Training Latent Dirichlet Allocation: Gibbs Sampling (Part 2 of 2)
Вставка
- Опубліковано 27 чер 2024
- This is the second of a series of two videos on Latent Dirichlet Allocation (LDA), a powerful technique to sort documents into topics. In this video, we learn to train an LDA model using Gibbs sampling.
The first video is here: • Latent Dirichlet Alloc... - Наука та технологія
This is the best tutorial that I could find around the web. I'm going to code and put it in my dissertation references. Also, I've watched it several times 😁 thank you
Exceptional skills in breaking down complex topics into simple and relatable words! Well done!
Words fail to thank you, your explanation is very, very wonderful. I hope you will continue explaining all the models of machine learning in this way and add subtitles to your videos.
Luis, you are very talented as tutor/teacher. I'am impressed! Glad to found you and I am your new subscriber :) Keep the good work and wish you all the best!
Fantastic content my man. Keep doing what you're doing.
God... Just loved your videos on LDA... I had been running all over the sources to understand LDA. You dropped in like a GOD.. Sigmoid(Thanks)
Man oh man Luis, you videos are a treat. You really deserve to know how much you're helping me and my fellow learners out, and how much it means to us. First saw you in the deep learning introductory course in udacity, and loved the content. Can't express how happy I was to discover you had a youtube channel ! Thank you !
PS : it was also fun rewatching the RNN/LSTM part of that course, and noticing that the person teaching us there was none other than Alexis Cook, who I had some occasions of briefly discussing with in Kaggle learn forums. Realizing that gave me a Machine Learning Cinematic Universe vibes hahah
Thank you for your kind message, it's so nice to hear the videos are helpful!
Yes, I enjoyed working from Alexis and I learned a lot from her. Check out her blog, she has some nice DL posts: alexisbcook.github.io/
best material on LDA online! this should have far more views.
A very simple and easy to understand explanation for a complex model. Amazing Video!
Muchas gracias Luis. You are very good at explaining complex things in a comprehensive way. Your videos help me a lot. Keep going!
Amazingly intuitive once again! Great series.
the best youtube videos are always the ones kept secret... thank you so so much the explanation is superb. i would not have gotten a role as a data scientist if not for amazing teachers like yo
Thank you very much, Luis Serrano! You are so talented teacher!
this is the right place I have been looking around for LDA. Thank you very much. It will be better when you also provide the lectures with the code.
Your explanation is so easy to understand and very clear. Thank you so much.
This was an amazing series. Thanks a ton!!
I found it clear and engaging. Thanks for your work!
Man, I think that our teacher bases his lessons and our projects themes of of your vids :D. Thanks. Really helped.
You are a true gem in teaching 😍. Please make more video as possible
Best intuitive explanation of LDA I have seen. Thank you for sharing ! Looking forward to Grokking NLP ;)
Wow...what an explanation! Great work man. Next time I sit for an interview, this is the best video to help me revise one of my internships :)
Awesome video for LDA! Thank you very much!
Amazing explanation. Very easy to understand. Thanks a lot!
Wonderful, and so clearly explained. Many thanks :)
Very nice video. The part where you explained where alpha and beta came from felt like a plot twist (because I had seen the prequel to this video). Please continue making videos like these!
Found very very helpful in understanding complex things in very easy and simple manner. YOU made my day :)
This is amazing content. Really enjoyed the work
Amazing, thank you so much. Would love to see a follow up explaining Collapsed Gibbs sampling.
What an awsum explanation... Thank you very much!!
Wonderful, and so clearly explained. Many thanks
You made me finally understand LDA after so many times of reading many materials 😂 Thank you.
Thank you very much for the clear and simple explanations !!
excellent explanation of LDA and Dirichlet distribution. Thank you very much
Man, for real, you explain better than 99,8% of the entire world (including the Indian hacks that help us w/ python)
Thank you for this video. I watched the first video on LDA, well presented. very talented with your pictorial presentation.
Pls do a video on Metropolis Hastings and Gibbs Gibbs sampling. I am yet to find a best video on the MH with Gibbs sampling. Your intuitive approach was marvelous.
This is a brilliant video, thank you so much!
Absolute genius. You can explain string theory to a baby. I wish I met you earlier in my life. Waiting more videos, especially maths based
Amazing! very well explained. Just outstanding. Thanks a lot!
So good. Much appreciated.
Thanks very much!! This is very helpful!!
Amazing explanation!
you videos is remarkable!
Fantastic explanation. Thank you
Appreciate a great video. Thanks a lot!
Thank you so much. You explain very well!
Thanks Serrano, You made LDA simple. Can you please make next part for Variational Inference and also the different variation/modification of LDA.
Super clear explanation, thx so much
Thanks for the information. It helped me a lot
Great content, thanks!
Too good, You are superb!
Thank you! this is a great video !!
Great video! thanks!
