inferring evolution, hidden Markov model

Sriram Sankararaman (Harvard) gave a great Computer Science Colloquium today about inferring the evolutionary tree (well, it isn't really a tree) from genetic information, particularly as regards humans and neandertals. He is able to show, using the statistics of DNA variability, that humans and neandertals had intermixing long after they separated (both geographically and as species). He was also able to show that there is statistical evidence for the sterility (infertility) of males after speciation. Awesome stuff, and very related to cosmology in many ways: The models are of two-point statistics of the DNA sequences, not the sequences themselves, and the probabilistic modeling methods (approximate Gaussian likelihood functions and MCMC) are very similar indeed.

Prior to that, in group meeting, McFee and Huppenkothen jointly proposed a plan for clustering black hole timing data using a hidden Markov model: The idea is that the data are generated by a probability distribution that is set by a state, and there are finite probabilities of transitioning from state to state at each time step. This is a well-understood idea in machine learning, but also very close to how we think about the generation of the timing data, fundamentally. Great plan! Huppenkothen's first order of business is to run k-means in a feature space (for initialization of the HMM).

upgrading The Cannon; finding single transits

The day started with a short conversation with Anna Ho (MPIA), Foreman-Mackey, Ness, and Rix about changing the polynomial model inside The Cannon into a Gaussian Process. This move (which Ho and Foreman-Mackey are attempting this week) brings a large number of advantages: It makes model complexity a continuous problem rather than a discrete problem, and it permits us to continuously tune model complexity at every wavelength. We are violating all the usual rules though, because although at training time we are doing standard Gaussian Process regression, at test time we are doing the inverse of Gaussian Process regression! That's a bit crazy.

In late morning, Susan Kassin (STScI) told us about simple measures of galaxy dynamics and the formation and evolution of the Tully–Fisher relation. She busted some galaxy-evolution myths. What she has been doing so successfully at high redshifts (out to one), the SDSS-IV MaNGA project will do at low redshifts (and much more).

I talked to Foreman-Mackey over lunch (and over 4000 miles) about single transits. He has developed (in partnership with Schölkopf) a machine-learning methodology for finding them and understanding both completeness and false-positive rates. We discussed the issues and scope for a first paper, which would create a complete catalog for the main-mission Kepler data.

I also gave comments on current drafts from Malz and Vakili, and had a long conversation with Wandelt (IAP) regarding probabilistic approaches to cosmology. I pitched my project to evaluate a non-approximate likelihood function in real space for CMB maps and he agreed that if we could show that a full likelihood function evaluation is computationally tractable, it might influence the next generation of analyses.

the meanings of sentences

It was a low-research day today, with the urgent blocking out the important! One highlight, however, was a talk in Computer Science by Yoav Artzi (UW) about natural language processing. Unlike other computer-science talks I have seen in this area, Artzi is trying to represent the meaning of sentences and conversations semantically. Usually in a machine-learning method, the "meaning" of the sentence is encoded in some abstract vector space. These meaning objects are generally uninterpretable and have no compositionality or other useful structure. Artzi wants compositionality, because he wants to be able to combine the meanings of multiple conjoined sentences or clarifications in a conversation and also combine with knowledge about the world. So he is learning a semantic model (and an ontology, and a lexicon, and so on) as sentences come in. This is an ambitious project! He is working in the limited domain of interactive systems to schedule meetings and travel, to make his project feasible.

editing and commenting

All I did today was comment on and edit reports by Wang, Malz, Vakili, the AAAC, and the Spitzer Oversight Committee.

optimism

At Friday group meeting, we got status reports from Wang, Vakili, Huppenkothen, and Sanderson. We came up with a well-defined path for Huppenkothen's first paper on classifying black-hole states from GRS 1915.

Over lunch, I had a brief, optimistic discussion with Ness. Yesterday she was concerned that our ability to see the red-clump stars (and separate them from the normal red-giant branch) might have been statistical differences in training set properties, but she reported that all such tests are now checking out okay: We really can see this difference spectrally. There is still a very important theoretical question of why, but that's going to be a good example of the data-driven model informing theory.

spectral age indicators, structured learning

Before group meeting, Ness and I discussed the scope of a paper that separates red-clump stars from ordinary red-giant stars, using the data-driven spectral model we call The Cannon. We discussed also the possibility that this could turn into a set of spectral age indicators: If we can separate the red clump from the red-giant branch, maybe we can split the red-giant branch into the three nearly overlapping branches on which stars rise and fall as they age.

