Explainable AI explained! | #3 LIME

▬▬ Resources ▬▬▬▬▬▬▬▬▬▬▬

▬▬ Timestamps ▬▬▬▬▬▬▬▬▬▬▬
00:00 Introduction
02:31 Paper explanation
03:34 The math / formula
08:35 LIME in practice

▬▬ Support me if you like 🌟

[sakshamgulati]

One of the best and wholesome explanations on this topic! Kudos!

[yangbai]

The objective function part is amazingly well explained.

[briggstwitchell]

Wow, very clear and concise explanation of a fairly complex technique. Thank you very much for the effort you put in to this. Very helpful!

[arturocdb]

i just want it to ongratulate you!... you are the only one in you tube that can have a solid background (books and papers) about what is teaching in a video. thaank you so much... your video help me a lot!...,

[dvabhinav]

Wow! Such a cool explanation. I had some intuition issues after reading the paper that are now solved due to this video. Thanks a lot!

[hlskojm]

great work! I am a lucky person cuz I found this channel, as My major interests are GNN and XAI. thank you so much for brilliant works.

[techjack]

Concise and great explanation on the topic.

[hqi]

Really cool video. Love the format that mix theory and coding. Just one point for disscussion: We need to measure Recall even for a binary classifier or for balanced data. It measrus a classifier's ability to find out all positive examples. And precison measures how well are the positive predictions. So they measure two different type of errors. In statistics we talk about Error of type one or type two

[ngwoonyee]

thank you so much for the splendid explanation!

[ayaeldeeb]

Really amazing!! It's very exciting and cool to conduct such research; it is also very explainable."

[borauyar]

13:20 I believe "edu" has something to do with "atheism" :) Thanks for the explanations!

[youssefbenhachem]

It is just clearly explained ! Thanks a lot

[MoMo-ibej]

Du bist sehr hilfreich! Quick questions from a relatively non tech person:
1) How does this scale for models with millions of params eg could you even use this on an LLM?
2) Am I right in understanding that you need a mini real data set to train locally every time then?

[ramalaccX]

Damn bro, this helped me a lot with my final thesis

[heejuneAhn]

LIME can explain how the Complex model behaviors about the interested sample point, but it cannot guarantee the prediction of the complex model around the sample is correct predictions. am I right? Also, we could use LIME to explain the incorrect prediction of the complex model. For example, we could take two samples one correct prediction and one incorrect one, then we could make a simple surrogate model abound the two samples so that we could see how the complex model predicts and what is wrong with the complex model. Am I right again?

[randalllionelkharkrang]

Thanks for this. Its funny that "ever_married" feature is inversely proportional to getting a stroke :)

[Anirudh-cfoc]

Thank you for the amazing explaination

[heejuneAhn]

I think in the last example, you cannot say the model is wrong or not. LIME just explain how it behaves near the sample. If you can explain so easily why do you use complex model. You only can say using the ground truth. Some features that you might think is not logical still might affect that specific local points.

[heejuneAhn]

When the feature is not continuous, how do we sample the neighborhood? For example, binary features, Just a combination of features? or consider the binary features to be continuous ones?

[kenalexandremeridi]

What is this for? I am intrigued ( im a musician)