Alphabet (United States)

Adversarial examples are perturbed inputs designed to fool machine learning models. Adversarial training injects such examples into training data to increase robustness. To scale this technique to large datasets, perturbations are crafted using fast single-step methods that maximize a linear approximation of the model's loss. We show that this form of adversarial training converges to a degenerate global minimum, wherein small curvature artifacts near the data points obfuscate a linear approximation of the loss. The model thus learns to generate weak perturbations, rather than defend against strong ones. As a result, we find that adversarial training remains vulnerable to black-box attacks, where we transfer perturbations computed on undefended models, as well as to a powerful novel single-step attack that escapes the non-smooth vicinity of the input data via a small random step. We further introduce Ensemble Adversarial Training, a technique that augments training data with perturbations transferred from other models. On ImageNet, Ensemble Adversarial Training yields models with strong robustness to black-box attacks. In particular, our most robust model won the first round of the NIPS 2017 competition on Defenses against Adversarial Attacks. However, subsequent work found that more elaborate black-box attacks could significantly enhance transferability and reduce the accuracy of our models.

Biomarkers are critical to the rational development of medical therapeutics, but significant confusion persists regarding fundamental definitions and concepts involved in their use in research and clinical practice, particularly in the fields of chronic disease and nutrition. Clarification of the definitions of different biomarkers and a better understanding of their appropriate application could result in substantial benefits. This review examines biomarker definitions recently established by the U.S. Food and Drug Administration and the National Institutes of Health as part of their joint Biomarkers, EndpointS, and other Tools (BEST) resource. These definitions are placed in context of their respective uses in patient care, clinical research, or therapeutic development. We explore the distinctions between biomarkers and clinical outcome assessments and discuss the specific definitions and applications of diagnostic, monitoring, pharmacodynamic/response, predictive, prognostic, safety, and susceptibility/risk biomarkers. We also explore the implications of current biomarker development trends, including complex composite biomarkers and digital biomarkers derived from sensors and mobile technologies. Finally, we discuss the challenges and potential benefits of biomarker-driven predictive toxicology and systems pharmacology, the need to ensure quality and reproducibility of the science underlying biomarker development, and the importance of fostering collaboration across the entire ecosystem of medical product development. Impact statement Biomarkers are critical to the rational development of medical diagnostics and therapeutics, but significant confusion persists regarding fundamental definitions and concepts involved in their use in research and clinical practice. Clarification of the definitions of different biomarker classes and a better understanding of their appropriate application could yield substantial benefits. Biomarker definitions recently established in a joint FDA-NIH resource place different classes of biomarkers in the context of their respective uses in patient care, clinical research, or therapeutic development. Complex composite biomarkers and digital biomarkers derived from sensors and mobile technologies, together with biomarker-driven predictive toxicology and systems pharmacology, are reshaping development of diagnostic and therapeutic technologies. An approach to biomarker development that prioritizes the quality and reproducibility of the science underlying biomarker development and incorporates collaborative regulatory science involving multiple disciplines will lead to rational, evidence-based biomarker development that keeps pace with scientific and clinical need.

1. The electrical activity of up to eight concurrently active motor units has been recorded from the human deltoid and first dorsal interosseous (f.d.i.) muscles. The detected myoelectric signals have been decomposed into their constituent motor-unit action potential trains using a recently developed technique.2. Concurrently active motor unit behaviour has been examined during triangular force-varying isometric contractions reaching 40 and 80% of maximal voluntary contraction (m.v.c.). Experiments were performed on four normal subjects and three groups of highly trained performers (long-distance swimmers, powerlifters and pianists).3. Results revealed a highly ordered recruitment and decruitment scheme, based on motoneurone excitability, in both muscles and in all subject groups.4. Differences were observed between the initial (recruitment) and final (decruitment) firing rates in each muscle. These parameters were invariant with respect to the force rates studied, although some differences were observed among subject groups.5. In general, firing rates of f.d.i. motor units increased steadily with increasing force (up to 80% m.v.c.). The firing rates of deltoid motor units rose sharply just after recruitment and then increased only slightly thereafter.6. Recruitment was found to be the major mechanism for generating extra force between 40 and 80% m.v.c. in the deltoid, while rate coding played the major role in the f.d.i.7. The potential of rate coding for increasing force levels up to m.v.c. is discussed.

