These are my recent Pin​board​.in links:

• [1006.5366] “Not only defended but also applied”: The per­ceived absur­dity of Bayesian inference

  “The mis­sion­ary zeal of many Bayesians of old has been matched, in the other direc­tion, by a view among some the­o­reti­cians that Bayesian meth­ods are absurd-​​not merely mis­guided but obvi­ously wrong in prin­ci­ple. We con­sider sev­eral exam­ples, begin­ning with Feller’s clas­sic text on prob­a­bil­ity the­ory and con­tin­u­ing with more recent cases such as the per­ceived Bayesian nature of the so-​​called dooms­day argu­ment. We ana­lyze in this note the intel­lec­tual back­ground behind var­i­ous mis­con­cep­tions about Bayesian sta­tis­tics, with­out aim­ing at a com­plete his­tor­i­cal cov­er­age of the rea­sons for this dismissal.”
  
  social-​​dynamics sta­tis­tics martial-​​arts-​​schools

• [1206.3268] Fea­ture Selec­tion via Block-​​Regularized Regression

  “In this paper, we con­sid­ered the prob­lem of find­ing a sub­set of covari­ates in a high-​​dimensional space that affect the out­put vari­able when there is a block struc– ture in the covari­ates. In the con­text of asso­ci­a­tion map­ping, we pro­posed a regression-​​based model with a Markov chain prior that encodes the infor­ma­tion in the cor­re­la­tion struc­ture such as dis­tance and re– com­bi­na­tion rate between adja­cent SNP mark­ers. We demon­strated on the sim­u­lated and mouse data that our pro­posed algo­rithm can be used to iden­tify groups of SNP mark­ers as a rel­e­vant block of causal SNPs. The idea of rep­re­sent­ing the cor­re­la­tion struc­ture as a Markov chain in a vari­able selec­tion method to learn grouped rel­e­vant vari­ables can be gen­er­al­ized to use a graph­i­cal model as a prior in a vari­able selec­tion prob– lem to rep­re­sent an arbi­trary cor­re­la­tion struc­ture in vari­ables in a high-​​dimensional space. Another inter– est­ing exten­sion of the model is to model a struc­ture in out­put vari­ables as well when mea­sure­ments of mul– tiple out­put vari­ables are available.”
  
  sta­tis­tics bioin­for­mat­ics algo­rithms data-​​mining feature-​​extraction

• Fil­ipe Kiss : A bet­ter git log

  “So, are you tired of this old and bored git log screen?”
  
  yes software-​​development git tricks-​​n-​​tips bash

• Neu­roskep­tic: Brains are Dif­fer­ent on Macs

  “The paper goes into lots more detail, but the les­son for researchers is extremely sim­ple: don’t cross the streams of data-​​analysis. Set up your analy­sis stream and then use it on all of your data. Same hard­ware, same soft­ware, same set­tings. Imag­ine you’re doing a study com­par­ing brain struc­ture in two groups. Halfway through ana­lyz­ing your data, you upgrade your MacOS. All of the brains you ana­lyze after that will be, say, 5% “big­ger”. That’ll cer­tainly make your data much nois­ier, and if you hap­pen to ana­lyze most of Group A before Group B, it’ll give you a false pos­i­tive find­ing. Some­times you just can’t avoid changes in hard­ware or soft­ware — IT techs have a habit of upgrad­ing things with­out ask­ing — but in these cases, you should run the same data under the old and the new regime to see if it’s mak­ing a dif­fer­ence. Finally, it would be wrong to blame FreeSurfer for this. I’d be sur­prised if they were any worse than the other soft­ware pack­ages. Mix­ing and match­ing ver­sions is some­thing that the FreeSurfer devel­op­ers specif­i­cally warn against. This paper shows why.”
  
  data-​​analysis repro­ducibil­ity technical-​​assumptions anomalies-​​are-​​where-​​you-​​find-​​them

