Hacker News Comments on
École normale supérieure
Statistical Mechanics: Algorithms and Computations
Hacker News Stories and CommentsAll the comments and stories posted to Hacker News that reference this url.
My personal favorite, Statistical Mechanics and Computations.. Excelled introduction to MCMC methods, lots and lots of python programs and filmed against a green screen.
Obligatory, Stanford CS231n: Convolutional Neural Networks for Visual Recognition  The assignments are excellent and will let you implement a deephish network from practically scratch, before diving into modern frameworks and applications.
My personal favorite introduction to MC(MC) is lecture 1 of statistical mechanics and computations 
You can not sample the distribution directly when “it is impossible to solve for analytically”. There are a few methods to overcome this issue, in particular “MCMC methods allow us to estimate the shape of a posterior distribution in case we can’t compute it directly”.
You only need to be able to compute the relative values of the function you want to integrate (in this case a probability density) at the current point and the proposed destination. Where is this function coming from is not relevant to understand MCMC methods, they can be applied in many problems unrelated to Bayesian statistics.
Here is another (non-zero-math) introduction to the topic: https://arxiv.org/pdf/cond-mat/9612186.pdf https://www.coursera.org/learn/statistical-mechanics/lecture...
BTW there is a nice related course on coursera: "Statistical Mechanics: Algorithms and Computations". Also notice there is a rosettacode entry for this.
The intrinsic important of MCMC is clear. What kinds of applications would be of excitement to startup-oriented readers here on YCombinator?
Also why is this called Quantum MCMC rather than just normal Markov Chain Monte Carlo?
By coincidence there is a Coursera Cousrse that just started on Statistical Mechanics and Algorithms, and the first exercise is to approximate pi.
⬐ msane> Also why is this called Quantum MCMC rather than just normal Markov Chain Monte Carlo?
I had the same question, so I looked it up. From https://en.wikipedia.org/wiki/Quantum_Monte_Carlo :
"Quantum Monte Carlo encompasses a large family of computational methods whose common aim is the study of complex quantum systems. One of the major goals of these approaches is to provide a reliable solution (or an accurate approximation) of the quantum many-body problem."⬐ ylemTaking a quick look at the blog, I believe that the author is simply trying to lead up to quantum monte carlo. I'm not a theorist, but the basic idea is that we have a pretty good understanding of quantum mechanics. But in materials, we don't just have one or two electrons roaming around, but rather have HUGE numbers. So, we can try to make a number of approximations to get tractable calculations, BUT a number of those approximations actually neglect the strong interactions between individual electrons. Quantum monte carlo tries to apply monte carlo techniques to this problem. You might enjoy the summary at: http://people.physics.illinois.edu/Ceperley/papers/213.pdf
(though it's rather geared towards physicists).
As for the hacker news appeal...hmmm....There are two levels. One is the curiosity appeal of how things work. The applied level that I can think of would be quite down the line. I went to a talk at a conference awhile back (again, I'm not an expert on this) that talked a little about this. Searching for new materials is hard. If you remember some years back, there was a scare about China restricted exports of rare earth materials (which make up high power permanent magnets) and people began to search for replacements. Well, how do you search for materials with new properties? A lot of what we've done in the past has involved heuristics. If you are doing synthesis, you make some guesses, over time, you build up intuition, and then you try to start making materials--but you don't know if they will form, or if they'll even have the properties you want if you make them.
One approach that people have taken is to use a method called Density Functional theory (DFT)--or ab initio methods (parameter free)-- that only works at 0 temperature and neglects many of those interactions between electrons that I told you about. And despite this, it can get you some really nice directions to search in for some materials. Now, to better capture some of those interactions that you left out (which you need to understand to get things like magnetism correct) you can try to add back in by including a parameter called "U" that is a measure of how much energy it takes for an electron to move from one atom to another in your material. Within DFT, this is a "fudge" factor and people look at it over some range of values to see how it effects material properties (keep in mind that these calculations are computationally expensive!!!).
Now, it turns out that that recently some researchers from Rutgers have figured out how to calculate this U parameter using something called "Dynamical Mean Field Theory". And, it turns out that there is a stage in this calculation that can be mapped to a problem that Quantum monte carlo simulations can solve quickly.
So....Long story short (ok, maybe too late for that), it can help us when we try to calculate the properties of materials--without synthesizing them--which saves a lot of time and money. But, it's still early days...
I hope this helps a bit...⬐ mauritsI highly recommend the coursera stat mech course. Professionally filmed against a green screen, clear explanations and a massive amount of python programs to illustrate it all.⬐ agalea91The python stuff was just vanilla MCMC sampling with the Metropolis algorithm.
Quantum is popping up because this is in introductory series aimed at people who want to understand QMC.⬐ evanbIn Lattice QCD  we basically execute huge MCMC in order to sample the partition function  on a discretized spacetime.
Convolutional Networks from Stanford . No video, but comes with a wonderful set of ipython notebooks to illustrate and work with cnn's.
Statistical Mechanics Algorithms and Computations . Very well done video's shot in a studio with a green screen. Comes with massive amounts of small python programs to illustrate the material.