Innovations like domain-specific hardware, enhanced security, open instruction sets, and agile chip development will lead the way.

We describe, analyze, and experiment with a framework for empirical loss minimization with regularization. Our algorithmic framework alternates between two phases. On each iteration we first perform an unconstrained gradient descent step. We then cast and solve an instantaneous optimization problem that trades off minimization of a regularization term while keeping close proximity to the result of the first phase. This view yields a simple yet effective algorithm that can be used for batch penalized risk minimization and online learning. Furthermore, the two phase approach enables sparse solutions when used in conjunction with regularization functions that promote sparsity, such as ℓ1. We derive concrete and very simple algorithms for minimization of loss functions with ℓ1, ℓ2, ℓ 2 2, and ℓ ∞ regularization. We also show how to construct efficient algorithms for mixed-norm ℓ1/ℓq regularization. We further extend the algorithms and give efficient implementations for very high-dimensional data with sparsity. We demonstrate the potential of the proposed framework in a series of experiments with synthetic and natural data sets.

1. The electrical activity of up to eight concurrently active motor units has been recorded from the human deltoid and first dorsal interosseous muscles. The resulting composite myoelectric signals have been decomposed into their constituent motor-unit action potential trains using a recently developed technique.2. A computer cross-correlation analysis has been performed on motor-unit firing rate and muscle-force output records obtained from both constant-force and triangular force-varying isometric contractions performed by normal subjects, and three groups of highly trained performers (long-distance swimmers, powerlifters and pianists).3. The temporal relationships between firing rate activity and force output have provided evidence that the deltoid of long-distance swimmers has a significantly higher percentage of slowly fatiguing fibres than that of normal subjects.4. Results showed that both muscles are incapable of producing a purely isotonic contraction under isometric conditions. Small, possibly compensatory force variations at 1-2 Hz result from a common drive to all active motoneurones in a single muscle pool.5. Rapid force reversals during triangular, force-varying isometric contractions appear to be accomplished through a size-related motor-unit control scheme. All firing rates decline prior to the force peak, but small motor units with slow-twitch responses tend to decrease their firing rates before large, fast-twitch motor units. This mechanism is not visually controlled, and does not depend on force rate in non-ballistic contractions.

Many of the recent successful methods for video object segmentation (VOS) are overly complicated, heavily rely on fine-tuning on the first frame, and/or are slow, and are hence of limited practical use. In this work, we propose FEELVOS as a simple and fast method which does not rely on fine-tuning. In order to segment a video, for each frame FEELVOS uses a semantic pixel-wise embedding together with a global and a local matching mechanism to transfer information from the first frame and from the previous frame of the video to the current frame. In contrast to previous work, our embedding is only used as an internal guidance of a convolutional network. Our novel dynamic segmentation head allows us to train the network, including the embedding, end-to-end for the multiple object segmentation task with a cross entropy loss. We achieve a new state of the art in video object segmentation without fine-tuning with a J&F measure of 71.5% on the DAVIS 2017 validation set. We make our code and models available at https://github.com/tensorflow/models/tree/master/research/feelvos.

Abstract Groupers are a valuable fishery resource of reef ecosystems and are among those species most vulnerable to fishing pressure because of life history characteristics including longevity, late sexual maturation and aggregation spawning. Despite their economic importance, few grouper fisheries are regularly monitored or managed at the species level, and many are reported to be undergoing declines. To identify major threats to groupers, the International Union for Conservation of Nature (IUCN) Red List criteria were applied to all 163 species. Red List assessments show that 20 species (12%) risk extinction if current trends continue, and an additional 22 species (13%) are considered to be Near Threatened. The Caribbean Sea, coastal Brazil and Southeast Asia contain a disproportionate number of Threatened species, while numerous poorly documented and Near Threatened species occur in many regions. In all, 30% of all species are considered to be Data Deficient. Given that the major threat is overfishing, accompanied by a general absence and/or poor application of fishery management, the prognosis for restoration and successful conservation of Threatened species is poor. We believe that few refuges remain for recovery and that key biological processes (e.g. spawning aggregations) continue to be compromised by uncontrolled fishing. Mariculture, through hatchery‐rearing, increases production of a few species and contributes to satisfying high market demand, but many such operations depend heavily on wild‐caught juveniles with resultant growth and recruitment overfishing. Better management of fishing and other conservation efforts are urgently needed, and we provide examples of possible actions and constraints.

