Visual Guide to Transformer Neural Networks - (Episode 3) Decoder’s Masked Attention

[@HeduAI]

Correction in the Video: A huge shout-out to @jackzead "What comes out of the encoder should be the key and value matrix and what comes out of the bottom part of the decoder is the query matrix." *TIMESTAMPS* Here are the timestamps associated with the concepts covered in this video: 0:00 - Recaps of Parts 1, 2 0:52 - Residual Connections 4:25 - Transformer Decoder 9:52 - Decoder's Masked Attention

[@muhammadsaadmansoor7777]

I was wondering if you would correct this. Thanks.

[@HeduAI]

@@muhammadsaadmansoor7777 I tried but was not able to find a way to add notes to a video which has been published. Apparently, KZbin has a few tools that get enabled to the channel owners with a certain number of subscribers only. Let me know if you know how to add notes to the video.

[@ivanmatas3884]

@@HeduAI You can pin this comment at the top. Excellent video, congratulations.

[@HeduAI]

@@ivanmatas3884 Thanks :) This comment is already pinned to the top.

[@torgoron9695]

@@HeduAI I think it's not anymore pinned to the top - don't know how this works though. Maybe you can also add to the video description that there's this error or the correction. Thanks a lot for this extremely awesome content! Saved me loads of time and was fun to watch at the same time!

[@LeManneCris]

These explanations are so good that they should be included into Tensorflow's and PyTorch's documentations, seriously !

[@HeduAI]

This made my day :) Thanks!

[@koustubhavachat]

True

This series is a gem. Really clear explanation, much easier to get back to the paper and the math with such overview. Huge kudos!

[@ИльяЯгупов-н4я]

absolutely agree!

[@dy8576]

This series deserves so much more "attention"

[@alsonyang230]

I have tried to learn transformer from various sources with several attempts in the past a couple of months, but I couldn't get the gist of it. I wish I have spotted your 3 episodes earlier, they were extremely helpful and easy to follow. I really appreciate the time, effort, love and joke you put into your videos, and sincerely wish you could keep making these types of videos on DL topics. I love the fact that each concept/technique is well explained with a concrete example as well as why it was done this way not the others. You are truely gifted at explaining complicated concept for people with limited DL background. Please keep up the good work

[@pra8495]

Please make series of NLP , Deep Learning. Your teaching style is amazing.

[@HeduAI]

Thank you! I am glad you enjoyed it!

[@ankoor]

You have an amazing gift of explaining complex things in an intuitive way!

[@HeduAI]

Thank you! Glad you found it useful :)

[@azerioauditore511]

Her way of explaining is so crystal clear. Its very evident she is a genius

[@HeduAI]

When Azerio from the Assassin's Creeds calls you a genius, you've gotta agree :D Thanks!

[@ahp-6785]

You are the mother of StatQuest and 3Blue1Brown. Both of these guys are awesome in explaining complex ideas in simple words. But you are the best.

[@ninjahunterx7497]

I don't know about StatQuest (haven't seen his ones) and 3Blue1Brown is good because of the visualization he brings with his advanced animations. But honestly, here she explained all these concepts using simple animations and had a good structure throughout the videos, each connecting well to the other. Very commendable if you ask me.

[@royeyono6512]

You are truly truly gifted in explaining very technical topics in an intuitive way! I often get embarrassed to go to KZbin to learn about DL concepts considering that I have studied the topic for a couple years now, but in all honesty, the scientific community desperately needs people like you. Thank you for investing your time in making the world a better place!

[@HeduAI]

Roy, you truly made my day. Thank you so much!

[@jb_kc__]

Super clear explanations. Really appreciate you putting this series together! (and love the pop culture references)

[@TanmayGejapati]

it hurts that ur skills are so under-rated, the ai community needs more of you!

[@SaiGopalCAP8978]

You are my savior. Went through a bunch of videos yesterday, didn't understand shit.. finally found these awesome short 3 videos today.

