The Math behind Transformers | Srijit Mukherjee | Computer Vision | Natural Language Processing

In this video, I explain the mathematics in the optimum possible way behind the Transformers, which is the widely used architecture behind the natural language processing tasks. Recently, transformers have been used to solve computer vision problems too. The explanation is based on the paper "Attention is all you need" [arxiv.org/abs/...] by Ashish Vaswani et al.
There are some errors in the video which are mentioned in the comments as well as on the following page. Many other learning resources are being shared on the page below.
May you find join in learning about the science of data.

[@arshadkazi4559]

Great Video! Liked the simplicity and honesty in the explanation.

[@mukherjeesrijit]

Thank you Arshad.

[@BiprojitNath]

This is amazing. Keep making more such tutorial.

[@mukherjeesrijit]

Thank you Biprojit. I will.

[@sayantan336]

Just a point the embeddings (512 dim vectorized representations) of the input/output words are also learned as a by-product of the training process.... i.e. The embedding layers both on the encoder side & the decoder side - are part of the learnable parameters of the model

[@mukherjeesrijit]

Thank you for the input.

[@annyd3406]

Please make more videos you are great thank you for this!!!
Make INDIA proud....

[@mukherjeesrijit]

Thank you! I will try my best.

[@michaelestrinone2111]

India Or America? His is a PhD student in American(!) Uni. If he wanted to make India proud he would stay in India.

[@sayantan336]

Well explained ... Thanks Srijit ... Enjoyed quite a lot..

[@mukherjeesrijit]

Thank you.

[@Rahulsircar94]

lol transformers for bengalis.loved it.

[@harishravi9936]

At 55:44,will the mask matrix be a lower triangular with all -inf ?

[@harishnayak976]

What is the equation for whole process in one mathematical form?

[@hari8568]

At 31:51 if we are finding similarities between two words, shouldn't QK^T matrix be symmetric?because relation score between any two words should be the same

[@gasun1274]

exactly

[@hari8568]

At 41:51 how exactly do we get Zi?Are we using PCA?

[@sbera07]

can you provide the notes u have made in this video

[@user-zn8rz3hu2k]

I would like to recommend a fantastic paper by DeepMine that provides comprehensive and detailed explanations regarding this topic.

[@mukherjeesrijit]

please do.

[@subhadipsarkar7692]

❤

[@gasun1274]

you have a tendency to overly raise your tone while you speak, i do this subconsciously too but you need to know that it is very annoying.

[@mukherjeesrijit]

Thank you for your suggestion. I will keep it in my mind.

[@vishruttalekar8626]

Disappointing!!

[@mukherjeesrijit]

Let me know how it can be improved.

[@emrahe468]

​@@mukherjeesrijit I'm a confused by the diagram in your video at @6:00. The left side seems to show inputs in a foreign language being fed into an encoder, while the right side displays multiple sequences in English. Is this setup discussing a specific type of decoder model like GPT-2, or is it more about an encoder-decoder architecture used for translation? The background of the diagram makes it hard to determine the exact context. With that diagram things get messy for me, hence couldn't stand much, sorry