Transformers for beginners | What are they and how do they work

Рет қаралды 148,288

Күн бұрын

This week we’re looking into transformers. Transformers were introduced a couple of years ago with the paper Attention is All You Need by Google Researchers. Since its introduction transformers has been widely adopted in the industry.
Get your Free Token for AssemblyAI Speech-To-Text API 👇
www.assemblyai.com/?...
Models like BERT, GPT-3 made groundbreaking improvements in the world of NLP using transformers. Since then model libraries like hugging face made it possible for everyone to use transformer based models in their projects. But what are transformers and how do they work? How are they different from other deep learning models like RNNs, LSTMs? Why are they better?
In this video, we learn about it all!
Some of my favorite resources on Transformers:
The original paper - arxiv.org/pdf/1706.03762.pdf
If you’re interested in following the original paper with the code - nlp.seas.harvard.edu/2018/04/0...
The Illustrated Transformer - jalammar.github.io/illustrate...
Blog about positional encodings - kazemnejad.com/blog/transform...
About attention - Visualizing A Neural Machine Translation Model - jalammar.github.io/visualizin...
Layer normalization - arxiv.org/abs/1607.06450
Some images used in this video are from:
colah.github.io/posts/2015-08...
jalammar.github.io/visualizin...
/ how-to-easily-build-a-...
/ elegant-intuitions-beh...

Пікірлер: 145

@pierluigiurru962 8 ай бұрын

This is clearest explanation of transformers I’ve found so far, and I personally have seen many trying to wrap my head around them. No skimming over details. Very well done!

@ashermai2962 2 жыл бұрын

This channel deserves more views and likes

@AssemblyAI 2 жыл бұрын

Thank you Asher!

@sarc007 Жыл бұрын

I agree

@Yaddu143 Жыл бұрын

I really want you talk about attention. Thank you, shinning in this video.

@testing3562 7 ай бұрын

I am a programmer, I have created many tools that were actually very useful. I even claim that I have 10+ years experience. But I feel very bad to realize that I am so dumb that I did not understand anything after the first 10 minutes of the video.

@sebastianmata9603 3 ай бұрын

They explain it with apples and pears but is actually a very mathematical and elaborated process, if you're not the kind of person that can remember easily how work the sine and cosine functions and do matrix multiplication for fun, is just a little bit harder to get it

@moonlight-td8ed 2 ай бұрын

BRUH JUST REWATCH IT AGAIN... THE VIDEO IS A 10/10

@moeal5110 9 ай бұрын

This is most clear and resourceful video I've seen. Thank you for your hard work and for sharing these resources

@yourshanky Жыл бұрын

Excellent explanation !! Sharp and clear. Thanks for sharing this.

@sivad2895 9 ай бұрын

The best video on transformer architecture with great explanations and charming presentation.

@reshamgaire4188 9 ай бұрын

Finally found a perfect video that cleared all my confusions. Thank you so much ma'am, may god bless you 🙏

@bdoriandasilva Жыл бұрын

Great video with a clear explanation. thank you!

@anandanv2361 Жыл бұрын

The way you explained the concept was awesome. It is very easy to follow.👍

@dooseobkim2100 11 ай бұрын

You are my savior for being actually able to get ready to read all of those AI related papers which I’m completely unaware of. I was stuck at the part of my thesis which I have to provide theoretical background of ChatGPT. As a business student I’m super grateful to learn these knowledges in computer science through your short lecture👍👍

@nikhilshrestha4711 Жыл бұрын

really love how you described the model. easier to understand 🙌

@AssemblyAI Жыл бұрын

Glad it was helpful!

@moonlight-td8ed 2 ай бұрын

cleanest and most informative video ever.. covered whole attention is all you need paper in 19 mins.. damn.. thank you MISRA TURP and assembly ai

@mohamadhasanzeinali3674 Жыл бұрын

I saw numerous videos about Transformers architecture. In my opinion, your video is the best among them. Appreciate that.