The standard recurrent neural network language model (rnnlm) generates sentences one word at a time and does not work from an explicit global sentence representation. In this work, we introduce and study an rnn-based variational autoencoder generative model that incorporates distributed latent representations of entire sentences. This factorization allows it to explicitly model holistic properties of sentences such as style, topic, and high-level syntactic features. Samples from the prior over these sentence representations remarkably produce diverse and well-formed sentences through simple deterministic decoding. By examining paths through this latent space, we are able to generate coherent novel sentences that interpolate between known sentences. We present techniques for solving the difficult learning problem presented by this model, demonstrate its effectiveness in imputing missing words, explore many interesting properties of the model's latent sentence space, and present negative results on the use of the model in language modeling. but now , as they parked out front and owen stepped out of the car , he could see True: that the transition was complete . RNNLM: it , " i said . VAE: through the driver 's door . you kill him and his True: men .

A key learning scenario in large-scale applications is that of federated learning, where a centralized model is trained based on data originating from a large number of clients. We argue that, with the existing training and inference, federated models can be biased towards different clients. Instead, we propose a new framework of agnostic federated learning, where the centralized model is optimized for any target distribution formed by a mixture of the client distributions. We further show that this framework naturally yields a notion of fairness. We present data-dependent Rademacher complexity guarantees for learning with this objective, which guide the definition of an algorithm for agnostic federated learning. We also give a fast stochastic optimization algorithm for solving the corresponding optimization problem, for which we prove convergence bounds, assuming a convex loss function and hypothesis set. We further empirically demonstrate the benefits of our approach in several datasets. Beyond federated learning, our framework and algorithm can be of interest to other learning scenarios such as cloud computing, domain adaptation, drifting, and other contexts where the training and test distributions do not coincide.

We present a new method for regularized convex optimization and analyze it under both online and stochastic optimization settings. In addition to unifying previously known first-order algorithms, such as the projected gradient method, mirror descent, and forward-backward splitting, our method yields new analysis and algorithms. We also derive specific instantiations of our method for commonly used regularization functions, such as `1, mixed norm, and trace-norm. 1

Thermoelectric materials for power generation are typically compared using the dimensionless figure-of-merit ZT because it relates directly to the device efficiency. However, for practical applications, the cost of power generation – as governed by material, manufacturing, and heat exchanger costs – is also a critical factor which is not captured in ZT alone. The necessary analysis, derived herein, optimizes the coupled thermoelectric and economic problem for the leg length, L, and system fill factor, F, as functions of these costs. Fuel, operating, and maintenance costs are excluded. This optimization yields the minimum $ per W value for thermoelectric power generation and a framework for comparing materials beyond ZT. This analysis shows that even very expensive thermoelectric materials have the potential to be the most cost effective at the system level, if incorporated with sufficiently short legs and small fill factor. An approximate scaling analysis, verified using numerical calculations, gives the first closed-form, analytical expressions for optimal L and F to minimize $ per W. The analysis also delineates various cost-dominant regimes with different priorities for materials development, including: (i) a heat exchanger cost dominated regime, where ZT should be increased regardless of material or manufacturing costs; (ii) an areal cost, C′′, dominated regime at fixed F, where ZT/C′′ should be maximized, and (iii) a volumetric cost, C′′′, dominated regime at fixed F, where ZT/(kC′′′) should be maximized, reinforcing the need for low thermal conductivity, k. The cost–performance framework derived here will be applied to a number of real materials and applications in a separate manuscript.

