This is a great video (as are the other 2) but one thing that needs to be clarified is that the embeddings themselves do not change (by attention @10:49). The gravity pull analogy is appropriate but the visuals give the impression that embedding weights change. What changes is the context vector.
Your videos in the LLM uni are incredible. Builds up true understanding after watching tons of other material that was all a bit loose on the ends. Thank you!
Truly amazing video! The published papers never bother to explain things with this level of clarity and simplicity, which is a shame because if more people outside the field understood what is going on, we may have gotten something like ChatGPT about 10 years sooner! Thanks for taking the time to make this - the visual presentation with the little animations makes a HUGE difference!
This is one of the best videos on UA-cam to understand ATTENTION. Thank you for creating such outstanding content. I am waiting for upcoming videos of this series. Thank you ❤
This was hands down the best explanation I've seen of attention mechanisms and multi head attention --- the fact I'm able to use these words in this sentence means I understand it
I appreciate your videos, especially how you can apply a good perspective to understand the high level concepts, before getting too deep into the maths.
This is one of the clearest, simplest and the most intuitive explanations on attention mechanism.. Thanks for making such a tedious and challenging concept of attention relatively easy to understand 👏 Looking forward to the impending 2 videos of this series on attention
This clarifies EMBEDDED matrices : - In particular the point on how a book isn't just a RANDOM array of words, Matrices are NOT a RANDOM array of numbers - Visualization for the transform and shearing really drives home the V, Q, K aspect of the attention matrix that I have been STRUGGLING to internalize Big, big thanks for putting together this explanation!
I always struggled with KQV in attention paper. Thanks a lot for this crystal clear explanation! Eagerly looking forward to the next videos on this topic.
Thank you for making this video series for the sake of a learner and not to show off your own knowledge!! Great anecdotes and simple examples really helped me understand the key concepts!!
Omg this video is on a whole new level . This is prolly the best intuition behind the transformers and attention. Best way to understand. I went thro' a couple of videos online and finally found the best one . Thanks a lot ! Helped me understand the paper easily
If I understand correctly, the transformer is basically a RNN model which got intercepted by bunch of different attention layers. Attention layers redo the embeddings every time when there is a new word coming in, the new embeddings are calculated based on current context and new word, then the embeddings will be sent to the feed forward layer and behave like the classic RNN model.
These videos where you explain the transformers are excellent. I have gone through a lot of material however, it is your videos that have allowed me to understand the intuition behind these models. Thank you very much!
Thank you for your explanation! I've always wondered why the attention mechanism in Transformers produces more effective embeddings compared to Word2Vec, and your video clarified this well. Word2Vec generates static embeddings, meaning that a word always has the same representation, regardless of the context in which it appears. In contrast, Transformers create context-dependent embeddings, where the representation of a word is influenced by the words around it. This dynamic approach is what makes Transformer embeddings so powerful.
Word embeddings Vectorial representation of a word. The values in a word embedding describe various features of the words. Similar words' embeddings have a higher cosine similarity value. Attention The same word may mean different things in different contexts. How similar the word is to other words in that sentence will give you an idea as to what it really means. You start with an initial set of embeddings and take into account different words from the sentence and come up with new embeddings (trainable parameters) that better describe the word contextually. Similar/dissimilar words gravitate towards/away from each other as their updated embeddings show. Multi-head attention Take multiple possible transformations to potentially apply to the current embeddings and train a neural network to choose the best embeddings (contributions are scaled by how good the embeddings are)
7:00 even with word embedding, words can be missing context and there’s no way to tell like the word apple. Are you taking about the company or the fruit? Attention matches each word of the input with every other word, in order to transform it or pull it towards a different location in the embedding based on the context. So when the sentence is “buy apple and orange” the word orange will cause the word apple to have an embedding or vector representation that’s closer to the fruit 8:00
I subscribe your channel immediately after watching this video, the first video I watch from your channel but also the first making me understand why embedding needs to be multiheaded. 👍🏻👍🏻👍🏻👍🏻
I didn't get it on why do we add linear transformation like earlier too we had embeddings in other planes then why do shear transformation ? Please someone answer
I've really enjoyed with that way of you described and demonstrated matrices as linear transformations. Thank you! Why, because I like Linear Algebra 😄
Unless I'm mistaken, I think the linear transformations in this video incorrectly show the 2D axis as well as the object changing position, but in fact the 2D axis would stay exactly the same but with the 2D object rotating around it for example.
This is an great explanation of attention mechanism . I have enjoyed your maths for machine learning on coursera. Thank you for creating such wonderful videos
Paraphrase: we weigh each embedding by its score, and then add up all these weighted embeddings to obtain a really good embedding. Question to think about: why not just take the best embedding? Is it because averaging improves robustness to noise?