[@吳信宏-z7i]

Your three episodes are the clearest and best explanation to Transformer that I've ever come across. They help me a lot and I thank you for that. I can see that you got talent to explain things complicated in a very simplistic way. Looking forward to watching more of them.

[@jackziad]

There might be a mistake in the video. What comes out of the encoder should be the key and value matrix and what comes out of the bottom part of the decoder is the query matrix. No? The query is the matrix that is looking for the right key and value pairs. The rest is very good. I always learn something new in your videos.

[@HeduAI]

Hmm... you maybe right. Let me confirm this and get back in a few days. If it is indeed a mistake, I will try my best to inform future viewers about it by either editing the video or pinning the edit comment to the top. Thanks for the critical feedback!

[@nivu]

The encoder output is used as the Query and Key input to the 2nd Multi-Headed Self Attention (also known as Encoder-Decoder Attention) of the Decoder Block. And the Value comes from the 1st Self-Attention Block of the Decoder. The explanation in the video is correct.

[@jackziad]

@@nivu Well, I have seen Jay Alammar's explanation and I have also seen 2 code implementations of Transformer and all 3 references explain and show that the encoder outputs the key and value matrices and the 1st self-attention block of the decoder outputs the query. Again this makes sense since Key and Value matrices always come in pairs (a key is a kind of an index into the value, they are supposed to be together). The query is the question that asks which key is the most relevant and so it makes sense it comes from the decoder since the decoder knows the ground truth. In case of language translation from english to german, the encoder would study the english langage, the 1st self-attention block of the decoder would study the german language, and the 2nd attention block of the decoder would study the interaction between the 2.

[@jackziad]

This is what Jay Alammar says about this point: "The “Encoder-Decoder Attention” layer works just like multiheaded self-attention, except it creates its Queries matrix from the layer below it, and takes the Keys and Values matrix from the output of the encoder stack."

[@HeduAI]

@@jackziad Thanks! Have added the corrective comment and pinned it to the top, accrediting you. Thanks again for pointing this out.

[@ushas3349]

The most difficult topic "Transformer Network" is explained very clearly with visual guide. Hats off to the speaker. I appreciate the efforts taken to prepare this guided tour.

[@HeduAI]

Thanks a ton Usha! :)

[@salonikothari7494]

i have never heard things being explained so clearly and articulated with all key details as much as possible ! thank you so much .... > :)

[@HeduAI]

Really glad to hear this :)

[@jackskellingtron]

Fantastic. Just finished the series without stopping. "Layer normalization simply means standardizing neuron activations along axis of the features" I've seen several descriptions of layer normalization, and they are usually paragraphs long. Yours is one sentence long and somehow 100x clearer. Animations are perfect as well. Yer a wizard 'edu

[@bendarodes61]

I've watched many videos about transformers, this is by far the best.

[@esn832]

You should get a Nobel Prize for this video series! Thank you very much!

[@HeduAI]

Hahah! :D I gladly accept the Nobel Prize from you :)

[@lazythor1716]

Decided to watch the 3 minutes ad without skipping, cuz you deserve it!

[@HeduAI]

This warmed my heart. Thank you!

[@lazythor1716]

@@HeduAI can you do another video covering attention in image recognition? I would really appreciate it

[@selimcanbayrak3278]

It is amazing how you explain these complicated models in a way that is so simple to understand. It reminds me that Einstein said "If you can’t explain it simply, you don’t understand it well enough."

[@ruslan124]

I don't know who you are but I do know that this series was the best explanation of Transformers I have seen or read and I have seen and read a lot. Thank you.

[@gesitsinggih]

Your video is worth more than 1000 page book. Your 3 videos allow me to understand and implement from scratch Transformer architecture. Any gap in explanation forces me to think for myself, deepening the understanding. Thank you.