A key learning scenario in large-scale applications is that of federated learning, where a centralized model is trained based on data originating from a large number of clients. We argue that, with the existing training and inference, federated models can be biased towards different clients. Instead, we propose a new framework of agnostic federated learning, where the centralized model is optimized for any target distribution formed by a mixture of the client distributions. We further show that this framework naturally yields a notion of fairness. We present data-dependent Rademacher complexity guarantees for learning with this objective, which guide the definition of an algorithm for agnostic federated learning. We also give a fast stochastic optimization algorithm for solving the corresponding optimization problem, for which we prove convergence bounds, assuming a convex loss function and hypothesis set. We further empirically demonstrate the benefits of our approach in several datasets. Beyond federated learning, our framework and algorithm can be of interest to other learning scenarios such as cloud computing, domain adaptation, drifting, and other contexts where the training and test distributions do not coincide.

Importance: Clinical decisions are ideally based on evidence generated from multiple randomized controlled trials (RCTs) evaluating clinical outcomes, but historically, few clinical guideline recommendations have been based entirely on this type of evidence. Objective: To determine the class and level of evidence (LOE) supporting current major cardiovascular society guideline recommendations, and changes in LOE over time. Data Sources: Current American College of Cardiology/American Heart Association (ACC/AHA) and European Society of Cardiology (ESC) clinical guideline documents (2008-2018), as identified on cardiovascular society websites, and immediate predecessors to these guideline documents (1999-2014), as referenced in current guideline documents. Study Selection: Comprehensive guideline documents including recommendations organized by class and LOE. Data Extraction and Synthesis: The number of recommendations and the distribution of LOE (A [supported by data from multiple RCTs or a single, large RCT], B [supported by data from observational studies or a single RCT], and C [supported by expert opinion only]) were determined for each guideline document. Main Outcomes and Measures: The proportion of guideline recommendations supported by evidence from multiple RCTs (LOE A). Results: Across 26 current ACC/AHA guidelines (2930 recommendations; median, 121 recommendations per guideline [25th-75th percentiles, 76-155]), 248 recommendations (8.5%) were classified as LOE A, 1465 (50.0%) as LOE B, and 1217 (41.5%) as LOE C. The median proportion of LOE A recommendations was 7.9% (25th-75th percentiles, 0.9%-15.2%). Across 25 current ESC guideline documents (3399 recommendations; median, 130 recommendations per guideline [25th-75th percentiles, 111-154]), 484 recommendations (14.2%) were classified as LOE A, 1053 (31.0%) as LOE B, and 1862 (54.8%) as LOE C. When comparing current guidelines with prior versions, the proportion of recommendations that were LOE A did not increase in either ACC/AHA (median, 9.0% [current] vs 11.7% [prior]) or ESC guidelines (median, 15.1% [current] vs 17.6% [prior]). Conclusions and Relevance: Among recommendations in major cardiovascular society guidelines, only a small percentage were supported by evidence from multiple RCTs or a single, large RCT. This pattern does not appear to have meaningfully improved from 2008 to 2018.

