วีดีโอ

I just find it counter-intutive (it could be just me) when describing the encoder-decoder portion by using the previously Q (search bar)/K (searched items)/V (searched results). When doing translation from English to French, it is like quering in English to find which French words to use? If I put the 'above analogy' among Q, K, V aside and purely looking at it from the Matrix perspective, I think I can see why. First, we need the 'attention" between the English and French words. To get the French text, we apply the 'attention' matrix with the English text (input - V) to perform the 'transformation' to generate the output French. To fulfill the matrix dimension across, we need to position French in Q, English in V and K. Oh, if I transpose the "English to French;' Attention matrix, input the French instead, will I get the English? I guess so but likely improper translation. Sorry about the long text. BTW, very good lecture, many thanks

I was looking for a deep dive to deep learning till I found this on the udlbook website: Video lectures for chapters 1-12 from Tamer Elsayed of Qatar University. I think it's a great combo!

You are an excellent lecturer

Jazakallah Khair. I'm a student at AUS (in the UAE) going through the book. Delighted to find these great video lectures!

Very helpful

Rational mean .do the right thing

From my heart Thank you so much ❤❤❤

مليون لايك يادكتور اسلوبك اكثر من رائع الله يوفقك

I know my suppose is not practical but wouldn't it be better if words are tokenized into +2d tensor instead of vector

اعتزل شرحك سيئ

ليه الشرح بالانجليزي ليه ليه نفسي افهم؟! مفيش مشكلة بالنسبة للكتاب ولا المشكله بالمصطلحات تكون باللغة الام لهذا العلم، لكن الشرح يكون بالانجليزية والله هذي انا ما فهمتها

Excellent lectures

where can I find theory question with answer for this course

ربنا يبارك في حضرتك ويجعل شرحك في ميزان حسناتك 💗💗💗

بارك الله فيك، ما اجمل أن يجتمع الدين مع العلم

رفع الله قدركم. The intentions for learning is fabulous. May Allah reward you for them.

Wow, what an amazing course Prof. Elsayed. Are you planning on continuing with the stable diffusion chapter of Prince's book? That would be awesome :)

Thank You Sir! for providing such a valueable Lectures ....Your each & very word open a new way towards the life .. Thanks again

Assalamu alaykum. Sir the example in the first poll quiz to find out whether a news article is political, economical or sports -- it is multiclass classification but why do we consider it univariate? Like in the previous examples of image segmentation into foreground and background, we considered each pixel value and then found two possible outputs out of them so it was multivariate binary classification problem, but in this news article problem, don't we consider all words indicating it multivariate ? Please can you guide. Jazakumullahu khayran

reinforcement learning ( I will be happy to watch this course from you professor)

This is the first course I completed from start to finish over a period of 3 months time in last 12 years. Thank you so so much Sir. I just hope and wish we can get your next sem course on DS and Information Retrieval :)

May God reward you, sir. From start till the end your lecture was very engaging and full of knowledge. Thank you for teaching us about having good intentions and collective goals as commanded to us in our religion.May Allah Subhanahu wa Ta'ala grant barakah in your knowledge & allow us to learn from you. Aameen

Thanks prof Tamer for this amazing playlist❤ Will you discuss the Residual Networks?

such wonderful lectures professor thank you for sharing them

absolutely exceptional.

First viewer.As usual excellent tutorial

awsome lecture.Best among youtube lectures

Thank you sir for sharing free such a valuable tutorials

Can you upload the slides Sir from Transformers (b) onwards ?

Thank you so so much sir. Every student on this planet deserve a teacher like you. This course motivated me to pursue my Masters from Spring 2025.

جزاك الله عنا خير الجزاء يا دكتور

Sir, a few thoughts: 1. Regarding multi-head attention, we use (D/H) and then concatenate all the outputs to get back the (D x N ) matrix at the end. This process seems perfect, but putting computational efficiency aside, if I use the same (D x N ) matrix size for all {Q, K, V}, can we expect the network to generate a better understanding of the text? This would mean performing transformations within the same vector space, whereas in the actual architecture, we lower the dimensionality to (D/H), which could result in some loss of information. This trade-off is made for computational efficiency and model performance. Is my understanding correct, or am I missing something obvious? 2. Regarding positional encoding, in the book (page 213, section 12.3.1), it says (and I quote): "The positional embedding vector may be added to the network inputs or at every network layer. Sometimes it is added to (X) in the computation of the queries and keys but not to the values." What’s the intuition behind adding positional encoding to queries and keys but not to values? If possible, could you share your thoughts on this in a future lectures? It would be very helpful. Thank you once again. Looking forward for the next class this week :)

أرجو إضافه تكلمه للكورس advance IR ، حيث أن المحاضر يملك تجربه طويله في هذا المجال ونرغب بخبرته

يا ريت تضيف تمارين لتثبيت أو إختبار المعلومه

Thank you sir!

Sir, Can you upload the Transformer lectures ? I guess I told this 19 times in last 19 lectures, but thank you so much once again for sharing your knowledge to the whole world. 🙏🙏

would you suggest me which book i follow?

Very well illustrated lecture. Your students are lucky with you as a very professional professor. I am graduated from 2 years and I which I were one of your students. You helped me alot to understand concepts were very complicated. Very excited to the next lecture and very excited to hands on projects. Good luck!

show of good pedagogical skills, great lecture.

"سير، قم بتقديم تطبيقًا حقيقيًا لمساق هذا حتى أستطيع إنشاء شيء يمكنه مساعدة إخوتي وأخواتي في مواجهة أعداء الإسلام"

Great lecture,Previously I have followed cs224n,But this is among the best

sir wonderful lecture

Thank you!

Sir please share the papers and data

Thank you so much, Sir! Could you please share the resources you mentioned on PyTorch, research papers, and other resources you mentioned as you share in the class? This is the first time I've been able to follow a course from the beginning and keep up while reading the UDL book. Teachers like you are truly special. I’ve started envisioning the day I can meet you in person and thank you directly.

Great work.

نفع الله بكم ..وجعل كلماتكم في ميزان حسناتك ان شاء الله

الله يقويك دكتور.....كلام صادق وواضح أنه من القلب....أسأل الله أن يجعله في ميزان حسناتك.....اللهم استعملنا ولا تستبدلنا

Thank u sir for this amazing clear tutorial. I have one question. He initialization is applicable for Relu activation.What about other activation functions? Please advise.

In the regularization term, we should not add the bias terms ? sum_of_phi_square should have only weights in the model and not the bias terms ? @1:00:03 And once again, thank you so so much Sir for the lecture, every friday, I just wait for the next lecture, :)