Math Behind LLMs, Transformers and Modern Computer Vision (Udemy.com)

I love how simply Patrik teaches about Image Segmentation with Vision Transformers, very clear and easy to follow as always. update: just finished the course. Amazing explanation of partial derivatives, gradients and backprop

A thorough and well-structured course that does an excellent job of demystifying the attention mechanism. What I appreciated most was the step-by-step approach - nothing felt rushed or skipped over. Whether you're a complete beginner or someone looking to solidify your foundations, this course delivers.

I think the first 5 sections were great, clearly explained, but the last section was too messy. I feel it was ok to give the interpretation after each main step using the computed matrices, but the rest of the explanation was confusing. Maybe the instructor can rely on a diagram to explain the conforming blocks of the architecture and to connect each one of them. More diagrams and images would be useful during the explanation.

Clear explanations and strong intuition-building around the math behind Transformers, LLMs, and computer vision. The instructor breaks down complex topics in a digestible way, and I feel more confident reading ML papers now. Recommended.

I think LLM side is perfect! Last chapter is more complex and so I cant focus. I have a lot of questions inside my brain about LLM. I should solve them. I want you to explain more detail for LLM as a GPT or basic features etc. You are a really good teacher. I enjoy absolutely and I got it. If you keep going tell us step by step, everything will be impeccable.

Brilliant wrap of 4.5 hours around Attention and Transformers in LLM. Course is worthy of its price and the duration. But there's much more that I wanted to learn, like neural networks and the maths behind them as well which were not covered here. But still a worthy course to start your mathematical journey around AI. I'd recommed to start from this one and then continue learning through other sources and courses as well.

In depth coverage of mathematics. Till now I have only covered the math for positional encoding, based on the experience I am sure the attention math will also be covered in depth. Please make more courses on math of deep learning/ML in general too, I will subscribe to those.

This was one of the first course in Udemy where I completed the course. In the past, I have taken courses and with some interruption, I lost the motivation to resume. The best thing about this instructor is that he hammers the main points repeatedly so even if your interest is flagging due to not remembering some important details after a hiatus, the instructor is able to nudge you back to resume your learning.

I appreciate this course and learning a lot from it. But there are a few areas where I wished to receive more empathy from the instructor. For example in section 4, chapter 16 where you define key matrix and take an example of 3 by 3 (I think in chapter 17), I did not understand what value add does key matrix (or key vector) bring that embeddings vector does not. I checked online and found that key matrix often "tones-down" the complexity of the embeddings to lower dimension (often to suit the task at hand). Otherwise the embeddings vector is an large 768 dimension vectors geared for all scenarios and key/query and value vectors are lower dimension. The example you should have taken is not 3 by 3, but probably 4 by 3 or 3 by 2 to illustrate this idea.

It's a very detailed (maybe too much at some point) explanation of math behind the transformer learning mechanisms. The course has a lot of repeats from one lecture to the other though, so I fast forwarded most of it and played only the parts when Patrick refers to his underlying document, sometimes not even playing but only reading it. What I liked the most is Patrick's ability to make complex things easy to understand thanks to relevant images and metaphors he uses to explain. It's in the Q&A part that things became crystal clear.