Andreas Mueller (NYU), one of the principal developers of scikit-learn joined my group meeting today. He told us about structured learning, in which you augment learning based on features with other kind of structural information, usually represented as graph edges or even graph edges with features themselves. Key example: If you want to know what pixels in an image are sky pixels, you are interested in their color, but also their proximity to neighboring pixels that are also labeled as sky pixels (or not). He is building, documenting, and maintaining an open-source package called pyStruct.

#astrohackny, candidate Heinrich, LHC

In the morning, I got a great email from Ness, showing that we can separate red-clump and non-red-clump red-giant stars at huge confidence, using APOGEE spectra and The Cannon. I also read and gave comments on Malz's first attempt at doing inference with probabilistic redshifts.

At #astrohackny Price-Whelan and I started on the crazy Gaussian-Process blind source separation plan. We also pitched a linear version too and divided up tasks among the various hackers interested in working on the Planck data. We didn't get very far, because we spent most of our hacking time understanding this Lawrence paper (PDF).

In the afternoon, I had the pleasure of being on a committee for the oral candidacy exam of Lukas Heinrich (NYU). He spoke about RECAST, which is a system to permit outsiders to re-interpret LHC ATLAS searches in terms of new or different physics models. The idea is: If a search has been done and it is relevant to some new or different physics, there is no need to do new searches ab initio until the existing searches have been checked for the relevant physics. This all also ties into ideas of preserving workflow and reproducibility and open science, all of which are very relevant to the Moore-Sloan Data Science Environments.

normal modes in stars

At lunch Huppenkothen gave the brown-bag talk, on neutron star normal modes and their possible use in constraining neutron-star equation of state (and thus nuclear physics). She was pessimistic in the end, because there are so few modes measured, but in a precision sense, the data (taken at face value) do rule out some models.

After the talk, Andrei Gruzinov and I argued about the relationship between Huppenkothen's normal-mode constraints and spin constraints on neutron stars (mentioned also last week by Kaspi in our Physics Colloquium). He made a nice argument, which I will butcher to this: The speed of sound at the surface of a neutron star (or really any gravitationally bound object) must be on the order of the gravitational orbit velocity at the surface. Why? Because otherwise the object would further compress under gravity! This all flows from the point that the sound speed is related to the compressibility through some kind of modulus. Simple! I should check this for the Sun.

I spent part of the morning prior to all this with Foreman-Mackey, discussing plans for his trip to Tübingen and Heidelberg. We want to work further on noise modeling or calibration, in the context of stellar variability, exoplanet search, and asteroseismology. We discussed Schölkopf's causal arguments and why we get overfitting despite them; I don't yet understand what is the appropriate "large data" limit at which the relevant theorems are going to hold.

IBM Watson

I spent my day today at the IBM T. J. Watson Research Center in Yorktown Heights, NY, hosted by Siyuan Lu (IBM). I had great discussions with the Physical Analytics team, and got in some quality time with Bruce Elmegreen (IBM), with whom I overlap on inferences about the initial mass function. I spoke about exoplanet search and population inference in my talk. The highlight of the trip was a visit to the Watson group, where I watched them talk to Watson, but we also looked into the data center, which contains the Watson that won on Jeopardy!. We made some plans to teach Watson some things about the known exoplanets; he is an expert in dealing with structured data.

Vicki Kaspi

Vicki Kaspi (McGill) gave the Physics Colloquium talk today. She compared the fastest-known millisecond pulsar (which her group discovered) to the fastest commercial blenders in spin period. The pulsar wins, but it wins far more in surface speed: The surface of a millisecond pulsar is moving a significant fraction (like 0.1) of the speed of light! She talked about the uses of pulsars for precision measurement and testing of general relativity. It is just incredible that nature delivers us these clocks! I got interested during the talk in the spin constraints on the equation of state: We often see constraints on equation of state from mass measurements, but there must be equally compelling limits from the spin: If you are spinning such that your surface is moving at or even near the sound speed in the material, I think (or I have an intuition) that everything goes to hell fast.

stellar modes, reading minds

At group meeting today Angus talked about her attempts to reproduce the asteroseismological measurements in the literature from Kepler short-cadence data. I think there is something missing, because we don't observe all the modes as clearly as they do. Our real goal is not just to reproduce the results of course; we discussed our advantages over existing methods: We have a more realistic generative model of the data; we can do multiple frequencies simultaneously; we can handle not just non-uniform time sampling but also non-uniform exposure times (which matters for high frequencies), and we can take in extremely non-trivial noise models (including ones that do detrending). I am sure we have a project and a paper, but we don't understand our best scope yet.