• Plug: What is infer­en­tial­ism? « Odontomachus’s Blog

  “I’ve been crit­i­cal of objects and the idea of ref­er­ence for a while now. To me sen­tences and propo­si­tions, by virtue of their role as “moves” in social inter­ac­tions, are likely to have pri­or­ity in a prop­erly objec­tive account of mean­ing. Many puta­tive objects (e.g. cor­po­ra­tions or muta­ble dig­i­tal doc­u­ments) bor­der on being fic­tional, gain­ing their object­hood only through what we say about them; and many refer­ring phrases seem to refer to dif­fer­ent things, depend­ing on what is being pred­i­cated. I think this opin­ion would make me what Pere­grin calls a “strong infer­en­tial­ist”. Even­tu­ally I hope that think­ing clearly about seman­tics ought to (among other things) help bring calm to the cur­rent mass hys­te­ria which is the Seman­tic Web and Linked Data, and help steer all of that energy expen­di­ture to improve its consequence.”
  
  prag­ma­tism indirect-​​links phi­los­o­phy talking-​​about-​​thinking-​​and-​​the-​​reverse

• [1206.3552] A Clas­si­fi­ca­tion for Com­mu­nity Dis­cov­ery Meth­ods in Com­plex Networks

  “In the last few years many real-​​world net­works have been found to show a so-​​called com­mu­nity struc­ture orga­ni­za­tion. Much effort has been devoted in the lit­er­a­ture to develop meth­ods and algo­rithms that can effi­ciently high­light this hid­den struc­ture of the net­work, tra­di­tion­ally by par­ti­tion­ing the graph. Since net­work rep­re­sen­ta­tion can be very com­plex and can con­tain dif­fer­ent vari­ants in the tra­di­tional graph model, each algo­rithm in the lit­er­a­ture focuses on some of these prop­er­ties and estab­lishes, explic­itly or implic­itly, its own def­i­n­i­tion of com­mu­nity. Accord­ing to this def­i­n­i­tion it then extracts the com­mu­ni­ties that are able to reflect only some of the fea­tures of real com­mu­ni­ties. The aim of this sur­vey is to pro­vide a man­ual for the com­mu­nity dis­cov­ery prob­lem. Given a meta def­i­n­i­tion of what a com­mu­nity in a social net­work is, our aim is to orga­nize the main cat­e­gories of com­mu­nity dis­cov­ery based on their own def­i­n­i­tion of com­mu­nity. Given a desired def­i­n­i­tion of com­mu­nity and the fea­tures of a prob­lem (size of net­work, direc­tion of edges, mul­ti­di­men­sion­al­ity, and so on) this review paper is designed to pro­vide a set of approaches that researchers could focus on.”
  
  via:cshalizi graph-​​theory com­mu­nity clas­si­fi­ca­tion algo­rithms nudge

• [1205.0792] Exact Wavelets on the Ball

  “We develop an exact wavelet trans­form on the three-​​dimensional ball (i.e. on the solid sphere), which we name the fla­glet trans­form. For this pur­pose we first con­struct an exact har­monic trans­form on the radial line using damped Laguerre poly­no­mi­als and develop a cor­re­spond­ing quad­ra­ture rule. Com­bined with the spher­i­cal har­monic trans­form, this approach leads to a sam­pling the­o­rem on the ball and a novel three-​​dimensional decom­po­si­tion which we call the Fourier-​​Laguerre trans­form. We relate this new trans­form to the well-​​known Fourier-​​Bessel decom­po­si­tion and show that band-​​limitness in the Fourier-​​Laguerre basis is a suf­fi­cient con­di­tion to com­pute the Fourier-​​Bessel decom­po­si­tion exactly. We then con­struct the fla­glet trans­form on the ball through a har­monic tiling, which is exact thanks to the exact­ness of the Fourier-​​Laguerre trans­form (from which the name fla­glets is coined). The cor­re­spond­ing wavelet ker­nels have com­pact local­i­sa­tion prop­er­ties in real and har­monic space and their angu­lar aper­ture is invari­ant under radial trans­la­tion. We intro­duce a mul­tires­o­lu­tion algo­rithm to per­form the fla­glet trans­form rapidly, while cap­tur­ing all infor­ma­tion at each wavelet scale in the min­i­mal num­ber of sam­ples on the ball. Our imple­men­ta­tion of these new tools achieves float­ing point pre­ci­sion and is made pub­licly avail­able. We per­form numer­i­cal exper­i­ments demon­strat­ing the speed and accu­racy of these libraries and illus­trate their capa­bil­i­ties on a sim­ple denois­ing example.”
  