That is a great question! Yes, one thing is because of robustness. Also, each embedding may capture different things, one could be good for a certain topic (say, fruits) but terrible at others (say, technology). Another reason is because of continuity. Let's say that you have embedding A, which has the highest score. The moment embedding B gets a higher score, you would switch abruptly from A to B, which creates a jump discontinuity. If you take the average, instead, you would smoothly go from, say 0.51*A + 0.49*B, into 0.49^A + 0.51*B, which is very similar.
Maybe the next video will clarify how the weighting is achieved. At first I thought the V matrix provides the weighting of the different embeddings, but now I am not sure.
@@tantzer6113 yes! I thought the exact same thing, but then someone showed me they it doesn’t, those weights are recorded inside the transformer. I’m seeing that the V matrix is another embedding in which the transformation is made (and the K and Q are used to find the distances). But I’ll clarify this more in the next video.
First of all thank you for making these great walkthroughs of the architecture. I would really like to support your effort on this channel. let me know how I can do that. thanks
Thank you so much, I really appreciate that! Soon I'll be implementing subscriptions, so you can subscribe to the channel and contribute (also get some perks). Please stay tuned, I'll publish it here and also on social media. :)
Thanks for your great effort to make people understand it. I, however, would like ask one thing such that you have explained V is the scores. scores of what? My opninion is that the V is the key vector so that the V makes QKT matrix to vector space again. Please make it clear for better understanding. Thanks!
I don't quite get at which stage the word "apple" is split into "apple (tech)" and "apple (fruit)". For both sentences the apple gets tokenized as a number. During training this number will be either shifted towards fruits or towards the tech space. In the video it is shown as two items (tech and fruit). But I don't quite get how this is done.
Thanks for the amazing videos! I am eagrly waiting for the third video. If possible please do explain the bit how the K,Q,V matrices are used on the decoder side. That would be great help.
0:55 I thought Attention mechanisms had been around for a while before this paper, e.g. Bahdanu et Al (2014) and likely even earlier than that in some form, and this paper really served as i) an illustration that attention was...well, all you needed and ii) the introduction of the Transformer model architecture?
At last someone explained the meaning of Q, K and V. I read original article and it just says "Ok, let's have 3 additional matrix Q, K and V to transform input embedding" ... What for? Thanks for explanation, this video really helps!
I have been reading the "attention is all you need" paper for like 2 years. Never understood it properly like this ever before😮. I'm so happy now🎉
This is a great video (as are the other 2) but one thing that needs to be clarified is that the embeddings themselves do not change (by attention @10:49). The gravity pull analogy is appropriate but the visuals give the impression that embedding weights change. What changes is the context vector.
Your videos in the LLM uni are incredible. Builds up true understanding after watching tons of other material that was all a bit loose on the ends. Thank you!
Best teacher on the internet, thank you for your amazing work and the time you took to put those videos together
Truly amazing video! The published papers never bother to explain things with this level of clarity and simplicity, which is a shame because if more people outside the field understood what is going on, we may have gotten something like ChatGPT about 10 years sooner! Thanks for taking the time to make this - the visual presentation with the little animations makes a HUGE difference!
This is one of the best videos on UA-cam to understand ATTENTION. Thank you for creating such outstanding content. I am waiting for upcoming videos of this series. Thank you ❤
I love your clear, non-intimidating, and visual teaching style.
Thank you so much for your kind words and your kind contribution! It’s really appreciated!
This channel is uderrated, your explainations is the best among other channels
What a beautiful way of explaining "Attention Mechanism". Great job Serano
absolutely loved the last part with explaining linear transformations of query key and values. thank you so much!
That was awesome, Thank you.
You saved me a lot of time reading and watching none-sense videos and texts
.
THE best explanation of this concept. That was genuinely amazing.
So glad to see you're still active Luis ! You and Statquest's Josh Stamer really are the backbone of more ml professionals than you can imagine
Just THANK YOU. This is by far the best video on the attention mechanism for people that learn visually
This was hands down the best explanation I've seen of attention mechanisms and multi head attention --- the fact I'm able to use these words in this sentence means I understand it
I appreciate your videos, especially how you can apply a good perspective to understand the high level concepts, before getting too deep into the maths.
This is one of the clearest, simplest and the most intuitive explanations on attention mechanism.. Thanks for making such a tedious and challenging concept of attention relatively easy to understand 👏 Looking forward to the impending 2 videos of this series on attention
best description ever! easy to understand. I've been suffered to understanding attention. Finally I can tell I know it!
The way you break down these concepts is insane. Thank you
This clarifies EMBEDDED matrices :
- In particular the point on how a book isn't just a RANDOM array of words, Matrices are NOT a RANDOM array of numbers
- Visualization for the transform and shearing really drives home the V, Q, K aspect of the attention matrix that I have been STRUGGLING to internalize
Big, big thanks for putting together this explanation!
