“We intro­duce a sim­ple cri­te­rion, the CAR score, for rank­ing and select­ing vari­ables in lin­ear regres­sion. The CAR score arises nat­u­rally in the best pre­dic­tor for­mu­la­tion of the lin­ear model, offers a canon­i­cal decom­po­si­tion of the pro­por­tion of explained vari­ance, and also takes account of cor­re­la­tion and group­ing struc­ture among explana­tory vari­ables. As pop­u­la­tion quan­tity the CAR score is not tied to any spe­cific infer­ence par­a­digm. Vari­able selec­tion based on AIC, $C_​p$, BIC, and other infor­ma­tion cri­te­ria is shown to be equiv­a­lent to thresh­old­ing CAR scores at a fixed level, whereas using false dis­cov­ery rates cor­re­sponds to an adap­tive cut­off. In com­puter sim­u­la­tions we show that CAR scores are highly effec­tive for vari­able selec­tion with a pre­dic­tion error that com­pares favor­able with the elas­tic net and sim­i­lar regres­sion pro­ce­dures. We illus­trate the approach by ana­lyz­ing dia­betes data as well as gene expres­sion data from the human frontal cortex.”

“A para­me­ter­i­sa­tion of gen­er­alised net­work clus­ter­ing, in the form of four-​​motif preva­lences, is pre­sented. This involves three real para­me­ters that are con­di­tional on one– two– and three-​​motif preva­lences. Inter­pre­ta­tions of these real para­me­ters are pre­sented that moti­vate a set of rewiring schemes to cre­ate appro­pri­ately clus­tered net­works. Finally, the dynam­i­cal impli­ca­tions of higher order struc­ture, as para­me­terised, for a con­tact process are considered.”

“We have devel­oped a frame­work to study the struc– ture and func­tion of com­plex net­works in purely geomet– ric terms. In this frame­work, two com­mon prop­er­ties of com­plex net­work topolo­gies, strong het­ero­gene­ity and clus­ter­ing, turn out to be sim­ple reflec­tions of the basic prop­er­ties of an under­ly­ing hyper­bolic geom­e­try. Het­ero­gene­ity, mea­sured in terms of the power-​​law degree dis­tri­b­u­tion expo­nent, is a func­tion of the neg­a­tive cur­va­ture of the hyper­bolic space, while clus­ter­ing reflects its met­ric property.”

“We describe the struc­ture of the graphs with the small­est aver­age dis­tance and the largest aver­age clus­ter­ing given their order and size. There is usu­ally a unique graph with the largest aver­age clus­ter­ing, which at the same time has the small­est pos­si­ble aver­age dis­tance. In con­trast, there are many graphs with the same min­i­mum aver­age dis­tance, ignor­ing their aver­age clus­ter­ing. The form of these graphs is shown with ana­lyt­i­cal argu­ments. Finally, we mea­sure the sen­si­tiv­ity to rewiring of this archi­tec­ture with respect to the clus­ter­ing coef­fi­cient, and we devise a method to make these net­works more robust with respect to ver­tex removal.”

“In this paper we have pro­posed and stud­ied a sim­ple model of con­tri­bu­tion games, in which agents can invest a fixed bud­get into dif­fer­ent rela­tion­ships. Our results show that col­lab­o­ra­tion between pairs of play­ers can lead to insta­bil­i­ties and non-​​existence of pair­wise equi­lib­ria. For cer­tain classes of func­tions, the exis­tence of pair­wise equi­lib­ria is even NP-​​hard to decide. This implies that it is impos­si­ble to decide effi­ciently if a set of play­ers in a game can reach a pair­wise equi­lib­rium. For many inter­est­ing classes of games, how­ever, we are able to show exis­tence and bound the price of anar­chy to 2. This includes, for instance, a class of games with gen­eral con­vex func­tions, or min­i­mum effort games with con­cave func­tions. Here we are also able to show that best response dynam­ics con­verge to pair­wise equilibria.”

“On a more philo­soph­i­cal level, our approach points at novel ques­tions that go beyond super­vised and semi-​​supervised learn­ing. What ben­e­fit do labels pro­vide over unsu­per­vised train­ing? Can our frame­work be extended to semi-​​supervised learn­ing where a few labels do exist? Can it be extended to non-​​classification sce­nar­ios such as mar­gin based regres­sion or mar­gin based struc­tured pre­dic­tion? When are the assump­tions likely to hold and how can we make our frame­work even more resis­tant to devi­a­tions from them? These ques­tions and oth­ers form new and excit­ing open research directions.”