Although population viability analysis (PVA) is widely used in setting conservation policy, there is disagreement about the usefulness of this method. Objections have been raised concerning the precision of predictions in view of the short time series of data available and the sensitivity of estimates of extinction risk to estimated parameters ( Hamilton & Moller 1995; Taylor 1995; Groom & Pascual 1998; Ludwig 1999). Beissinger and Westphal (1998) reviewed the use of demographic models for endangered-species management. They pointed out that poor data cause difficulties in parameter estimation, which in turn lead to unreliable estimates of extinction risk. There are additional problems with model validation, especially if all available data have been used to estimate parameters. Beissinger and Westphal (1998) recommend that PVA be used to evaluate relative rather than absolute extinction risk, that projections be made only over short time periods, and that simple models be used rather than complicated ones. Fieberg and Ellner (2000) showed that values of the quasi-extinction probability—the probability of decline to a lower population threshold—for a simple model range between 80% and 5% as the value of the intrinsic growth rate r varies between −0.03 and +0.02. Such a range in estimates of r is common for data sets of moderate size. They also show that a precise estimate of extinction probability over a horizon of t years requires between 5t and 10t years of data, and that similar results hold for age-structured models. In a recent article, Brook et al. (2000) used field data on declining species to test the accuracy and bias of PVA models for predicting extinction risk and concluded that “PVA is a valid and sufficiently accurate tool for categorizing and managing endangered species.” We examined the reasons for these differing assessments of the value of PVA. Brook et al. (2000) considered 21 long-term data sets (11–57 years, mean = 24.9). They used the first half of each set to estimate parameters, with a variety of PVA software packages. They used the second half of each set to test the predictions of each package. The predictions were tested by comparing the actual and predicted numbers of species that declined below a given threshold abundance, which was defined by specifying a target risk level and using the PVA model to identify the corresponding threshold. They applied a significance test to these differences between the predicted and actual number of species falling below the thresholds, and failed to detect any significant differences (their Table 1 & Fig. 1). They applied similar tests to final population sizes with analogous results. Results from simulated population viability analyses (PVA) using parameters based on Tables 2 and 3 in the supplementary material of Brook et al. (2000). (a) Comparison of observed and predicted total number of extinctions for a set of 21 species, as in Fig. 1 of Brook et al. (2000), based on the unstructured model described in the text. Five replicates are plotted. For each of the 105 trials (21 species × 5 replicates), we simulated a PVA as described in the text. (b) Comparison of actual and estimated extinction risks for each of the 105 trials used in panel (a). Actual extinction risks were calculated by running 25,000 simulations of the model using the true parameter values and recording the fraction of runs crossing below each estimated threshold from (a). Dashed lines show the tenth and ninetieth percentiles of the distributions of estimated extinction risks. (c) Comparison of observed and predicted total number of extinctions, as in panel (a), based on the age-structured, density-dependent model described in the text. All simulations began with a population in stable age distribution for the mean matrix with a population density of 500. (d) Comparison of actual and estimated extinction risks, as in panel (b), for the 105 trials with the age-structured, density-dependent model. Brook et al. (2000) tested PVA models only on actual field data, whereas other authors used simulated data. Various methods for testing PVA predictions with field data are reviewed by McCarthy and Broome (2000) and McCarthy et al. (2001). As McCarthy et al. (2001) emphasize, a valid test must be based on data that were not used to fit the model. In view of the amount and quality of data necessary for parameterizing a complex population model, field data sets adequate for both parameterizing and testing a model will generally be scarce. In contrast, simulated data allow for the replication necessary to evaluate the precision of model-derived estimates relative to the true extinction risk or population growth rate, which are known exactly for a simulated population. Simulated data may also have shortcomings. Taylor et al. (2000) discuss the merits of using simulated data for model testing. Modern statistical practice requires that every statistical estimate be accompanied by a measure of its precision if inferences are to be drawn from these estimates (Sokal & Rohlf 1981). This general principle has special force for extinction-risk estimates based on PVA, for which investigations have repeatedly shown lack of precision. For complicated statistical models such as those used in PVAs, there may be no way to derive confidence intervals analytically, but they are readily obtained from simulations. White (2000) and Sæther et al. (2000) show how to use simulations to obtain confidence intervals for complicated models. Alternatively, confidence intervals can be obtained for some models by repeatedly resampling real data (Sokal & Rohlf 1981). Brook et al. (2000) do not estimate confidence intervals for their extinction-risk estimates, leaving the precision of their estimates unclear. The tests of Brook et al. (2000) are applied to an ensemble of species rather than to individual species. Such a test provides information about the bias in the risk estimates, but it provides little information about their precision because the expected total number of extinctions depends only on the average risk over the ensemble. We illustrate this distinction with two models ( Fig. 1) based on Tables 2 and 3 in the supplementary material of Brook et al. (2000). The first model ( Fig. 1a & 1b) is the unstructured density-independent growth model n(t+ 1) = n(t)exp(r (t )) (the model used for theoretical analyses by Dennis et al. [1991] and Fieberg & Ellner [2000]), with r(t) a Gaussian(μ,σ2) random variable with μ = −0.044, σ = 0.3. The value of σ is a rounded average over species of the values in Brook et al.'s Table 2, and the absolute value of μ (0.044) equals the average of |μ| over species. The second model ( Fig. 1c & 1d) is an age-structured Leslie matrix model with logistic density dependence in neonate survival. The mean and coefficient of variation of each vital rate for this model were derived by taking the average of the corresponding values for each species in Brook et al.'s Table 3, and rounding slightly. Our model species had juvenile and adult stages, with first breeding at 2 years of age and a maximum age of 15 years. The mean (coefficient of variation) of vital rates were as follows: adult annual fecundity, 0.6 (40%), juvenile survival, 0.6 (15%), adult survival, 0.75 (20%). Gaussian distributions truncated at 0 were used for random variations in vital rates. We assumed that neonate survival (or, equivalently, adult fecundity) was a function of adult density, with the value given above holding at 500 adults and decreasing linearly to zero at 1000 adults—hence, a maximum adult fecundity of 1.2/year at low adult densities. We assumed that this form of density dependence was known a priori, but the mean and variance of the survival rates and maximum fecundity had to be estimated from data. For each model we simulated a PVA as