I always struggled with KQV in attention paper. Thanks a lot for this crystal clear explanation!
Eagerly looking forward to the next videos on this topic.
Thank you for making this video series for the sake of a learner and not to show off your own knowledge!! Great anecdotes and simple examples really helped me understand the key concepts!!
Omg this video is on a whole new level . This is prolly the best intuition behind the transformers and attention. Best way to understand. I went thro' a couple of videos online and finally found the best one . Thanks a lot ! Helped me understand the paper easily
If I understand correctly, the transformer is basically a RNN model which got intercepted by bunch of different attention layers. Attention layers redo the embeddings every time when there is a new word coming in, the new embeddings are calculated based on current context and new word, then the embeddings will be sent to the feed forward layer and behave like the classic RNN model.
Can anyone confirm this?
One of the best intuitions for understanding multi-head attention. Thanks a lot!❣
These videos where you explain the transformers are excellent. I have gone through a lot of material however, it is your videos that have allowed me to understand the intuition behind these models. Thank you very much!
Thank you for your explanation! I've always wondered why the attention mechanism in Transformers produces more effective embeddings compared to Word2Vec, and your video clarified this well. Word2Vec generates static embeddings, meaning that a word always has the same representation, regardless of the context in which it appears. In contrast, Transformers create context-dependent embeddings, where the representation of a word is influenced by the words around it. This dynamic approach is what makes Transformer embeddings so powerful.
I really enjoyed how you give a clear explanation of the operations and the representations used in attention
Fantastic !!! The explanation itself is a piece of art.
The step by step approach, the abstractions, ... Kudos!!
Please more of these
This is such a good, clear and concise video. Great job!
Wow, clearest example yet. Thanks for making this!
Wooow thanks so much. You are a treasure to the world. Amazing teacher of our time.
The most easy to understand video for the subject I've seen.
Great explanation. After watching a handful of videos this one really makes it real easy to understand.
This is the most amazing video on "Attention is all you need"
Excellent video. Best explanation on the internet !
best explanation of embeddings I've seen, thank you!
this video is really teaching you the intuition. much better than the others I went through that just throw formula to you. thanks for the great job!
I did not even realize this video is 21 minutes long. Great explanation.
Kudos to your efforts in clear explanation!
amazing explanation Luis. Can't thank you enough for your amazing work. You have a special gift to explain things. Thanks.
Word embeddings
Vectorial representation of a word. The values in a word embedding describe various features of the words. Similar words' embeddings have a higher cosine similarity value.
Attention
The same word may mean different things in different contexts. How similar the word is to other words in that sentence will give you an idea as to what it really means.
You start with an initial set of embeddings and take into account different words from the sentence and come up with new embeddings (trainable parameters) that better describe the word contextually. Similar/dissimilar words gravitate towards/away from each other as their updated embeddings show.
Multi-head attention
Take multiple possible transformations to potentially apply to the current embeddings and train a neural network to choose the best embeddings (contributions are scaled by how good the embeddings are)
Hey Louis, you are AMAZING! Your explanations are incredible.
Thanks Luis, been following your contents for a while. This video about attention mechanism is very intuitive and easy to follow
7:00 even with word embedding, words can be missing context and there’s no way to tell like the word apple. Are you taking about the company or the fruit?
Attention matches each word of the input with every other word, in order to transform it or pull it towards a different location in the embedding based on the context. So when the sentence is “buy apple and orange” the word orange will cause the word apple to have an embedding or vector representation that’s closer to the fruit
8:00
Deep respect, Luis Serrano! Thank you so much!
El mejor video que he visto sobre la materia. Muchísimas gracias por este gran trabajo.
Well the gravity example is how I understood this after a long time. you are true legend.
Nicely done! This gives a great explanation of the function and value of the projection matrices.
amazing, love your channel. It's certainly underrated.
What a great explanation on this topic! Great job!
I subscribe your channel immediately after watching this video, the first video I watch from your channel but also the first making me understand why embedding needs to be multiheaded. 👍🏻👍🏻👍🏻👍🏻
You're my fav teacher. Thank you Luis 😊
This is amazingly clear! Thank for your your work!
It's so great, I finally understand these qkvs, it bothers me so long. Thank you so much !!!
Great explanation with the linear transformation matrices. Thanks!
This video helps to explain the concept in a simple way.
Thanks, the explaination is so intuitive. Finally understood the idea of attention.
Amazing explanation Luis! As always...
Merci Louis! :)
I didn't get it on why do we add linear transformation like earlier too we had embeddings in other planes then why do shear transformation ? Please someone answer
It was the most useful video explaining attention mechanism. Thank you
I watched a lot about attentions. You are the best. Thank you thank you. I am also learning how to explain of a subject from you 😊
Amazing video... Thanks sir for this pictorial representation and explaining this complex topic with such an easy way.
