, el vecindario que creas con perturbaciones requiere evaluar la predicion de cada perturbacion. Si tienes dudas adicionales puedes comunicarte por este medio.

@@cristianarteaga gracias por contestar, me refiero a la localidad, si tengo 200 registros para entrenar, miro algunos para darle interpretabikidad al modelo, o solo lo uso para cada nueva predicción?

Entiendo tu pregunta. Dado que LIME es un methodo para explicaciones locales (un registro al tiempo), para poder darle una interpretacion global al modelo (usando todos tus registros) tienes que recurrir a estrategias adicionales. Por ejemplo : 1. Una estrategia es usar Submodular-Pick LIME (SP-LIME) que es mencionada en el paper original de LIME (section 4). SP-LIME busca seleccionar una muestra representativa de tus registros para asi observar un comportamiento general del modelo. 2.Otra estrategia es agregar las explicaciones individuales para producir una explicacion global como se sugiere en este paper: arxiv.org/pdf/1907.03039.pdf

@@cristianarteaga muchas gracias!!!

@@cristianarteaga muy bueno todo! mil gracias

You have a good point. It would be interesting to test the effect of different segmentation strategies on the quality of the explanations.

Hello. Thanks for your comment. Sorry for the late reply. 1) Your interpretation for areas far from the decision boundary makes sense. In fact, due to issues like the one you mention, LIME may not perform well for certain continuous features. For that reason, the python library LIME uses discretization by default for continuous variables (see github.com/marcotcr/lime/issues/68). 2) I am not sure whether it is a good idea to increase the kernel size to a point where a more "global" explanation is obtained. I think sticking to a "local" scope is better because the linear surrogate model you are estimating may not behave very well when you move very far from the surrounding area of the point you want to explain.

Hello! Thanks for your comment. With categorical features it is even easier. To create perturbations you simply create a random vector of zeroes and ones instead of sampling from the mean and std. dev. of your features. Let's say you have a binary categorical feature called "gender" (1: female, 0: male). The LIME random perturbation for the feature "gender" will use 0 to represent that you perturb your feature (flip it to the opposite value) and 1 represents you don't perturb it. You can clearly see this concept on another tutorial where I apply to LIME explanations of image classifications (th-cam.com/video/ENa-w65P1xM/w-d-xo.html). You can assume that each superpixel in the image is a binary categorical feature that is turned on and off with random perturbations and then it is evaluated how the presence of absence of these features affect the predictions. This exact concept of turning on and off categorical features applies to your question.

Cristian Arteaga thank you very much. your explanation is so useful. I have one more question. If i discretize continuous feature, the process is same with categrical feature?

That is correct! When you discretize a continuous feature you generate LIME perturbations using the same process used for categorical features. A discretized feature is just a feature converted to categorical using percentiles.

Hi Hera. You are right. Thank you for pointing this out. The error was that the import for skimage.io should be "import skimage.io" instead of "import skimage". I have updated the notebook and made sure it works. You can check the updated version in the blog section of my website arteagac.github.io/blog.html . You can also open it in Google Colab for quick testing.

Hey David. Yes, it will work once you replace the model with your hotdog classifier. If you want to do a quick test, you can use the Jupyter notebook I wrote arteagac.github.io/blog.html. However, for more stable results, I recommend you to check the LIME Python library: github.com/marcotcr/lime.

Thanks!!

Gracias :)