“High-​​dimensional cor­re­lated data pose chal­lenges in model selec­tion and pre­dic­tive learn­ing. In this paper, we derive an iter­a­tive thresh­old­ing tech­nique for gen­er­al­ized lin­ear mod­els (GLMs) with pos­si­bly nonorthog­o­nal designs. We pro­pose a fam­ily of $\Theta$-estimators which are asso­ci­ated with penal­ized like­li­hoods and can be com­puted by thresholding-​​based iter­a­tive pro­ce­dures. It can also be used to robus­tify GLMs and extend the canon­i­cal $M$-estimators.…”

“The prob­lem of arriv­ing at a prin­ci­pled method of pric­ing goods and ser­vices was very sat­is­fac­to­rily solved for con­ven­tional goods; how­ever, this solu­tion is not applic­a­ble to dig­i­tal goods. After tak­ing into con­sid­er­a­tion idio­syn­crasies of the dig­i­tal realm, we give a mar­ket model that is appro­pri­ate for the dig­i­tal set­ting, and a notion of equi­lib­rium for it. We also prove exis­tence of equi­lib­rium for our mar­ket model.”

“Non­lin­ear bilat­eral fil­ters (BF) deliver a fine blend of com­pu­ta­tional sim­plic­ity and blur-​​free denois­ing. How­ever, lit­tle is known about their nature, noise-​​suppressing prop­er­ties, and opti­mal choices of fil­ter para­me­ters. Our study is meant to fill this gap-​​explaining the under­ly­ing mech­a­nism of bilat­eral fil­ter­ing and pro­vid­ing the method­ol­ogy for opti­mal fil­ter selec­tion. Prac­ti­cal appli­ca­tion to CT image denois­ing is dis­cussed to illus­trate our results.”

“For every list of inte­gers x_​1, …, x_​m there is some j such that x_​1 + … + x_​j — x_{j+1} — … — x_​m \approx 0. So the list can be nearly bal­anced and for this we only need one alter­na­tion between addi­tion and sub­trac­tion. But what if the x_​i are k-​​dimensional inte­ger vec­tors? Using results from topo­log­i­cal degree the­ory we show that bal­anc­ing is still pos­si­ble, now with k alter­na­tions.
This result is use­ful in multi-​​objective opti­miza­tion, as it allows a polynomial-​​time com­putable bal­ance of two alter­na­tives with con­flict­ing costs. The appli­ca­tion to two multi-​​objective opti­miza­tion prob­lems yields the fol­low­ing results:
– A ran­dom­ized 1/​2-​​approximation for multi-​​objective max­i­mum asym­met­ric trav­el­ing sales­man, which improves and sim­pli­fies the best known approx­i­ma­tion for this prob­lem.
– A deter­min­is­tic 1/​2-​​approximation for multi-​​objective max­i­mum weighted satisfiability.”

“In this paper we present an effi­cient com­puter aided mass clas­si­fi­ca­tion method in dig­i­tized mam­mo­grams using Arti­fi­cial Neural Net­work (ANN), which per­forms benign-​​malignant clas­si­fi­ca­tion on region of inter­est (ROI) that con­tains mass. One of the major mam­mo­graphic char­ac­ter­is­tics for mass clas­si­fi­ca­tion is tex­ture. ANN exploits this impor­tant fac­tor to clas­sify the mass into benign or malig­nant. The sta­tis­ti­cal tex­tural fea­tures used in char­ac­ter­iz­ing the masses are mean, stan­dard devi­a­tion, entropy, skew­ness, kur­to­sis and uniformity.…”

“In this paper fusion of visual and ther­mal images in wavelet trans­formed domain has been pre­sented. Here, Daubechies wavelet trans­form, called as D2, coef­fi­cients from visual and cor­re­spond­ing coef­fi­cients com­puted in the same man­ner from ther­mal images are com­bined to get fused coef­fi­cients. After decom­po­si­tion up to fifth level (Level 5) fusion of coef­fi­cients is done. Inverse Daubechies wavelet trans­form of those coef­fi­cients gives us fused face images. The main advan­tage of using wavelet trans­form is that it is well-​​suited to man­age dif­fer­ent image res­o­lu­tion and allows the image decom­po­si­tion in dif­fer­ent kinds of coef­fi­cients, while pre­serv­ing the image information.…”

