Another perfect lecture, finally we can understand such beautiful subject and not just memorize it like mindless robots. Thank you so much Ritvik, your our hero! Gratitude from Brazil
This a huge gem! I love all your videos, they’re always a beautiful mix of theory, applied, and visual examples. I also think they’re the perfect length as well as depth and breath of connected material covered. That’s a delicate balance most technical RU-vid videos fail at and what makes yours special. 👍
This is the best linear algebra explanation I've ever heard and I've watched basically everything. The only thing you missed was the geometric interpretation, the point of the basis axes don't change. Still, absolutely excellent. 3b1b is the one everyone praises when actually he confuses simple things. You did the reverse.
Thank you so much, you're great at explaining and I appreciate you including the application of the concept in the real world, that helps to connect the points!
PERFECT! As a programmer, I found the process just like "data normalization" which is indeed recommended and useful, amazing. One stupid question, so what's the difference between the column-column check you did, and echelon(row-row) form? I've seen some use echelon
Another great video, thanks RItvik! Could you please make one about the determinant / trace / diagonalization? Because many happen to see these stuff in Linear Algebra courses, I specifically wonder how are they used in Data Science.
Good topic. It turns out that a deep neural network framework is pretty convenient for solving for the two low rank approximation matrices, or finding the exact solution matrices if they exist. I came up with the following technique: In Tensorflow you use two Embeddings layers with your choice of k and one Lambda layer to do a matrix multiply. Your loss function can be a typical choice like L2 distance between the result of the Lambda layer and the entry of the original big matrix. Each entry of teh original big matrix constitutes one training example. The optimizer is your choice like Adam, everyone loves Adam optimizer. So I came up with this arrangement to do movie recommendations on the MovieLens dataset. And it's better than Alternating Least Squares algorithm for many reasons, one big one being with the DNN technique, you will completely avoid making the dumb assumption that there are zero values in the original matrix entries that are missing values. Of course if you are not missing any values then ALS is probably fine.
Off topic, but you should make a video on implementing linear bayes/bayesian logistic regression/similar. Would be on-topic for your channel and would also compliment your non-bayesian implementations.
Hi :) thank you for this video. I wish Ive watched this video before svd video . Would you pls make a video about latent factor Decomposition and CUR model for approximation?
And which math book do you recommend to have an in_depth concept about data science, ml and ai at the same time with practical concept ? Just the way you teach (not pure useless math formula without any data sience related explanation )