Gradient Descent : Data Science Concepts

[@buihai8371]

man, you video is better than my 3 hour lecture at uni. finally i find a video that go straight to the key point

[@cambridgebreaths3581]

This episode is utterly amazing. Please make a comprehensive complete series of every essential concept in data Science, for absolute beginnes. Truly brilliant explanation. Thank you kindly

[@ritvikmath]

Thanks!

[@barkinkaratay7951]

man this must be the best explanation on gradient descent I found so far. thank you very much

[@amithnambiar9818]

Ritvik this is so well explained ! Thank you so much man.

[@ritvikmath]

of course! thanks for watching!

[@MrMoore0312]

Excellent! Great new twist on standard GD, love the practical objective-oriented decision making process explained with math and code!! Thanks!

[@rahulahuja1412]

Very helpful. Would love to see another video that incorporates advanced concepts like momentum etc. Thanks a lot for your videos. They have helped me enhance my understanding of complex concepts.

[@ritvikmath]

Great suggestion! And thanks!

[@andrewjolly319]

This is really great, and having the notebooks to follow along with is so helpful. I'm at the start of my PhD in physics and we got taught basically zero stats and now suddenly everything is gradient descent and MCMC and despite other textbooks/youtube videos I've not been able to get my head around it until now.

[@DidierDurand]

Great explanations: very easy to understand!

[@JO-vj9kn]

Thanks for a great tutorial! As many others I was stuck with cryptic lecture notes by a genius but not very pedagogic professor and this type of material is way more approachable.

[@richardqualis4780]

Very well explained!!!

[@katieangelopoulos4706]

These 10 minutes were worth more than 3 hours of lectures in a masters.

[@steve8180]

dude thanks for this, it was really really well explained

[@amnont8724]

6:05 the minus is given that we want to find a minimum point right? Because if the gradient is negative, we want to move forward toward the minimum, and if it's positive, we want to move backwards to find the minimum. So if we wanted to find a maximum, we should've turned the minus sign to a plus?

[@otmarjr]

I feel so luck I found your content! You've got a serious talent to explain stuff. In the future could you please shoot a video recomment books and courses for someone interested in becoming a data scientist ?

[@sairaj6875]

Thank you so much. This was a really nice demonstration.

[@aniket1983]

Thank you so much...you are the only one clarified about defining own loss function...I was stuck on if there is only L1 and L2 loss why bother as all gradiant discent will be same...so thanks a ton for the clarity...

[@EW-mb1ih]

of course we want videos on how the step of the descent is obatined and how you treat not convex function :)

[@martand_05]

Nicely explained bro🔥🔥

[@ritvikmath]

Thanks!

[@chacmool2581]

I was able to, more or less, follow this, and I haven't even gotten started with my data science studies!

[@ritvikmath]

Nice!

[@chacmool2581]

@@ritvikmath My engineering background and those punishing four semesters of calculus (integral, differential, diff eqs, multivariate) are still alive in my aging brain. How much advanced calculus do you really need to know, though? I mean, max/min problems and differentiation is the stuff that high school seniors tackle in Advanced Placement courses.

[@LimeObeans]

do one with adaptive momentum and one where you have no function or derivative please!! u rock

[@ragstoriches919]

Amazing video🎉

[@ritvikmath]

Thanks 😁

[@othsaj1467]

Amazing vid 💯

[@chunchen3450]

Any topics on adding weights to time series analysis? For instance higher weights to more recent observations, to get a better forecasting? Is this often used in practice?

[@ritvikmath]

I've definitely seen cases of higher weights for more recent examples

[@danielwiczew]

@ritvikmath Are you planning maybe an introduction to Cholesky, LU and SVD matrix decompositions ?

[@ritvikmath]

Good suggestion!

[@srinivasanbalan2469]

Good content. Thanks

[@ritvikmath]

Thanks!

[@rutu.dances.to.express]

Great intuition to play with loss function in the 2nd part of the video. I had a small doubt here.. The idea of adding e^i to the loss function is to give more weights to smaller values of x right?? What if in a case, we have multiple input features (x1,x2,x3... and so on)- In such a case will it indicate that we care about predicting accurately for smaller values of x1,x2,x3... (irrespective of high or low values in a particular record..? for e.g. one single record can have high value of x1, where as very low value of x2.. and so on. ) Or since you've used the variable i in e^i, does it indicate row numbers 1,2,3,... (initial rows) will get more weight than further rows?

[@ihgnmah]

Thank you for the vid. I just wonder what kind of loss function would take care of the data points in the curve rather than the small cluster below.

[@areskapoor1371]

Interesting video! How did you choose what model to use for the data initially, and also for the second part why did you choose e specifically for e^-i? Is this a common mathematical thing to do?

[@buckylc9345]

Thank you so much for explaining a couple of concepts I didn't find in other more popular videos! Is there a cost function we can use to effectively treat the left lower corner as outliers?

[@houyao2147]

Does this mean we should identify those outliers first and then remove them before training?

[@user-or7ji5hv8y]

but don't you also derive the loss function using MLE? or did I get that part wrong with something else?

[@aRoryBorealis]

Thank you so much

[@ritvikmath]

You're most welcome

[@muhammadal-qurishi7110]

can you please explain CRF

[@ritvikmath]

Great suggestion!

[@adaloveless3904]

I'm dumb to this stuff... you lost me at "derivative of the loss function with respect to W"... all of a sudden the X dimension comes into play... what happens in a neural network where there are many dimensions?

[@nachomacho7027]

perfect