[@HughMcBrideDonegalFlyer]

I think this is the best explanation I have ever seen , believe I have looked for quite a while . Amazing Job. Thank you for your efforts

[@HughMcBrideDonegalFlyer]

Ignore the above , I have found it , again your explanation is the best on the web , hands down . Thank you for the time and effort you put into this. Its amazing

[@baskaisimkalmamisti]

Truly it is the clearest explanation of the transformer architecture. Please keep posting that kind of great explanations!

[@picassojeus6114]

You are a ROCKSTAR 💚💚💚. So lucidly explained... Shows your eloquence. Glory to you.

[@paweosek5009]

I was trying to understand it for quite a long time, and for the first time I don’t feel like a child lost in the woods 😂 you really have a gift!

[@HeduAI]

This has got to be the sweetest comment :)

[@ravirajm]

This is the easiest to understand explanation of Transformers with amazing graphics. I would pay for this content :)

[@kieudangvan5692]

I have gone through dozens of explanations, yours is a life saver. You are just a genius Batool, keep it up!

[@HeduAI]

Cheers 🙌

[@zhitaochen5665]

You explain the difference between training and inference so well. This video deserves so many more views.

[@snmemich]

as an expert on education and (human) learning, I would like to tell you that you explanation over those three episodes is outstanding! thanks a lot for your effort! Listening to you was pure joy!

[@HeduAI]

That is truly an honor to know! Thank you :)

[@lalluyoutub]

Finally & finally, I understood Transformers!!! Best playlist that I found on Transformers!!! Thank you @Hedu AI. Hope to see you make more videos.

[@norchiegaming]

Please make more videos explaining stuff intuitively! These are the best videos I've found on the topic and helped tremendously

[@saifuddinnabilmohammedzain9608]

More videos please, if this channel continues at the same level, it will become one of the most important references in the world

[@HeduAI]

Wow, this brightened my day :)

[@mbunooo]

this series is AWESOME and this channel should have more suscribers!!!

[@yassine20909]

You got the magic, of explaining very complicated material and making it so obvious, within you. Thank you for the series on the attention model, you did a Great job. 👍

[@aayushjariwala6256]

Some video includes concept only and other includes maths. But your way to teach is just incredible. You haven't just included maths and concepts but also examples to show each detail without skipping in very less time.

[@HeduAI]

Your comment made my day :)

[@chaitanyachhibba255]

these are the best videos available, I hope millions watch this. I have been through many videos in the past week and I wish I would have seen this sooner. Thanks lot, Please keep on posting more videos..

[@HeduAI]

So glad to hear this!

[@weihongchin8140]

this is my 3rd comments for my entire life using KZbin! The best video for explaining transformer! Hope you will make more videos in future!

[@SaurinShah1]

This is one of the best explanations I've ever seen. Thank you so much!

[@SathvikKullolli]

Amazed by the way of explaination!!.Most intuitive and brilliant articulation of one of the hardest concepts!!A huge THANK YOU!!!

[@fritstuininga8159]

These explaination videos are so good. It would be a crime not to subscribe. You earned yourself a sub today!

[@justdo_it3699]

wow... this video give me perfect understand about transformer. Your teaching style is amazing, perfect, no question left. plz make more video. I'm very thanks for your effort

[@rajv4509]

If I said thank you a million times that would not be enough!!! You have simplified a complex topic in such a way that anyone can understand it. Keep up the good work! The world of AI needs more people like you :)

[@HeduAI]

Thank you so much for your million thank you's :) Should be enough fuel to keep me inspired for the next several videos :)

[@brandonkim4675]

Wow, you absolutely killed it! Your explanation is mind-blowingly easy to follow for even one who does not have any background in AI, like me. I have watched a lot of videos on this topic ever since ChatGPT hit the headline, but none of them managed to make me feel like I understood anything. No video has come close to the way you broke down the process step by step. Your use of visual aids was simply fantastic and made the whole concept so much clearer. Thank you soooooo much.