  wavelets geom­e­try representation-​​theory signal-​​processing answer-​​languages

• [1205.3077] Efficiency-​​Revenue Trade-​​offs in Auctions

  “When agents with inde­pen­dent pri­ors bid for a sin­gle item, Myerson’s opti­mal auc­tion max­i­mizes expected rev­enue, whereas Vickrey’s second-​​price auc­tion opti­mizes social wel­fare. We address the nat­ural ques­tion of trade-​​offs between the two cri­te­ria, that is, auc­tions that opti­mize, say, rev­enue under the con­straint that the wel­fare is above a given level. If one allows for ran­dom­ized mech­a­nisms, it is easy to see that there are polynomial-​​time mech­a­nisms that achieve any point in the trade-​​off (the Pareto curve) between rev­enue and wel­fare. We inves­ti­gate whether one can achieve the same guar­an­tees using deter­min­is­tic mech­a­nisms. We pro­vide a neg­a­tive answer to this ques­tion by show­ing that this is a (weakly) NP-​​hard prob­lem. On the pos­i­tive side, we pro­vide polynomial-​​time deter­min­is­tic mech­a­nisms that approx­i­mate with arbi­trary pre­ci­sion any point of the trade-​​off between these two fun­da­men­tal objec­tives for the case of two bid­ders, even when the val­u­a­tions are cor­re­lated arbi­trar­ily. The major prob­lem left open by our work is whether there is such an algo­rithm for three or more bid­ders with inde­pen­dent val­u­a­tion distributions.”
  
  algo­rithms Pareto-​​front performance-​​measure multiobjective-​​optimization

• Sym­bol­set

  “Sym­bol­sets are seman­tic sym­bol fonts. They work in mod­ern browsers and any­where Open­Type fea­tures are supported.”
  
  typog­ra­phy uni­code

• [1204.6653] Elim­i­na­tion of Glass Arti­facts and Object Segmentation

  “Many images nowa­days are cap­tured from behind the glasses and may have cer­tain stains dis­crep­ancy because of glass and must be processed to make dif­fer­en­ti­a­tion between the glass and objects behind it. This research paper pro­poses an algo­rithm to remove the dam­aged or cor­rupted part of the image and make it con­sis­tent with other part of the image and to seg­ment objects behind the glass. The dam­aged part is removed using total vari­a­tion inpaint­ing method and seg­men­ta­tion is done using kmeans clus­ter­ing, anisotropic dif­fu­sion and water­shed trans­for­ma­tion. The final out­put is obtained by inter­po­la­tion. This algo­rithm can be use­ful to appli­ca­tions in which some part of the images are cor­rupted due to data trans­mis­sion or needs to seg­ment objects from an image for fur­ther processing.”
  
  image-​​segmentation image-​​processing nudge-​​targets algo­rithms

• The whole of the law — Things from your life

  “But it’ll be your deci­sion, not iner­tia or fate. The ongo­ing cadence of ask­ing these ques­tions (and, maybe, the con­tent of any answers you come up with) will con­vene an open space for you to live in. A world where what­ever you do is right.”
  
