We investigate the task of 2D articulated human pose estimation in unconstrained still images. This is extremely challenging because of variation in pose, anatomy, clothing, and imaging conditions. Current methods use simple models of body part appearance and plausible configurations due to limitations of available training data and constraints on computational expense. We show that such models severely limit accuracy. Building on the successful pictorial structure model (PSM) we propose richer models of both appearance and pose, using state-of-the-art discriminative classifiers without introducing unacceptable computational expense. We introduce a new annotated database of challenging consumer images, an order of magnitude larger than currently available datasets, and demonstrate over 50% relative improvement in pose estimation accuracy over a stateof-the-art method.

No abstract

Why are large, complex ecosystems stable? Both theory and simulations of current models predict the onset of instability with growing size and complexity, so for decades it has been conjectured that ecosystems must have some unidentified structural property exempting them from this outcome. We show that trophic coherence-a hitherto ignored feature of food webs that current structural models fail to reproduce-is a better statistical predictor of linear stability than size or complexity. Furthermore, we prove that a maximally coherent network with constant interaction strengths will always be linearly stable. We also propose a simple model that, by correctly capturing the trophic coherence of food webs, accurately reproduces their stability and other basic structural features. Most remarkably, our model shows that stability can increase with size and complexity. This suggests a key to May's paradox, and a range of opportunities and concerns for biodiversity conservation.food webs | May's paradox | diversity-stability debate | dynamical stability | complex networks I n the early seventies, Robert May addressed the question of whether a generic system of coupled dynamical elements randomly connected to each other would be stable. He found that the larger and more interconnected the system, the more difficult it would be to stabilize (1, 2). His deduction followed from the behavior of the leading eigenvalue of the interaction matrix, which, in a randomly wired system, grows with the square root of the mean number of links per element. This result clashed with the received wisdom in ecology-that large, complex ecosystems were particularly stable-and initiated the "diversity-stability debate" (3-6). Indeed, Charles Elton had expressed the prevailing view in 1958: "the balance of relatively simple communities of plants and animals is more easily upset than that of richer ones; that is, more subject to destructive oscillations in populations, especially of animals, and more vulnerable to invasions" (7). Even if this description were not accurate, the mere existence of rainforests and coral reefs seems incongruous with a general mathematical principle that "complexity begets instability," and has become known as May's paradox.One solution might be that the linear stability analysis used by May and many subsequent studies does not capture essential characteristics of ecosystem dynamics, and much work has gone into exploring how more accurate dynamical descriptions might enhance stability (5,8,9). However, as ever-better ecological data are gathered, it is becoming apparent that the leading eigenvalues of matrices related to food webs (networks in which the species are nodes and the links represent predation) do not exhibit the expected dependence on size or link density (10). Food webs must, therefore, have some unknown structural feature that accounts for this deviation from randomness-irrespectively of other stabilizing factors.We show here that a network feature we call trophic coherence accounts for much of the varianc...

Why are most empirical networks, with the prominent exception of social ones, generically degree-degree anticorrelated? To answer this long-standing question, we define the ensemble of correlated networks and obtain the associated Shannon entropy. Maximum entropy can correspond to either assortative (correlated) or disassortative (anticorrelated) configurations, but in the case of highly heterogeneous, scale-free networks a certain disassortativity is predicted--offering a parsimonious explanation for the question above. Our approach provides a neutral model from which, in the absence of further knowledge regarding network evolution, one can obtain the expected value of correlations. When empirical observations deviate from the neutral predictions--as happens for social networks--one can then infer that there are specific correlating mechanisms at work.

The temporal envelope of speech is important for speech intelligibility. Entrainment of cortical oscillations to the speech temporal envelope is a putative mechanism underlying speech intelligibility. Here we used magnetoencephalography (MEG) to test the hypothesis that phase-locking to the speech temporal envelope is enhanced for intelligible compared with unintelligible speech sentences. Perceptual "pop-out" was used to change the percept of physically identical tone-vocoded speech sentences from unintelligible to intelligible. The use of pop-out dissociates changes in phase-locking to the speech temporal envelope arising from acoustical differences between un/intelligible speech from changes in speech intelligibility itself. Novel and bespoke whole-head beamforming analyses, based on significant cross-correlation between the temporal envelopes of the speech stimuli and phase-locked neural activity, were used to localize neural sources that track the speech temporal envelope of both intelligible and unintelligible speech. Location-of-interest analyses were carried out in a priori defined locations to measure the representation of the speech temporal envelope for both un/intelligible speech in both the time domain (cross-correlation) and frequency domain (coherence). Whole-brain beamforming analyses identified neural sources phase-locked to the temporal envelopes of both unintelligible and intelligible speech sentences. Crucially there was no difference in phase-locking to the temporal envelope of speech in the pop-out condition in either the whole-brain or location-of-interest analyses, demonstrating that phase-locking to the speech temporal envelope is not enhanced by linguistic information.