[@martinhlosta8249]

This was the best explanation of Transformers and Self-Attention I have come across so far! Thanks a lot for this series of videos!

[@HeduAI]

You are very welcome! :)

[@alankarshukla4385]

This is JUST EPIC!! This is the best explanation out there!! I looked everywhere but always seems something is missing. Thank you so much for making this great Video.

[@tinski]

Your series is by far the best explanation of transformers I have come across. Thank you!

[@HeduAI]

Glad to hear this :)

[@arjuntheprogrammer]

One of the best explantion process I have ever seen

[@lassadsmaoui2929]

I passed a long time searchig a good explanation untill i find these series . awesome !! keep going 😍😍😍

[@whitelady1063]

havn't seen such a good series of vids in a realy long time

[@stevehe5713]

Amazing videos! The best out of all the transformer videos ever watched. Thank you!

[@EnderKira73PC]

Please, go on making videos. This is the best content I've found on Transformers

[@yusun5722]

The best detailed explanation of the transformer. Hope to see more of this type videos.

[@munugotiswapna8357]

Your explanation is too good..I am able to understand these transformers because of the way explained

[@ahmadsawalhah9620]

Really I wonder how could you do that, simple , clear , straightforward, beautiful ..... thank you very much,,,, Keep going

[@vsucc3176]

This series was amazing! Thank you very much for these videos, they helped me a bunch :)

[@Ayanshandseals]

I just coded whole Transformer toy implementation with help of these video. Kudos to you!

[@HeduAI]

Sounds great!!

[@HiggsPatch]

Nicely done. Very clear and succinct. Loved the narration, and the acting parts from Game of Thrones. Huge thank you for making me understand Transformers!!🙏

[@HarshitSingh-xx8ey]

These videos were sooo good. You should continue making more

[@corvo1068]

As many others have said, I think your videos are the best explaination for Tranformers on KZbin. I guess/hope your channel is going to explode :D I really hope you can make more videos like that!

[@HeduAI]

You made my day. Thank you so much!

[@saruaralam2723]

Your all videos are best I came across (explaining with matrices, example, etc). Kindly make more contents.

[@TranVanHue-b5n]

Best video series ever! Thanks is not enough.

[@yashrajverma8412]

Best explanation ever, Please continue to upload videos about new research papers in the area. Keep up the good work!

[@maysammansor]

Batool you are really a great teacher. Thanks for the content. Please provide more vid on LLMS and AI

[@subhabratabanerjee3087]

one of the best explanations so far I have visited on transformer , all 3 episodes are really awesome , great work !! Please cover BERT , advanced NLP topics etc in future btw subscribing your channel :)

[@HeduAI]

It appears the hard work put into creating this video paid off! :)

[@rushirajparmar9602]

@@HeduAI It really did, me too! Awesome explanation!! Please keep going!! :)

[@subhabratabanerjee3087]

MIT recently posted video on transformer (kzbin.info/www/bejne/h6fOgmR4aKt1p6M) ,trust me for each sec it remind me your video only , awesome work , request you please a create new video with correction. Thanks !

[@kamalamarepalli1165]

Very very wonderful content, voice, explanation, animations...so good...thank you so much.

[@ruantwice]

This video series is the best one I have found about transformers! Seriously, a job well done! Thank you for your effort!

[@rajarshideeplearning6500]

Literally THE best 3 video playlist 👌👌

[@prajwalanagani5590]

This is a great playlist and one of the best I have come across for transformers. Thanks! And looking forward to your next set of videos.

[@martinsoderstrom449]

Thank you! This is the best guide I have seen on here. :)

[@benak495]

Thank you for all your videos about transformers. They are incredible. It's obvious you've worked hard on them, and they've been really instructive.

