Uncategorized

Machine Learning

I have become interested in the intersection of machine learning with science.  Two of my students will explore this intersection over the summer.    One of our project involves classification, the other involves reinforcement learning.   I’m excited!

Posted by william, 0 comments

Plot.ly

I went to a pretty exciting meet up last night at George Washington University.  It was by put on my Statistical Programming DC and was my first meet up with them.  The speaker was Matt Sundquist, who is one of the founders of plot.ly, which is sort of the github of plotting.   The talk was pretty good and I was impressed by the package.  It’s free and they have bindings to Julia, Python, Excel, Matlab, Igor, R, etc.   You can view graphs, edit legends, etc. from within the browser and send them directly from your python scripts, which is really cool!   You can also stream real time data over.  One really nice feature is that you can see the data used to generate a plot and the code (in multiple languages) that will generate it.  This would be really cool to see in journals!   It ties in nicely with the recent posting by Nature about hosted iPython consoles.


Given the variety of graphs, I can understand why they went with d3.js—we went with jqplot for a number of our online plotting solutions because I thought the learning curve would be shallower for our high school/college interns.   Also, I think we started back in the days of Flot, before d3 was out there.    I somehow thought that d3.js was only SVG, but one of my coworkers told me that it now supports canvas as well, so it could be fast—but apparently producing pdf files from canvas is painful...

I hope their company survives and I can definitely see using it for sharing plots with my collaborators.   Because of our use of interactors, I doubt that I will be able to replace a number of existing projects though due to our need for custom interactors.   Hopefully they develop a plugin system in the future!



Posted by william, 0 comments

Instabilities

I went for a jog today and thought back to an insight that I had in graduate school.   During my phD, I studied frustrated magnetism.   In frustrated magnets, you consider a material where the atoms are arranged in triangular motifs.  If these atoms have unpaired spins on them and only nearest neighbor antiferromagnetic interactions, then the interactions will be unable to be simultaneously satisfied and the material will never order, despite having strong interactions.   However, in practice, I found that Nature is tricky and that a number of these materials would order.  In some cases, thermal or quantum fluctuations would result in ordering.  In other cases, order in the system could be stabilized by a structural phase transition which would favor some interactions over others.   However, one thing that gradually became clear was that there were interesting phases near such frustrated magnets and the interesting question is how a small perturbation might result in interesting physics.

 

I think this could be a general strategy for searching for switchable phenomena.  Find a phase which is subject to an instability through some type of perturbation.  For example, pressure, electric field, etc.   If you can move the system towards the instability using one control parameter you’re half way there.  The next challenge is whether you can use another perturbation to move it back across the boundary using another control parameter.   If so, then you have a chance to build something interesting...

Posted by william, 0 comments

Uncontrolled!

I just finished reading Uncontrolled by Jim Manzi.  He starts the book spending a surprising amount of time in the early chapters of the book trying to give the reader a crash course in the philosophy of science.   The TL; DR is that I’m rather lucky to be a physicist.   The longer answer is that in the hard sciences such as physics and chemistry, we often have basic models of how the world work which we constantly test against new observations.   If we’re lucky, we find that our models don’t describe something in nature and that there is new science to discover.   But, at a deep level, we have a few articles of faith.   For example, we believe that the laws of physics don’t change from place to place.  We also believe that the laws of nature don’t change with time (even if our understanding of them does).   Other fields such as medicine are not so lucky.   We still have a very rudimentary understanding of the human body.  We don’t have strong enough models to say whether a given compound should cure a given disease.   There are also many confounding factors.   So, randomized trials are necessary to see whether the effects observed are intrinsic or accidental.   Even if a randomized trial indicates that an effect is present, it can be difficult to generalize it to a different context or population.   Sociology and economics are even more difficult.   The author brings up a number of cases in economics where authors have overreached and made predictions that really should have been tested in randomized experiments—otherwise, they are far too overreaching with too little support from data.

After this introduction, Manzi delves into perhaps the strongest section of the book where he outlines how he used randomized experiments in business to determine strategy.   When doing this, he looked into the literature of random trials in medicine, as well as random trials in social sciences (for example, leading up to welfare reform in the 90s, the federal government required a number of random trials for states which wanted to try different strategies.).  

There is another section of political trials, which reminded me of Rick Perry and His Eggheads and the better Get Out the Vote, which deal with experimental tests in politics of what strategies work and which don’t in get out the vote efforts.   

 

Perhaps the weakest section of the book is where the author attempts to make policy suggestions.   But overall, I would recommend reading this book.  If nothing else, it is helpful in increasing skepticism of predictions which lack experimental support.

Posted by william, 0 comments