“In this paper we present a sim­ple novel approach to tackle the chal­lenges of scal­ing and rota­tion of face images in face recog­ni­tion. The pro­posed approach reg­is­ters the train­ing and test­ing visual face images by log-​​polar trans­for­ma­tion, which is capa­ble to han­dle com­pli­ca­cies intro­duced by scal­ing and rota­tion. Log-​​polar images are pro­jected into eigen­space and finally clas­si­fied using an improved multi-​​layer per­cep­tron. In the exper­i­ments we have used ORL face data­base and Object Track­ing and Clas­si­fi­ca­tion Beyond Vis­i­ble Spec­trum (OTCBVS) data­base for visual face images. Exper­i­men­tal results show that the pro­posed approach sig­nif­i­cantly improves the recog­ni­tion per­for­mances from visual to log-​​polar-​​visual face images. …”

“Here an effi­cient fusion tech­nique for auto­matic face recog­ni­tion has been pre­sented. Fusion of visual and ther­mal images has been done to take the advan­tages of ther­mal images as well as visual images. By employ­ing fusion a new image can be obtained, which pro­vides the most detailed, reli­able, and dis­crim­i­nat­ing infor­ma­tion. In this method fused images are gen­er­ated using visual and ther­mal face images in the first step. In the sec­ond step, fused images are pro­jected into eigen­space and finally clas­si­fied using a radial basis func­tion neural net­work. In the exper­i­ments Object Track­ing and Clas­si­fi­ca­tion Beyond Vis­i­ble Spec­trum (OTCBVS) data­base bench­mark for ther­mal and visual face images have been used. Exper­i­men­tal results show that the pro­posed approach per­forms well in rec­og­niz­ing unknown indi­vid­u­als with a max­i­mum suc­cess rate of 96%.”

“The effi­ciency of a sim­ple model of cross­flow fan is max­i­mized when the geom­e­try depends on a design para­me­ter. The flow field is numer­i­cally com­puted using a Galerkin method for solv­ing a Pois­son par­tial dif­fer­en­tial equation.”

“In this paper, we tackle the prob­lem of inno­va­tion spread­ing from a mod­el­ing point of view. We con­sider a net­worked sys­tem of indi­vid­u­als, with a com­pe­ti­tion between two groups. We show its rela­tion to the inno­va­tion spread­ing issues. We intro­duce an abstract model and show how it can be inter­preted in this frame­work, as well as what con­clu­sions we can draw form it. We fur­ther explain how model-​​derived con­clu­sions can help to inves­ti­gate the orig­i­nal prob­lem, as well as other, sim­i­lar prob­lems. The model is an agent-​​based model assum­ing sim­ple binary attrib­utes of those agents. It uses a major­ity dynam­ics (Ising model to be exact), mean­ing that indi­vid­u­als attempt to be sim­i­lar to the major­ity of their peers, bar­ring the occa­sional purely indi­vid­ual deci­sions that are mod­eled as ran­dom. We show that this sim­plis­tic model can be related to the decision-​​making dur­ing inno­va­tion adop­tion processes. …”

“This paper pro­poses some exten­sions to the work on ker­nels ded­i­cated to string align­ment (bio­log­i­cal sequence align­ment) based on the sum­ming up of scores obtained by local align­ments with gaps. The exten­sions we pro­pose allow to con­struct, from clas­si­cal time-​​warp dis­tances, what we called sum­ma­tive time-​​warp ker­nels that are pos­i­tive def­i­nite if some sim­ple suf­fi­cient con­di­tions are sat­is­fied. Fur­ther­more, from the same for­mal­ism, we derive a time-​​warp inner prod­uct that extends the usual euclid­ean inner prod­uct, pro­vid­ing the capa­bil­ity to han­dle dis­crete sequences or time series of vari­able lengths in an Hilbert space. The clas­si­fi­ca­tion exper­i­ment we con­ducted, using either first near neigh­bor clas­si­fier or Sup­port Vec­tor Machine clas­si­fier leads to con­clude that the pos­i­tive def­i­nite elas­tic ker­nels we pro­pose out­per­form the dis­tance sub­sti­tut­ing ker­nels for the clas­si­cal elas­tic dis­tances we tested.…”

