Markov Chains + Monte Carlo = Really Awesome Sampling Method. Markov Chains Video : • Markov Chains : Data S... Monte Carlo Video : • Monte Carlo Methods : ... Markov Chain Stationary Distribution Video : • Markov Chain Stationar...
My dude, I don't often need your teachings, but when I do you are able to single-handedly overshadow most of my past professors. I've watched in the past 4 years a good chunk of your videos and you didn't do a single one in which I didn't add some new view, even if small, on the topic. Keep it up with the work.
Really excellent series of videos - been scratching my head over sampling methods for ages, but you explain it so succinctly and clearly it is finally making sense. Thanks for these!
Thank you so much, I'm a scientist myself and have used some mcmc package blindly. Now, applying what I have been doing to every step of this video made me understand the full concept super clearly.
I gotta say your videos have been super helpful for a stats subject I took last semester (which involved time series, ARIMA model, stationarity etc.) and now MCMC came out at the perfect timing. You have such a gift for explaining the intuition behind statistical concepts, and I'm looking forward to future videos from you. Your channel is a treasure!
Thanks for making this video. Finally came across the one that explain MCMC in plain words without dumping math formulas. Hope other videos and articles in follow this.
Hi Ritvik, your explanations are great in many ways. One of the best things is they are very logically coherent, leaving no gaps that require the listener to figure out. Please do keep up the splendid work. This is a major good deed for so many.
I'm speechless; your presenting style and explanatory power is insane!!! Thank you so much, I'm just getting into this stuff and the reading is tricky Liked, subbed, etc. 👍👌😁
I expect by watching this video, the percent successful uptake of this material for me is so much better than any textbook alone. YT and presenters like ritvikmath is the way to learn new STEM stuff for sure. Much faster and easier, this way. It's like when they finally translated the Bible from Latin to English, and now I'm not needing to suffer with the Latin version any more. haha
a complement about why detailed balanced condition is valid if a distribution is stationary, it's because of bayesian statistics. recall the equation P(a|b) = P(b|a)p(a)/p(b), some rearrangement we get: p(b)P(a|b) = p(a)P(b|a) if it's in stationary, p(a) and p(b) are const, then the equation holds, we call it detailed balanced conditon.
Thank you so much for this video. This is really helpful for my undergraduate research work. One thing I'm finding difficult to understand is, why do we use "thinning" in MCMC ? From what I have read so far, it aims to reduce autocorrelation - but why? Please tell me your thoughts on this problem. I appreciate it a lot. TIA
fantastic. are you just going through chris bishops book and making videos to help us out? i'm reading it atm and keep finding content on your channel. it really is quite helpful in providing intuition for a very dense subject
I wish Ian Goodfellow's book explained MCMC like you do. And I wish my professors back in university can teach and give intuition like this video. I would have been much more interested in stats and data science if it was taught properly.
Great video! Really liked the high-level explanation to get us comfortable with the ideas behind these methods. Quick question: I'm assuming we don't know p(x), so how do we construct a stationary distribution about p(x)?
Really great video. A quick question though, what if I want to approximate f(x)? Currently I am using a form of MCMC to do this to estimate the state probability of n samples.
Thank you, you are always the best. I am working on Bayesian network structure learning using Gibbs sampling, Could you suggest the best book or video which will help me to go through this please. Thank you.
@ritvikmath by any chance would you happen to have some notes presenting the topic in more depth? I have a general idea of the method but having trouble wrapping my head around some methods presented in papers. If not, its okay!
So the Monte Carlo part refers to the eventual sampling from the stationary Markov Chain? I kind of missed where it comes in, except for the board title.
The Monte Carlo part refers to simulating steps through the Markov Chain. So we design a Markov Chain with some transition probabilities and then we start at some x0 and step from one state to the next which is the Monte Carlo part.
Also, could you maybe make a video on where in Data Science sampling techniques like MCMC (Gibbs, Metropolis ...) are useful? Missing data imputation? Would be highly appreciated!
I get a philosophy from here. The objective is actually is to design the appropriate transition probability. It's like to build work out and healthy eating habit if you want a body goals.
Loved your explanation but can you please organise the videos I need to see serially before watching the "Markov Chain Monte Carlo (MCMC) : Data Science Concepts" video. All the videos are scattered all over the place.
How exactly should the end of the burn in be detected and decided by an iterative algorithm, when it's a random variable that is being monitored, and it is therefore jumping around (so you can't see if it goes flat compared to prior values) and you don't even have the truth value to compare with, because otherwise you'd already have your goal in hand at the very beginning?
How do we know the p(x) that should be the steady state of our MC? because I think the p(x) is the black box that we do not know and wants to sample from it to find it. If we have p(x), what is the obstacle against us that prevent us from sampling from it? This is a little bit confusing for me in all sampling videos on RU-vid.
To be clear the following comment is in no way a criticism; rather it's a line of thinking as to illuminating how I can use this tool on some project. Can you also demonstrate a powerful application or two of this powerful method, on real data from a business, institution, or science dataset? So then is this machinery intended for making better simulations? Such as...? Compared against baseline case that does not use it, how much better is the answer to the problem? Accordingly, in this vein, some excellent looking software frameworks to help use MCMC were recently very well described by Kapil Sachdeva also on RU-vid, particularly PYMC3, Stan, NumPyro, and TFProb. (Sorry for YT, but I expect YT will interfere with it if I provide a URL in this comment linking directly to Kapil Sachdeva.)