  this

• The Pirate Uni­ver­sity | Pirate university

  “The Pirate Uni­ver­sity is an on-​​line bul­letin board on which stu­dents post requests for aca­d­e­mic pub­li­ca­tions. You can com­pare it to an aca­d­e­mic wish list. Oth­ers, who know where to find these pub­li­ca­tions, reply and if pos­si­ble, pro­vide links to the resources searched. The Pirate Uni­ver­sity is not pro­vid­ing, stor­ing or shar­ing copy­righted mate­r­ial. An impor­tant ques­tion is if the upload­ing of arti­cles, pub­li­ca­tions is legal. If you are the copy­right holder of the arti­cle requested, there should be no prob­lem. Also in cer­tain cases, if you or your insti­tute have acquired the rights of the pub­li­ca­tion, or if it is free of rights, there shouldn’t be a prob­lem. It is prob­a­bly best to con­sult with your librar­ian to see which kind of pub­li­ca­tion is okay to share on the Internet.”
  
  academic-​​culture pub­lish­ing col­lab­o­ra­tion crowd­sourc­ing librar­i­ans open-​​access schol­ar­ship

• [1206.3793] A dis­trib­uted classification/​estimation algo­rithm for sen­sor networks

  “…We pro­pose a novel coop­er­a­tive iter­a­tive algo­rithm which copes with the com­mu­ni­ca­tion con­straints imposed by the net­work and shows remark­able per­for­mance. Our main result is a rig­or­ous proof of the con­ver­gence of the algo­rithm and a char­ac­ter­i­za­tion of the limit behav­ior. We also show that, in the limit when the num­ber of sen­sors goes to infin­ity, the com­mon unknown para­me­ter is esti­mated with arbi­trary small error, while the clas­si­fi­ca­tion error con­verges to that of the opti­mal cen­tral­ized max­i­mum like­li­hood esti­ma­tor. We also show numer­i­cal results that val­i­date the the­o­ret­i­cal analy­sis and sup­port their pos­si­ble gen­er­al­iza­tion. We com­pare our strat­egy with the Expectation-​​Maximization algo­rithm and we dis­cuss trade-​​offs in terms of robust­ness, speed of con­ver­gence and imple­men­ta­tion simplicity.”
  
  distributed-​​processing collective-​​behavior sensor-​​networks algo­rithms nudge-​​targets

• [1204.6391] Extend­ing par­tial rep­re­sen­ta­tions of func­tion graphs and per­mu­ta­tion graphs

  “Func­tion graphs are graphs rep­re­sentable by inter­sec­tions of con­tin­u­ous real-​​valued func­tions on the inter­val [0,1] and are known to be exactly the com­ple­ments of com­pa­ra­bil­ity graphs. As such they are rec­og­niz­able in poly­no­mial time. Func­tion graphs gen­er­al­ize per­mu­ta­tion graphs, which arise when all func­tions con­sid­ered are lin­ear. We focus on the prob­lem of extend­ing par­tial rep­re­sen­ta­tions, which gen­er­al­izes the recog­ni­tion prob­lem. We observe that for per­mu­ta­tion graphs an easy exten­sion of Golumbic’s com­pa­ra­bil­ity graph recog­ni­tion algo­rithm can be exploited. This approach fails for func­tion graphs. Nev­er­the­less, we present a polynomial-​​time algo­rithm for extend­ing a par­tial rep­re­sen­ta­tion of a graph by func­tions defined on the entire inter­val [0,1] pro­vided for some of the ver­tices. On the other hand, we show that if a par­tial rep­re­sen­ta­tion con­sists of func­tions defined on subin­ter­vals of [0,1], then the prob­lem of extend­ing this rep­re­sen­ta­tion to func­tions on the entire inter­val [0,1] becomes NP-​​complete.”
  
  graph-​​theory math-i-didn’t-know representation-​​theory ontol­ogy inter­est­ing