“Improve­ments in tech­nique in con­junc­tion with an evo­lu­tion of the the­o­ret­i­cal and con­cep­tual approach to neu­ronal net­works pro­vide a new per­spec­tive on liv­ing neu­rons in cul­ture. Orga­ni­za­tion and con­nec­tiv­ity are being mea­sured quan­ti­ta­tively along with other phys­i­cal quan­ti­ties such as infor­ma­tion, and are being related to func­tion. In this review we first dis­cuss some of these advances, which enable elu­ci­da­tion of struc­tural aspects. We then dis­cuss two recent exper­i­men­tal mod­els that yield some con­cep­tual simplicity.…”

“A sin­gle tar­get is hid­den at a loca­tion cho­sen from a pre­de­ter­mined prob­a­bil­ity dis­tri­b­u­tion. Then, a searcher must find a sec­ond prob­a­bil­ity dis­tri­b­u­tion from which ran­dom search points are sam­pled such that the tar­get is found in the min­i­mum num­ber of tri­als. Here it will be shown that if the searcher must get very close to the tar­get to find it, then the best search dis­tri­b­u­tion is pro­por­tional to the square root of the tar­get distribution…”

“Net­works por­tray a mul­ti­tude of inter­ac­tions through which peo­ple meet, ideas are spread, and infec­tious dis­eases prop­a­gate within a soci­ety. Iden­ti­fy­ing the most effi­cient “spread­ers” in a net­work is an impor­tant step to opti­mize the use of avail­able resources and ensure the more effi­cient spread of infor­ma­tion. Here we show that, in con­trast to com­mon belief, the most influ­en­tial spread­ers in a social net­work do not cor­re­spond to the best con­nected peo­ple or to the most cen­tral peo­ple (high between­ness cen­tral­ity). Instead, we find: (i) The most effi­cient spread­ers are those located within the core of the net­work as iden­ti­fied by the k-​​shell decom­po­si­tion analy­sis. (ii) When mul­ti­ple spread­ers are con­sid­ered simul­ta­ne­ously, the dis­tance between them becomes the cru­cial para­me­ter that deter­mines the extend of the spreading.…”

“We study the prob­lem of sched­ul­ing peri­odic real-​​time tasks so as to meet their indi­vid­ual min­i­mum reward require­ments. A task gen­er­ates jobs that can be given arbi­trary ser­vice times before their dead­lines. A task then obtains rewards based on the ser­vice times received by its jobs. We show that this model is com­pat­i­ble to the impre­cise com­pu­ta­tion mod­els and the increas­ing reward with increas­ing ser­vice mod­els. In con­trast to pre­vi­ous work on these mod­els, which mainly focus on max­i­mize the total reward in the sys­tem, we aim to ful­fill dif­fer­ent reward require­ments by dif­fer­ent tasks, which offers bet­ter fair­ness and allows fine-​​grained trade­off between tasks. We first derive a nec­es­sary and suf­fi­cient con­di­tion for a sys­tem, along with reward require­ments of tasks, to be fea­si­ble. We also obtain an off-​​line fea­si­bil­ity opti­mal sched­ul­ing policy.…”

“Work­ing in Winfree’s abstract tile assem­bly model, we show that a constant-​​size tile assem­bly sys­tem can be pro­grammed through rel­a­tive tile con­cen­tra­tions to build an n x n square with high prob­a­bil­ity, for any suf­fi­ciently large n. This answers an open ques­tion of Kao and Schweller (Ran­dom­ized Self-​​Assembly for Approx­i­mate Shapes, ICALP 2008), who showed how to build an approx­i­mately n x n square using tile con­cen­tra­tion pro­gram­ming, and asked whether the approx­i­ma­tion could be made exact with high prob­a­bil­ity. We show how this tech­nique can be mod­i­fied to answer another ques­tion of Kao and Schweller, by show­ing that a constant-​​size tile assem­bly sys­tem can be pro­grammed through tile con­cen­tra­tions to assem­ble arbi­trary finite *scaled shapes*, which are shapes mod­i­fied by replac­ing each point with a c x c block of points, for some inte­ger c. …”