[@chirpy_data9885]

Thanks a ton for making this extremely useful and a well explained video. For me there hasn't been any video/blog or course that brought the underlying concepts right from vector spaces all the way to softmax function so succinctly and clearly. Looking forward to many more videos

[@HeduAI]

So glad to hear this :)

[@mehmetakifozkanoglu2259]

The best explanation that I have ever watched, thank you so much

[@HeduAI]

You are very welcome! :)

[@AnuRag-fo2ts]

Wonderfully explained. Looking forward for more

[@yevhendiachenko3703]

Thank you for the great videos! There might a small mistake at 9:02. Probabilities after softmax should sum up to 1 but in your example, their sum is greater.

[@vahanhovhannisyan7603]

This was such an awesome explanation! Thank you very much!

[@mohitmeena541]

This is epic content on youtube. Please keep uploading!

[@_divya_shakti]

finally wait is over , thankyou maam

[@diddydd]

I think the normalization should be done w.r.t. the features (axis=1), but in the video (3:19) it shows a norm w.r.t. the data. It seems from the code I looked at, that tf.keras.layers.BatchNormalization is used which actually by default normalizes the feature columns. Besides this, it was an AMAZING video series, thank you.

[@vjaiic]

Top class content on Transformers! I just refer these episodes every time whenever I need them! Please create more content on other Deep Learning topics! You’ll surely be a top KZbinr one day!

[@HeduAI]

So glad to hear this! :) You made my day.

[@markpadley890]

Notwithstanding jackzead's observation, another very good video. My understanding of this very important NLP architecture is way, way better as a direct result of this entertaining video series - time for celebration indeed - well done!!

[@HeduAI]

It appears the hard work put into creating this video paid off! :)

[@yiweijia6922]

Thank you! It's a really incredible tutorial. Everything is explained in detail and intuitively!! Love your teaching style and it's really worth a subscription. :)

[@goodlifenepal]

You have created so great learning asset. Hats off to you.

[@mydigitalwayia956]

Nunca había visto una explicacion tan excelente como la suya. Muchas felicitaciones. Por favor publique mas videos.

[@Kalel69150]

I am not used to like or comment youtube videos but I would like to tell you that your explanations are amazing. Please make other videos !

[@HeduAI]

It appears the hard work put into creating this video paid off! :)

[@aaryannakhat1842]

Stupendous! Thank you for this intuitive explanation, I was really struggling to get my head around Transformers.

[@shadabalam2122]

The best transformer explanation 👏👏

[@mbunooo]

There is an error in 8:20. As some already pointed out, if you concat all the embeddings, you would get a variable lenght vector. Instead, the linear layer is applied to each embedding and the logits from just the last embedding are taken into account to generate next's word logits.

[@davidrod2865]

Can you explain it a little bit better? would help a lot. Thanks

[@oprangler]

Best explanation on transformers I've found

[@csescholar8890]

mind blowing explanation i have seen ever

[@abilashnair1745]

Best teacher in the world

[@csescholar8890]

its mind blowing explanation i have seen ever

[@leo_prem]

Brilliant, I was lost in the paper, but you saved me.

[@sharjeel_mazhar]

You have ny utmost respect, ma'am!

[@pollopallo7401]

You probably saved me for tomorrow exam, thank you so much!

[@isaaczheng2769]

This serie is so good! PLEASE make more AI tutorials

[@alexandrbyzov3581]

This is an extraordinary good explanation! Thank you very much! Instant subscribe!

[@yogendranayak5585]

if you were my teacher in high school, I would not be the failure I am right now. Thank you so much for uploading these videos.

[@HeduAI]

Transformers are pretty much state of the art in AI. If you are trying to understand these concepts, I can definitely say one thing about you - you are no way a failure :)

[@sajansudhir1859]

@@HeduAI You are a motivator too.Thanks

[@cest_heng]

Best videos I have ever watched. Huge thanks! How long did it take you to make the three videos?

[@HeduAI]

So glad to hear this! Around 2 months :D

[@mohitpilkhan6808]

@@HeduAI Hats off !!!