CSCI 5722: Computer Vision | Spring 25 | Special Lecture - DeepSeek

[@kaidongliang3483]

非常棒

[@hasainahmed1883]

I don't know how this 1 hr passed in reel era and now I am looking for more such great content.

[@Coconut-Crusted-French-Toast]

Thank you for the detailed walkthrough. Very helpful.

[@GGWPTrader]

Mad respect sir.. i'm gonna having fun & learn so much from your channel..

[@windmaple]

Thank you for making this video available!

[@Artem-c1p9q]

Excellent work! Love it! ❤❤❤

[@ai-by-hand]

Thank you! 😊

[@maycodes]

Thank you professor ❤❤❤

[@dongwoo113]

감사합니다!

[@camille-t7z]

amazing content, thanks a lot!!!

[@mueezurrehman8572]

Really informative and detailed. Thanks.

[@ihcnehc]

Thank YOU!

[@pastrop2003]

vere well done, thank you!

[@mohsinkhalid2375]

Professor what about the fine-tuning part? How RL is utilized to fine-tune the model.

[@anilshinde8025]

Great lecture Professor. Would like to know role of Group Relative Policy optimization role in DeepSeek

[@airesearch2024]

Could you create a play list for this course so we can keep track of

[@fintech1378]

how to order the book

[@deeal]

why do we do this dimension reduction? from 5=> 4?

[@zeeshanashraf4502]

@deeal tldr explanation - reducing dimensions increases the speed of training and inference. It also reduces the size of the model. 5 is the model size/hidden dimension(H.D), 4 is the per head dimension(P.H.D) and 6 is the Context Length(C.L). If the weight matrix does not reduce the dimensionality , it's size will be (H.D x H.D) instead of current dimension (P.H.D x P.H.D). This will result in larger number of parameters being learnt. Also, if dimensionality is not reduced, K, V, Q matrices will also have size (H.D X C.L) which is much larger than (PHD X CL). Multiplying such large matrices is very expensive, approximately O(n^3). So smaller matrices are used.

[@deeal]

Thanks for the explanation, much appreciated. Interesting, so why not start with that input size to begin with? I understand the input are the embedding right? I am missing something 😅