follows. We assumed n = 24 years of data and generated 12 simulated years of data to simulate the data-collection process (i.e., 12 simulated r (t) values for the unstructured model, 12 values each of adult survival, juvenile survival, and maximum adult fecundity for the age-structured model ). We parameterized the models by computing the sample mean and standard deviation of the simulated data for each vital rate. We then simulated the parameterized model 25,000 times to determine a series of quasi-extinction thresholds yielding extinction risks of 5%, 10%, 20%, and so forth, over a 12-year time period. We performed one model run with the true parameter values for each species, and we recorded the total number of species crossing below each threshold. Figures 1a and 1c are analogous to Fig. 1 of Brook et al., showing good agreement between actual and PVA–estimated total number of extinctions over ensembles of 21 test cases. Figures 1b and 1d compare the actual and estimated risks for each species individually. The spread in individual risk estimates is wide, so these estimates would not be reliable for assessing or comparing individual species. Fig. 1d illustrates that using a more realistic (and therefore more complicated) model only aggravates the problem, even though the amount of data was increased in exact proportion to the larger number of parameters in the more complex model and the density dependence was assumed to be known a priori. These results show that the ensemble-level tests of Brook et al. (2000) were inadequate to assess the precision of PVA risk estimates. The results of Fieberg and Ellner (2000) suggest that PVA will not be precise unless the sample size greatly exceeds the prediction interval. There is an additional reason for caution in applying the results of Brook et al. (2000): their conclusions were drawn from failure to reject null hypotheses. In such a case, proper inference requires that the size of the Type II error be examined by a power analysis ( Peterman 1990; Thompson et al. 2000), but Brook et al. do not provide such an analysis. The bottom line of their Table 1, where errors of a factor of two too high or too low are not statistically significant, suggests that extremely large differences would be required to reject the null hypothesis for the 5% extinction risk typical of published PVAs. This lack of power is not due to poor choice of methods but is an unavoidable consequence of the scarcity of long-term data sets. How useful is PVA, in view of its limitations? Thompson et al. (2000) provide an example to be emulated. They use PVA and power analysis to explore the consequences of some management strategies. They base their analysis on a range of assumptions about the rate of population decline, rather than relying on a single estimated rate. They use power analysis to determine the length of data series required to detect the decline. An important feature of this analysis and others presented in the same special section of Conservation Biology is careful accounting for uncertainty and its consequences for management. Similar uses of PVA for comparative purposes take advantage of its ability to summarize diverse data and explore the consequences of alternative actions (Groom & Pascual 1998; Burgman & Possingham 2000). For example, Lindenmayer and Possingham (1996) used PVA to compare timber-management options for conservation of Leadbetter's possum (Gymnobelideus leadbeateri ) in southeastern Australia and found that the ranking among the options was robust to parameter and model uncertainties. This is quite different from attempting to make quantitative predictions of extinction risk based on small data sets. It is limited, however, to within-species comparisons of relative risk under different management scenarios. When a comparison between species is attempted—for example, to assay which has the greatest need for immediate intervention—the uncertainties in the absolute risk estimates for each species are likely to be too high for such comparisons to be meaningful. Brook et al. (2000) tested predictions over an average time interval of about 13 years, so their results are relevant to 10- and 20-year time frames used for World Conservation Union listing of critically endangered and endangered species. Such short-term predictions can be important for formulating a management framework. But published PVAs generally have used much longer time intervals, 50–200 years, with 100 years the most common ( Beissinger & Westphal 1998; Fieberg & Ellner 2001). As our results indicate, risk estimates for longer time intervals are increasingly imprecise, with most estimates near zero or 1 because the predicted long-term risk is extremely sensitive to the estimated mean growth rate ( Dennis et al. 1991; Ludwig 1999; Fieberg & Ellner 2000). Coulson et al. (2001) raise additional concerns about the conclusions of Brook et al. (2000). They caution that data for most threatened or endangered species will be sparse and of lower quality than the data sets analyzed by Brook et al. Furthermore, they argue that predictions are likely to be accurate only if future mean and variation of vital rates or population growth will be similar to the data used to parameterize the model. Populations that are subject to rare high-recruitment events or catastrophic mortalities will therefore provide additional difficulties with regard to model parameterization and reliability. In summary, the results of Brook et al. (2000) are not sufficient to justify PVA as an accurate tool for categorizing individual species, even for short (10- to 20-year) time intervals. Their results provide evidence (subject to concerns about power) that risk and growth rate estimates are unbiased, which implies that PVA could be useful in predicting the total loss rate for a large group of species. For assessment of individual species, it is essential to account for imprecision in parameter estimates and its consequences for risk assessment. A variety of tools are available. We have already mentioned confidence intervals on the risk of extinction within a given time horizon. An analogous tool is prediction intervals for the time to extinction ( Engen & Sæther 2000), but methods to compute these are available only for very simple models. Alternatively, extinction probabilities can be calculated for the range of plausible parameter values by Bayesian methods ( Ludwig 1996). Similarly, using frequentist methodology, one can calculate the level of confidence that the true probability of extinction is less than any value (essentially a p value associated with the true probability of extinction). One may then display the range of likely extinction probabilities or weight them by a measure of their plausibility in light of the data. A weighting procedure has the merit of producing a single measure of risk, but this measure is sensitive to various assumptions made in the assessment process. Perhaps a better strategy would be to produce a prediction interval for the population size over the entire time horizon of interest, taking into account uncertainty in parameter estimates (as described by Sæther et al. 2000. This eliminates the subjective choices of a specific time horizon and quasi-extinction threshold for computing an extinction risk. Population viability analysis may then be one useful tool among a variety of decision-making aids, which might include historical and predicted future habitat losses, recent population trends, and genetic considerations. As is often the case with important environmental problems, even the best available science may be unable to provide the level of predictability and accuracy we might wish. We thank M. Mangel for his stimulating input in the preparation of this paper, and H. Possingham and an anonymous referee for comments on the manuscript.