@AssemblyAI Жыл бұрын

Thank you, that is great to hear. :)

@dannown Жыл бұрын

This is a really lovely video -- very specific and detailed, but also followable. Thanks!

@AssemblyAI Жыл бұрын

Glad it was helpful!

@stevemassicotte4068 Жыл бұрын

@16:14,, the binary table is wrong, there are two sevens. The second column should start with 8 and not a second 7. Attention is all you need ;) Thanks for the video !

@BCSEbadulIslam 4 ай бұрын

Came here to comment the same 👍

@GeorgeZoto 9 ай бұрын

Great and both low and high level descprition of transformers, thank you for creating this useful resource :)

@otsogileonalepelo9610 Жыл бұрын

Just WOW! You broke down these concepts nicely. Thank you. Live long and prosper 🖖🖖

@AssemblyAI Жыл бұрын

Thank you!

@nikhil182 Жыл бұрын

Thank you so much!💓this has to be the best introduction video to Transformers. We are planning to use Transformers for our Video Processing project.

@AssemblyAI Жыл бұрын

Glad it was helpful!

@mudasserqayyom9947 2 ай бұрын

I'm watching lot of videos of Transformers, But that is exactly I want. Thank You So Much Ma'am. And also AssemblyAl.

@abinav92 Жыл бұрын

Best video on intro to transformers!!!

@kalyandey5195 5 ай бұрын

Awesome!! crystal clear explanation!!!

@andybrice2711 2 ай бұрын

Positional encodings are not that weird when you think of them as being similar to the hands on a clock: It's a way of representing arbitrarily long periods of time, within a confined space, with smooth continuous movement and no sudden jumps. Picture the tips of clock hands. Their vertical position follows a sine wave, their horizontal position follows a cosine wave. And we add precision with more hands moving at different speeds.

@imagnihton2 Жыл бұрын

This made the concept sound incredibly simple compared to some other sources... Amazing!

@AssemblyAI Жыл бұрын

Great to hear, thank you!

@pyaephyo3633 Жыл бұрын

i love it. Your explanation is easy to understand.

@geekyprogrammer4831 2 жыл бұрын

This high quality video deserves a lot more views!

@AssemblyAI 2 жыл бұрын

Thank you!

@goelnikhils Жыл бұрын

Amazing Explanation. Vow. Thanks a lot

@PeterKoman Жыл бұрын

Finally a transformer video that actually explains the theory in understandable way. Many thanks.

@AssemblyAI Жыл бұрын

That's great to hear, thank you Peter!

@malayali_thaaram Жыл бұрын

Yes!!! I agree! Finally!

@archowdhury007 Жыл бұрын

Beautifully explained. Loved it. First time I understood the transformer model so easily. Great work. Please keep creating more such content. Thanks.

@carlosroquesuarezgurruchag8681 Жыл бұрын

Thx for the time. Very clear the explanation

@sanketdeshmukh491 Жыл бұрын

Thank You for in depth explanation. Kudos!!!

@AssemblyAI Жыл бұрын

You're very welcome!

@maryammoradbeigi6690 Жыл бұрын

Incredible explanation on the transformer... Amazing video. Thanks a lot

@AssemblyAI Жыл бұрын

Glad you liked it!

@hussainsalih3520 Жыл бұрын

amazing keep doing this amazing tutorials :)

@amitsingh7684 2 ай бұрын

very nicely explained with clear details

@shubham-pp4cw 2 жыл бұрын

clear explanation of quiet complex topic and explained easily in shorted period time

@AssemblyAI 2 жыл бұрын

Glad to hear you liked it!

@vivekpetrolhead 6 ай бұрын

Best explanation for beginners I've seen besides statquest

@_joshwalter_ 11 ай бұрын

This is phenomenal!