Very impressed with this channel and presenter
I've really enjoyed with that way of you described and demonstrated matrices as linear transformations. Thank you! Why, because I like Linear Algebra 😄
Unless I'm mistaken, I think the linear transformations in this video incorrectly show the 2D axis as well as the object changing position, but in fact the 2D axis would stay exactly the same but with the 2D object rotating around it for example.
Luis Serrano you have a gift for explain! Thank you for sharing!
Outstanding, thank you for this pearl of knowledge!
Amazing! Loved it! Thanks a lot Serrano!
What a great video man!!! Thanks for making such videos.
Great video and very intuitive explenation of attention mechanism
Explained very well. Thank you so much.
This is an great explanation of attention mechanism . I have enjoyed your maths for machine learning on coursera. Thank you for creating such wonderful videos
Great explanation. Thank you very much for sharing this.
You are great at teaching Mr. Luis
Wooow. Such a good explanation for embedding. Thanks 🎉
Incredible explanation. Thank you so much!!!
Thank you so much for the attention to the topic!
Thanks! Lol, I see what you did there! :D
Sir please tell why did you do a shear transformation , i watched it like 10 times but still didn't geti it
This is amazing explanation! Thank you so much 🎉
Thanks for sharing. Your videos are helping me in my job. Thank you.
Paraphrase: we weigh each embedding by its score, and then add up all these weighted embeddings to obtain a really good embedding. Question to think about: why not just take the best embedding? Is it because averaging improves robustness to noise?
That is a great question! Yes, one thing is because of robustness. Also, each embedding may capture different things, one could be good for a certain topic (say, fruits) but terrible at others (say, technology).
Another reason is because of continuity. Let's say that you have embedding A, which has the highest score. The moment embedding B gets a higher score, you would switch abruptly from A to B, which creates a jump discontinuity. If you take the average, instead, you would smoothly go from, say 0.51*A + 0.49*B, into 0.49^A + 0.51*B, which is very similar.
Thanks for the answer, and for the wonderful video.
Maybe the next video will clarify how the weighting is achieved. At first I thought the V matrix provides the weighting of the different embeddings, but now I am not sure.
@@tantzer6113 yes! I thought the exact same thing, but then someone showed me they it doesn’t, those weights are recorded inside the transformer. I’m seeing that the V matrix is another embedding in which the transformation is made (and the K and Q are used to find the distances). But I’ll clarify this more in the next video.
Excellent explanation. Thank you very much.
First of all thank you for making these great walkthroughs of the architecture. I would really like to support your effort on this channel. let me know how I can do that. thanks
Thank you so much, I really appreciate that! Soon I'll be implementing subscriptions, so you can subscribe to the channel and contribute (also get some perks). Please stay tuned, I'll publish it here and also on social media. :)
Superb explanation❤ please make more videos like this
Wow wow wow! I enjoyed the video. Great teaching sir❤❤
Thanks. I saw also your "Math behind" video, but still missing the third in the series.
Thanks! The third video is out now! ua-cam.com/video/qaWMOYf4ri8/v-deo.html
amazing explanation! What software is used to make the visuals (graphs, transformations etc.) Thanks!
Thank you so much! I use Keynote for the slides.
Thanks for your great effort to make people understand it. I, however, would like ask one thing such that you have explained V is the scores. scores of what? My opninion is that the V is the key vector so that the V makes QKT matrix to vector space again. Please make it clear for better understanding. Thanks!
Outstanding video. Amazing to gain intuition.
I don't quite get at which stage the word "apple" is split into "apple (tech)" and "apple (fruit)".
For both sentences the apple gets tokenized as a number. During training this number will be either shifted towards fruits or towards the tech space. In the video it is shown as two items (tech and fruit). But I don't quite get how this is done.
Thanks for the amazing videos! I am eagrly waiting for the third video. If possible please do explain the bit how the K,Q,V matrices are used on the decoder side. That would be great help.
0:55 I thought Attention mechanisms had been around for a while before this paper, e.g. Bahdanu et Al (2014) and likely even earlier than that in some form, and this paper really served as i) an illustration that attention was...well, all you needed and ii) the introduction of the Transformer model architecture?
Thank you for this amazing explanation
Amazing explanation 🎉
Amazing video, thank you very much for sharing!
Was the example of attention using the apples self-attention or just attention?
Yeah!!!! Looking forward to the second one!! 👍🏻😎
Thank you so much for making these videos!
At last someone explained the meaning of Q, K and V. I read original article and it just says "Ok, let's have 3 additional matrix Q, K and V to transform input embedding" ... What for? Thanks for explanation, this video really helps!