“e devel­oped a low-​​cost, high-​​throughput micro­biome pro­fil­ing method that uses com­bi­na­to­r­ial sequence tags attached to PCR primers that amplify the rRNA V6 region. Ampli­fied PCR prod­ucts are sequenced using an Illu­mina paired-​​end pro­to­col to gen­er­ate mil­lions of over­lap­ping reads. Com­bi­na­to­r­ial sequence tag­ging can be used to exam­ine hun­dreds of sam­ples with far fewer primers than is required when sequence tags are incor­po­rated at only a sin­gle end.…”

“Tem­plate match­ing is one of the sim­plest meth­ods used for eyes and mouth detec­tion. How­ever, it can be mod­i­fied and extended to become a pow­er­ful tool. Since the patch itself plays a sig­nif­i­cant role in opti­miz­ing detec­tion per­for­mance, a study on the influ­ence of patch size and shape is car­ried out. The opti­mum patch size and shape is deter­mined using the pro­posed method. Usu­ally, tem­plate match­ing is also com­bined with other meth­ods in order to improve detec­tion accu­racy. Thus, in this paper, the effec­tive­ness of two image pro­cess­ing meth­ods i.e. grayscale and Haar wavelet trans­form, when used with tem­plate match­ing are analyzed.”

“In this paper we present a com­par­a­tive study on fusion of visual and ther­mal images using dif­fer­ent wavelet trans­for­ma­tions. Here, coef­fi­cients of dis­crete wavelet trans­forms from both visual and ther­mal images are com­puted sep­a­rately and com­bined. Next, inverse dis­crete wavelet trans­for­ma­tion is taken in order to obtain fused face image. Both Haar and Daubechies (db2) wavelet trans­forms have been used to com­pare recog­ni­tion results. For exper­i­ments IRIS Thermal/​Visual Face Data­base was used. Exper­i­men­tal results using Haar and Daubechies wavelets show that the per­for­mance of the approach pre­sented here achieves max­i­mum suc­cess rate of 100% in many cases.”

“As we have shown, Clus­ter Rever­ber­a­tion is a mech­a­nism avail­able to neural sys­tems for robust short-​​term mem­ory with­out synap­tic learn­ing. To the best of our knowl­edge, this is the first mech­a­nism pro­posed which has these charac– ter­is­tics – essen­tial for, say, sen­sory mem­ory or cer­tain working-​​memory tasks. All that is needed is for the net­work topol­ogy to be highly clus­tered or modu– lar, and for small groups of neu­rons to store one bit of infor­ma­tion, as opposed to the con­ven­tional view which assumes one bit per neu­ron. Con­sid­er­ing the enor­mous num­ber of neu­rons in the brain, and the fact that real indi­vid­ual neu– rons are prob­a­bly too noisy to store infor­ma­tion reli­ably, these hypothe­ses do not seem farfetched.…”

“A pop­u­lar method for select­ing the num­ber of clus­ters is based on sta­bil­ity argu­ments: one chooses the num­ber of clus­ters such that the cor­re­spond­ing clus­ter­ing results are “most sta­ble”. In recent years, a series of papers has ana­lyzed the behav­ior of this method from a the­o­ret­i­cal point of view. How­ever, the results are very tech­ni­cal and dif­fi­cult to inter­pret for non-​​experts. In this paper we give a high-​​level overview about the exist­ing lit­er­a­ture on clus­ter­ing sta­bil­ity. In addi­tion to pre­sent­ing the results in a slightly infor­mal but acces­si­ble way, we relate them to each other and dis­cuss their dif­fer­ent implications.”

“We estab­lish con­crete math­e­mat­i­cal cri­te­ria to dis­tin­guish between dif­fer­ent kinds of writ­ten sto­ry­telling, fic­tional and non-​​fictional. Specif­i­cally, we con­structed a seman­tic net­work from both nov­els and news sto­ries, with $N$ inde­pen­dent words as ver­tices or nodes, and edges or links allot­ted to words occur­ring within $m$ places of a given ver­tex; we call $m$ the word dis­tance. We then used mea­sures from com­plex net­work the­ory to dis­tin­guish between news and fic­tion, study­ing the min­i­mal text length needed as well as the opti­mized word dis­tance $m$. The lit­er­a­ture sam­ples were found to be most effec­tively rep­re­sented by their cor­re­spond­ing power laws over degree dis­tri­b­u­tion $P(k)$ and clus­ter­ing coef­fi­cient $C(k)$; we also stud­ied the mean geo­desic dis­tance, and found all our texts were small-​​world networks.…”