How to Master Deep Learning Math?

[@deeplearningexplained]

30 days free trial of Brilliant + 20% off premium subscription over here: brilliant.org/deeplearningwithyacine/ also, do check out these three very cool visual mathy channels: Two cool visual math channel: @3blue1brown‬ @algorithmicsimplicity‬ ‪@CodeEmporium‬

[@fuhodev9548]

I think the hardest point is not to understand the math, it is about how can you derive new stuff with what you learnt.

[@deeplearningexplained]

that's definitely the hardest form of understanding yes.

[@erikanderson1402]

That’s pretty much what PhD’s in math do.

[@vickyt.819]

As a undergraduate math student, this video is a hidden gem.

[@deeplearningexplained]

Best of luck with your finals! 🔥🔥🔥

[@christiangreisinger2339]

Can you also do a video for learning how to write good code (OOP in python, pytorch, jax, ...) in order to contribute to complex projects

[@deeplearningexplained]

Great recommendation Christian, will add it to my list!

[@EmmanDesh]

Yessss

[@examgpt]

Brilliant boy!! Very crisp and clear!! - From a guy from remote part of India

[@deeplearningexplained]

Thanks! I’m glad it was useful! 👍👍👍

[@profenevarez]

I agree with everything you shared in this video and I really like the spatial connection mathematics has! That’s really neat to know and see why practicing mathematics IS inextricably connected with LEARNING mathematics!

[@deeplearningexplained]

💯

[@RomPvP]

do you have a blank version of the spreadsheet? it looks great for visualizing the topics you have to learn and where you failed to understand things previously

[@deeplearningexplained]

Hello there, Yes I do it's over here: docs.google.com/spreadsheets/d/1amCdV-uUWjGoIuJKq1I1zCaFzX--uuJaOX9fUELKj9E/edit?usp=sharing

[@Chadpritai]

Kindly create a refresher videos on maths for deep learning and then make videos on paper implementation with code. Thanks Jon Snow, merry Christmas and happy new year 2025🎄🎁⛄

[@deeplearningexplained]

Hey thanks for the recommendation, will do 🫡! Merry Christmas and happy new year to you too!

[@karavan_narzan]

In your experience, when solving problems to understand the “motion” and the “shape of the problem”, is it required to solve the problems end-to-end, or just enough so you grok the intuitions?

[@deeplearningexplained]

Good question, it really depends on your prior knowledge and how similar it is to the problem at hand. If it's a field of mathematics you never saw before, you need to do the motion end-to-end multiple time. If it's a connex field, you might not need to drill as much end-to-end. The lated is a bit like trying to swim in a lake versus a pool. There's a lot of stuff that transfer, but you need to still figure out how to breath properly through the waves and follow more open-ended circuit.

[@dalton4035]

Could you give an example of how you would memorize a problem/topic in your mind palace?

[@deeplearningexplained]

Very solid question. It depends really on the purpose of the memorization and the timeframe I'm looking at. One thing I did when I was studying biochemistry is I made heavy use of visualization encoding to remember abstract element that had a logic like the Krebs Cycle: microbenotes.com/krebs-cycle/ What I did was I setup spatially every 9 steps into my house I lived in with 9 loci. Then at each loci I've put a visualization of the end production of an element. For instance, Acetyl CoA was encoded at the start of my staircase as a very old projector that use acetate (had them in my primary school). Then when I climbed the steps I was splashing around lemon juice with a very strong smell. When I got to my door there was a huge lemon pearling with water I had to displace by hand, it was slippery, but had a nice smell to it. The encoding above means for me: 1. Acetyl CoA 2. Through Citrate synthase 3. Generate Citrate Then I went on through each of the 9 loci. If you setup your visualization properly you will memorize that full chart in like 2-3h max and it's very sticky. After a while, I didn't need anymore direct access to the memory palace to remember the steps and I just remembered the Krebs Cycle as something natural.

[@Chadpritai]

Would you mind sharing the spreadsheet link 🙏🏻?

[@deeplearningexplained]

Yes for sure it’s here: docs.google.com/spreadsheets/d/1amCdV-uUWjGoIuJKq1I1zCaFzX--uuJaOX9fUELKj9E/htmlview

[@BladeAurora]

i like the spreadsheet. why do you select 5 tries instead of 10? what's the reason for choosing that number of tries?

[@deeplearningexplained]

Great question! Remember after each try you need to study the theory and make sure you understand what is going. From experience, going through the motion 5 times is more than enough to clear out any yellow and red. Take into consideration you are doing ALL the problem here. Usually there is a lot of overlap between the problem sets, so as you are clearing a yellow you are also strengthening yourself for the upcoming problem. If I have a red for 5 consecutive tries, it usually means that there is something wrong with the problem (not my solution).

[@Chadpritai]

What do you think about this approach " you don't to learn all maths first hand you can learn it when it's required as you go deep down" ?

[@deeplearningexplained]

Great question, I’m a firm believer of the top down approach. It’s just the most efficient way to learn. However, once you are already doing that and you flagged that a lack of formal calculus is hindering your progress I would do the green, yellow and red method specifically on that topic. Doing it this way is more efficient and more motivating because you know exactly why you’re learning calculus.

[@wombodombo9005]

Thank you for the green yellow and red method thats a very good method!!! However, I have a question, let's take linear algebra as an example, I read the first 200 pages of Gilbert Strang's book and I took about 6 to 12 hours per sub-chapter, I did all the exercises thanks to the correction and the help of GPT to help my intuition and I've improved a lot since then and i can do 90% exercices of the problems sets. However, I think I've done at least 500 exercises, should I go through them again to see my weak points or apply this method from now on? Thank you

[@deeplearningexplained]

Awesome work! 500 exercises is already better than most ever do for a subject. I wouldn’t necessarily do them all, however I would take 1-2h to find out the one you struggled with and add them into your list of exercises. Apply the method to them and future exercise from now on, you’ll see you will become much more efficient at finding your weak points and focusing on improving those!

[@wombodombo9005]

@@deeplearningexplained Merci beaucoup pour ta réponse, ta chaine est une mine d'or 🙏

[@IgorKovacs02]

please keep it up! :D

[@deeplearningexplained]

Thanks! I will! :)

[@chuchyi]

also mathacademy worth mentioning, i wish i knew bout it before i started

[@deeplearningexplained]

Thanks for mentioning, didn’t know about them!

[@chuchyi]

recommend to look into, they’ve optimized process of learning to new level

[@Chadpritai]

They're hella expensive for me 😢