@Zulu369 Жыл бұрын

This video is the best technical explanation I have seen in years. Although Transformers are a breakthrough in the field in NLP, I am convinced that they do not describe completely and satisfactorily, the way humans process language. For all civilizations, spoken language predates written language in communications. Those who do not read and write, still communicate clearly with others. This means humans do not represent natural language in their brains in terms of words, syntax and position of tokens but rather in terms of symbols, images and multimedia shows that make up stories we relate to. Written language comes only later as an extra layer of communication to express transparently these internal representations that we carry within ourselves. If AI is able to access and decode these internal representations, then the written language, the extra layer, becomes a lot easier to understand, organize, and put on paper with simple techniques rather than using these intricate Transformers that I consider as temporary and unnatural ways of describing natural languages.

@rokljhui864 Жыл бұрын

Your idea is represented above , in words, existing separately from your mind. Surely most intelligence is contained within written language, mathematical expression and images.

@Zulu369 Жыл бұрын

@@rokljhui864 As I explained above, written words make up THE extra layer that is actually not necessary once you learn more persuasive communications techniques.

@evetsnilrac9689 7 ай бұрын

@@rokljhui864 "Surely" is not how you start an intelligent hypothesis. You must explain the rationale for your belief since it is not at all readily apparent that the intelligence to process written language was not already in our brains so that we could conceive of and learn written language.

@evetsnilrac9689 7 ай бұрын

This is a crucial point to understand for all of us interested in fully harnessing what we perceive to be the true potential of this technology. I would start with the Adamic symbol-based language.

@bysedova 9 ай бұрын

Please make a detailed video about self-attantion! Thank you for your explanation! I like you haven't used difficult math terms and you have tried to explain for understanding with easy material supply.

@krishnakumarik208 8 ай бұрын

VERY GOOD EXPLANATION.

@VaibhavPatil-rx7pc Жыл бұрын

smile and learn and clean explaniation!!!

@amigospot 2 жыл бұрын

Nice video for a fairly complex architecture!

@AssemblyAI 2 жыл бұрын

Thanks Hyder! - Mısra

@jayanthkothapalli9.2 11 ай бұрын

Great work mam. You made it simple to understand.

@wasifrock687 Жыл бұрын

very well explained. thank you!

@AssemblyAI Жыл бұрын

Glad it was helpful!

@keithwins 6 ай бұрын

Thank you that was excellent

@donevo1 Жыл бұрын

very nice presentation! in 12:18 you say that attention is on 8 words. from reading the paper I think that attention is on ALL the words, and 8 is the number of heads: each word vector (D=512) is split to 8, i.e vector dimention in each head is 64.

@AddisuSeteye Жыл бұрын

Amazing explanation. I can't wait to watch your explanation on another AI related topic.

@AssemblyAI Жыл бұрын

More to come!

@thebiggerpicture__ Жыл бұрын

Great video. Thanks!

@AssemblyAI Жыл бұрын

You're welcome :)

@rodi4850 Жыл бұрын

best explanation!

@JayTheMachine 9 ай бұрын

thank you soo much, damn, love your explainations

@devraj241 Жыл бұрын

great video, well explained!

@near_. Жыл бұрын

What's the purpose of output embedding?? What are we feeding in that???

@user-fp5sx2vr4z Жыл бұрын

Thank you for the presentation, it has been so insightful. I wish you made a video about the word embeddings of the transformers. Thanks

@AssemblyAI Жыл бұрын

Great suggestion!

@6001navi 11 ай бұрын

awesome explanation

@juliennoel3061 4 ай бұрын

hi! oh yeah please a specific video on 'attention' 🙂 - And also : 'great job you are doing! Congrats! Thumbs !!'

@0Tyr 2 жыл бұрын

Very informative channel, and well presented..

@AssemblyAI 2 жыл бұрын

Thank you! - Mısra

@niyatisrivastava4-yearb.te820 6 ай бұрын

best explanation

@talktovipin1 5 ай бұрын

Very nice explanation. Incorporating animations into the images while explaining would enhance comprehension and make it even more beneficial.