Great explanation
Very clear explaination thank you so much
Awesome teaching!!! Made LDA look very simple. Could you please make a video on Gaussian Mixture model
Very well explained
Amazing explaination
Brilliant!!
That is sick!
amazingly clear.......
amazing content thank you!
GOAT, a colombian legend
Thanks so much!
I thought that I understood LDA, until now. Luis Serrano is the Richard Feynman of Machine Learning. Thanks,
I understand the analogy of the room where you iteratively bring objects together assuming the others are already correctly positioned. But how did you know in the first place that the coat hanger and the pants had the same color? How do you know what the color of the words is?
Thank you so much
hahha, exellent illustration the language of machine! human cannot understand what it means but we all reach the same goal in the end. Stunning!
awesome
thank you for this, I have been looking for a simple example of LDA with a visual element to show how the process works. However, in your example the "random" colors are actually very close to an accurate model. I would have loved to see it with a start point that was not pre-biased toward the correct answer.
My doubt is at what point in this algorithm do we compare the generated docs and the actual docs.
It is awesome! Could you explain more about dividers in Gibbs sampling (in demo)?
I didn't expected quechua, hahaha.. que grande q eres!
Thank you so much for this explanation. Very intuitive. I have one more question about this tiny alpha and beta add to avoid zero-out. Since we have no idea what these two 'real dirichlet' look, how could we estimate or determine the value for that? Appreciate it!
Thanks for the video! Do you have any resources to read about LDA and Gibbs sampling? Would specifically like to learn more about the math behind it
Very interesting!
How is this different from Naive Bayes based document classification?
you need labels for naive bayes, its supervised
thank you sir, a great video, how do you come up with best values for alpha and beta?
Very nicely explained. How about cosine similarity vs Gibbs sampling for topic modelling ?
This has to be the best video.
Do you think i can use this to organize files using the metadata and name?
Wonderful explanation, still can't understand the equations though. Having great difficulty with that leap.
Very sharp explanation. How to build search a documents with lda? How to get best query which could locate cluster of documents?
This video explains I need to know about LDA but one small doubt @4:08min of the video. How did you assign color to each word? Is it random or some rules we need to follow.
I wish you were my teacher!
Thanks Luis.. Awsum work.
Just had one confusion. at @4:13 you assigned the topics to the words. Is it assigned randomly or is there any logic behind this?
Thanks again for the amazing work.
i think it was assigned randomly, and then will be update with the gibbs sampling process until get the right one
How do you assign colors to each words at the very beginning? Is it random or follows some distribution?
I believe it’s uniformly random.
Could say more about the situation when articles have different number of words? In the previous video you mentioned Poisson distribution, but could you develop this topic, please?
Hi, one question I have is how does Gibbs sampling eventually correctly label every word after many iteration? In the tidying up a room example, we were able to put things together because we know the association between shirt and pants, but for Gibbs sampling, we randomly assign every word a label. In other words, a more accurate example would be we have all these objects in the room and we randomly assign them labels, how do we sort them now? Thanks for your help!
Can you do a video for LDA and Gibbs Sampling implementation using python to identify topics like in this video?
at 12-48 when we try to allocate a color to the word 'ball', we then check what colours are there in this document and what are coolers of word 'ball' in other documents. But where do we take these colours? Considering we just started with all black words in all documents. Do we initialise them randomly and then they converge to more definitive colours?
are those four documents in the example the documents generated by the "LDA machine" in the first video?
I have a question, while documents can be monochromatic, but same cannot be said for words as they can have different contexts. For example, 'war' can refer to either world war I or world war II... what can be done if we run into this situation?
Will the algorithm converge to the true posterior when frequency stop changing significantly if we assign the colour with highest probability instead of sampling a colour from all the topic probabilities computed?
Would this correspond to MAP vs Bayesian approaches to LDA training?
Jeeeez this is an awesome tutorial! Amazing work, thank you very much.
faith_in_humanity += 1
What is the state of art of choosing number of topics K? Cross validation AIC/BIC? or use non parametric method?
Now with all these advanced neural architectures (transformers etc) does this technique is considered depreciated?🤔
It look like the more prevailing a word is, the more monochromatic that word is. Frequent words tend to reinforce themselves. However, each of them might have more different meanings than obsolete words do. How to get rid of such bias?
How do you stop the words from all being the same colour? That would solve both learning objectives
just look into the dirichlet distribution, we can color right away all words in a document. Why we need to reference other documents to color each word in one document??
what do you mean by parameters @18:02, its really confusing how you connect the math with the layman explaination, would be great if you can elaborate in the comments
u r ~ human ur god!
Actually, something seems, for me, not right here, but probably (or definitely) it is right, but I need some explanation. The very first coloring of the words is completely random (4:35) and based on this, the rest coloring happens.. does it makes sense? Since the very first coloring was random, it could have been completely wrong, so why is it a benchmark for the further coloring? Please, could you or anyone else explain it to me?
I have this same question.