Just before lunch, Kyunghyun Cho (Montreal) gave a Computer Science colloquium about deep learning and translation. His system is structured as an encoder, a "meaning representation", and then a decoder. All three components are interesting, but the representation in the middle is a model system for holding semantic or linguistic structure. Very interesting! He has good performance. But the most interesting things in his talk were about other kinds of problems that can be cast as machine translation: Translating images into captions, for example, or translating brain images into sentences that the subject is currently reading! Cho's implication was that mind reading is just around the corner...

Planck and needles in haystacks

On the way to #astrohackny I learned that I can write usefully on the subway on my mobile phone, which possibly justifies its immense cost (in dollars and in valuable personal attention). At the meeting I pitched my proposal for foreground source separation in the Planck imaging. Price-Whelan pointed out that the incredibly flexible model I am proposing could be a model for absolutely anything in any domain. It is also massively degenerate. Undaunted, I proposed that we perform a first experiment using fake data. The idea is to generate data using a physical model and then fit it with this flexible data-driven model, and see what happens. We had various ideas about generating the data. I also spent some time interviewing Colin Hill (Columbia) about the Planck data analysis and data products.

At Columbia Pizza Lunch, there was much arguing about binary black holes, which is one of my favorite subjects. This paper (in Nature, so it is suspect out of the gate) gives indirect evidence of a binary quasar based on a very peculiar light curve. The authors looked at 250,000 quasars to find this beauty, which makes me wonder if the result is a fluke. A lot can happen when you take 250,000 draws from a stochastic process! Worth thinking about and checking.

high-energy astrophysics; inference without likelihoods

[Lull in posting because of vacation in Quebec. Slept one (exceedingly cold) night in an actual, real-life igloo!]

At lunchtime in the NYU Center for Data Science there was a great talk by Daniela Huppenkothen about x-ray and gamma-ray astrophysics, for non-astronomers. She talked about imaging, spectroscopy, and time series, with a focus on the latter. She did a great job explaining the differences between astrophysics and other data-science domains. At the end there were good questions from (among others) neural scientist Bijan Peseran, who (comparing perhaps to his own domain) was interested in non-trivial time correlations among photon events. After all, neurons are all about non-trivial time correlations in spike trains.

Earlier in the day, Foreman-Mackey and I spoke about K2 projects and exoplanet population projects. The plan is to try some likelihood-free inference; we spent some time talking about technical details. In likelihood-free inference (ABC) one performs repeated simulations of the data; there are fundamental parameters, and then there are (usually) also random-number draws. We might want to sample separately in these, in some Gibbs-like way. Thanks to Brewer for getting us thinking along these lines.

non-Gaussianity in the CMB

Kendrick Smith (Perimeter) gave the astrophysics seminar today. He talked about non-Gaussianity. Various impressive things in his talk. One was the fact that they can measure, at high signal-to-noise, even four-point functions in the Planck data, and confidently see expected deviations from the Gaussian. This may be the only data set in the physical sciences in which a four-point function can be measured at very high precision. Another impressive thing is that, in the context of various inflationary scenarios (in which small non-Gaussian effects arise), he can evaluate the probability of a non-Gaussian initial condition, and therefore use MCMC to generate a realization. He is thinking in terms of initializing n-body simulations, but this is also critical for CMB data analysis: It means, potentially, that he could do full, non-approximate inference of inflationary parameters given the CMB data. It was a great talk, filled with good ideas about using very good mathematical physics to tractably connect theory to observations.

At group meeting, Sanderson asked us about potential statistics we could be using to see the effect of dark-matter substructure on old accreted stellar debris in the Milky Way. She is thinking about hot, old structures, not cold, young streams. In the afternoon, Malz and I talked about redshift likelihoods and photometric redshifts and the overlaps between what we have been thinking about and HETDEX with the thought that he might be able to become a collaborator on that project, given his past work there.

undisclosed location

I spent the day at an undisclosed (wintry) location, sitting by the fireplace, watching the snow falling, and putting words into Foreman-Mackey and Montet's paper about new planets found in the K2 data.