@MrTheyosyos Жыл бұрын

"attentions for beginners" will be great :)

@salamander7715 10 ай бұрын

Seeing all the comments of people saying that this video made things simple just makes me feel stupid ahah! This video is amazing and the explanations are great, but i can't say i've understood more then 35% of the concepts. I'll have to watch this several times for sure

@rufus9322 Жыл бұрын

Thank you for your video 🤗 How to understanding more details about word embedding method in Transformer model?

@nogur9 11 ай бұрын

Thanks :)

@kartikgadad9285 7 ай бұрын

Thanks for explaining Transformers, can we have a video on Embeddings, seems super interesting. The Positional Encoding part was difficult to understand, as it has been just taken from abstract level, can we find better video on positional encoding?

@user-ft2jx8io9g 9 ай бұрын

geart work, may allah bless you and guide you 🥰🥰😍😍

@user-tt6tg3eb7n 9 ай бұрын

I overall liked the video a lot. I just do not thing is enough to understand the whole concept. Especially masked multi head attention layer was missing and how the actually outcome of the model is created (translation etc)

@lexflow2319 Жыл бұрын

I don't understand why there are 6 decoders and encoders. The diagram shows 1 each. Also, what is the output as input to the decoder. Is that the last output from final softmax

@wenshufan Жыл бұрын

Thank you for explaining the transformer in detail. However, I still don't get how do you train the Q,K,V matrix. The attention mechanism is calculated by from them. What type of feedback/truth can one use to train those matrix values then?

@andersonsystem2 2 жыл бұрын

Good video

@AssemblyAI 2 жыл бұрын

Glad you enjoyed it :)

@actorjohanmatsfredkarlsson2293 Жыл бұрын

Great video. I’m missing how the attation layers: queries, keys and values and the output weights are trainee? Also what was the values matrix for?

@MrAmgadHasan Жыл бұрын

They are trained just like any neural network: we have a loss function that compares the model's output with the desired output, and then this loss is propagated backwards to the weights and biases and we use gradient descent to update the weights. Lookup "back propagation" for more info or just look up"how neural networks are trained"

@helgefredriksen 22 күн бұрын

Hi, could anyone explain how the Feed Forward part of the transformer learns? How does the loss function work? By masking out some of the input from the self-attention part during training and then compare the real value with predicted value?

@rokljhui864 Жыл бұрын

Interesting. Sounds like a Fourier transform; Obtaining a frequency distribution from a time-series, reveals the underlying frequency components and amplitudes. Are you essentially distilling the 'word cycles' from the sentences to obtain meaning from the word patterns across different word combination lengths (from single word to many thousand) And, optimising the predictability of the next word automatically optimises for the appropriate word combination lengths, that align with actual meaning. i.e Understanding 'peaks' are optimised similar to the fundamental frequencies in a Fourier transform. ?

@mbrochh82 Жыл бұрын

I wish someone would explain how exactly the backpropagation works and what values exactly get nudged and tweaked during learning (and by which means)

@ankit9401 2 жыл бұрын

You are awesome and I appreciate your efforts. After watching your video, I can say now I understand the transformer architecture. I have a query. According to original BERT paper, two objectives used during training: Masked Language Model and Next Sentence Prediction. Are these training objectives present in original or all transformer models or they are specifically used for BERT ? I hope you make video to explain attention and BERT model in future 😊

@AssemblyAI 2 жыл бұрын

Great to hear the video was helpful Ankit! These are not the tasks that were in the original transformer model. But I think they are not specific to BERT. Other architectures also use same/similar tasks to train their models. We have a BERT video in the channel by the way. Here it is: kzfaq.info/get/bejne/bMeYq7OFscDchWw.html - Mısra

@strongsyedaa7378 2 жыл бұрын

@@AssemblyAI So instead of using RNN & LSTM we directly use Transformers?