• [1206.3294] Flex­i­ble Pri­ors for Exemplar-​​based Clustering

  “Exemplar-​​based clus­ter­ing meth­ods have been shown to pro­duce state-​​of-​​the-​​art results on a num­ber of syn­thetic and real-​​world clus­ter­ing prob­lems. They are appeal­ing because they offer com­pu­ta­tional ben­e­fits over latent-​​mean mod­els and can han­dle arbi­trary pair­wise sim­i­lar­ity mea­sures between data points. How­ever, when try­ing to recover under­ly­ing struc­ture in clus­ter­ing prob­lems, tai­lored sim­i­lar­ity mea­sures are often not enough; we also desire con­trol over the dis­tri­b­u­tion of clus­ter sizes. Pri­ors such as Dirich­let process pri­ors allow the num­ber of clus­ters to be unspec­i­fied while express­ing pri­ors over data par­ti­tions. To our knowl­edge, they have not been applied to exemplar-​​based mod­els. We show how to incor­po­rate pri­ors, includ­ing Dirich­let process pri­ors, into the recently intro­duced affin­ity prop­a­ga­tion algo­rithm. We develop an effi­cient max­prod­uct belief prop­a­ga­tion algo­rithm for our new model and demon­strate exper­i­men­tally how the expanded range of clus­ter­ing pri­ors allows us to bet­ter recover true clus­ter­ings in sit­u­a­tions where we have some infor­ma­tion about the gen­er­at­ing process.”
  
  clus­ter­ing algo­rithms

• Mag­a­zine — The Case Against Cre­den­tial­ism — The Atlantic

  ’”ALL OF OUR WORK HAS GIVEN ME A VERY STRONG view,” Richard Boy­atzis told me one after­noon. The con­sult­ing firm Boy­atzis heads, McBer and Com­pany, was founded by David McClel­land in 1963. Its spe­cialty has been ana­lyz­ing what peo­ple actu­ally do in busi­ness jobs—not what their job descrip­tions say, but how they spend their time and which skills seem most impor­tant to their suc­cess. “I’ve come to see that when­ever a group insti­tutes a cre­den­tial­ing process, whether by licens­ing or insist­ing on advanced degrees, the espoused rhetoric is to enforce the stan­dards of pro­fes­sion­al­ism. This is true whether it’s among accoun­tants or plumbers or physi­cians. But the observed con­se­quences always seem to be these two: the exclu­sion of cer­tain groups, whether by inten­tion or not, and the estab­lish­ment of mediocre per­for­mance standards.“‘
  
  pro­fes­sion­al­iza­tion cre­den­tial­ing Andrew-​​Abbott-​​smiles-​​in-​​Chicago author­ity exper­tise cultural-​​assumptions disintermediation-​​targets

• [1205.2483] Edge-​​clique graphs of cock­tail par­ties have unbounded rankwidth

  “In an attempt to find a polynomial-​​time algo­rithm for the edge-​​clique cover prob­lem on cographs we tried to prove that the edge-​​clique graphs of cographs have bounded rankwidth. How­ever, this is not the case. In this note we show that the edge-​​clique graphs of cock­tail party graphs have unbounded rank width.”
  
  open-​​questions nudge-​​targets graph-​​theory algo­rithms

• [1206.3235] Iden­ti­fy­ing rea­son­ing pat­terns in games

  “We present an algo­rithm that iden­ti­fies the rea­son­ing pat­terns of agents in a game, by iter­a­tively exam­in­ing the graph struc­ture of its Multi-​​Agent Influ­ence Dia­gram (MAID) rep­re­sen­ta­tion. If the deci­sion of an agent par­tic­i­pates in no rea­son­ing pat­terns, then we can effec­tively ignore that deci­sion for the pur­pose of cal­cu­lat­ing a Nash equi­lib­rium for the game. In some cases, this can lead to expo­nen­tial time sav­ings in the process of equi­lib­rium cal­cu­la­tion. More­over, our algo­rithm can be used to enu­mer­ate the rea­son­ing pat­terns in a game, which can be use­ful for con­struct­ing more effec­tive com­put­er­ized agents inter­act­ing with humans.”
  
  game-​​theory infer­ence strat­egy nudge-​​targets learning-​​by-​​watching