@abrahamowos Жыл бұрын

A question @ 11:30 : if for instance the values v are really large and you multiple them by the results from the softmax layer. Won't the resulting weighted be too high after adding them together?

@JackoMcW Жыл бұрын

I'm not sure I understand your question or what you mean by "too high," but consider that all of those softmax values will be

@amparoconsuelo9451 10 ай бұрын

I have read books and watched videos on Transformers. I still don't understand Transformers. I want to order from Amazon an assembly Transformer kit, work on it and have a Transformer I understand the way I undestand how Lotus 123 and Wordstar were created.

@kellenswain2049 Жыл бұрын

11:06 from reading the paper, 64 is not the square root of the length of QKV vectors, it looks like it is d_model/h where h is the number of heads used in multihead attention. And so then I assume d_model is the length of the QKV vectors?

@RAZZKIRAN Жыл бұрын

thank u

@AssemblyAI Жыл бұрын

You're welcome!

@wp1300 Жыл бұрын

13:35 Positional encoding

@near_. Жыл бұрын

What's the purpose of output embedding?? What are we feeding in that???

@manjz7hm 6 ай бұрын

You explained well , but my brain not digesting it 😂

@near_. Жыл бұрын

What's the purpose of output embedding?? What are we feeding in that???

@EmanueleOlivetti 6 ай бұрын

Around 16:00 the binary representation repeats twice 7 so the right part of the binary encoded numbers is incorrect

@user-bu3ds4dc6d 9 ай бұрын

easiest explanation.

@guimaraesalysson Жыл бұрын

Theres any video about attention mechanism ?

@AssemblyAI Жыл бұрын

Not yet but it's a good idea!

@roshanverma1123 11 ай бұрын

Great simplified content! Thanks! Btw, you look beautiful!

@nirmesh44 5 ай бұрын

make attention video

@titusfx Жыл бұрын

I'm still concern how all these papers don't have any mathematical rigour, there isn't one theorem, there is nothing. And it works....🤯 I can't imagine when the rigourosity start coming in, what would be the results. I'm starting to believe that deep learning is Physics for knowledge 😅

@juanpimentel4567 7 күн бұрын

Why are there 6 encoders and 6 decoders. Someone please explain.

@nikbl4k 28 күн бұрын

great video, very interesting

@marcfruchtman9473 Жыл бұрын

Very interesting and informative. Thank you for providing a very detailed explanation of Transformers. One note: The word "Query" is pronounced like Qw-eerie (USA English). The beginning sounds sort of like the sound of "Quarry", or "Quack" but rhyming with dearie.

@AverageOrdinaryEverydaySupDog 2 күн бұрын

☝ question.... hmmm.. what?

@M7mdal7aj 9 ай бұрын

thanks but the explanation is not detailed enough. but nice explanation for the positional embedding. thanks

@robl39 Жыл бұрын

What is disappointing about this video is that you have to know about or understand 50 other concepts first

@NielsSwimberghe 4 ай бұрын

"You might need to watch this multiple times". You don't say. 😅

@denwo1982 4 ай бұрын

Chatgpt “explain this video to me as if I was an 8 year old”

@JaredEdwardsPM Жыл бұрын

I feel like you just described how a ouija board works…

@AssemblyAI Жыл бұрын

Comment of the year :D

@frizzsupertramp6434 Жыл бұрын

At 16:44 the binary representations on the right side are wrong (number 7 comes twice, should start with 8 on the right side). (Sorry for being anal 😀)

@AssemblyAI Жыл бұрын

Thanks for the heads up! Video editing gets tedious sometimes :)

@prdeshnaser5297 4 ай бұрын

needs pauses in speech, after 15min all i hear is a vector of blblblahs. a good video and human illustration of how text gets generated by machines, but that is not the point.