BACKGROUND: The risk of transfusion-transmitted human immunodeficiency virus-1 (HIV-1), hepatitis C virus (HCV), and hepatitis B virus (HBV) infections is predominantly attributable to donations given during the early stage of infection when diagnostic tests may fail. In 1997, nucleic acid amplification technique (NAT)-testing was introduced at the German Red Cross (GRC) blood donor services to reduce this diagnostic window period (WP). STUDY DESIGN AND METHODS: A total of 31,524,571 blood donations collected from 1997 through 2005 were screened by minipool NAT, predominantly with pool sizes of 96 donations. These donations cover approximately 80 percent of all the blood collected in Germany during that period. Based on these data, the WP risk in the GRC blood donor population was estimated by using a state-of-the-art mathematic model. RESULTS: During the observation period, 23 HCV, 7 HIV-1, and 43 HBV NAT-only-positive donations were detected. On the basis of these data and estimated pre-NAT infectious WPs, the residual risk per unit transfused was estimated at 1 in 10.88 million for HCV (95% confidence interval [CI], 7.51-19.72 million), 1 in 4.30 million for HIV-1 (95% CI, 2.39-21.37 million), and 1 in 360,000 for HBV (95% CI, 0.19-3.36 million). Based on observed cases of breakthrough infections, the risk of transfusion-related infections may be even lower. CONCLUSION: The risk of a blood recipient becoming infected with HCV, HIV-1, or HBV has reached an extremely low level. Introduction of individual donation testing for HCV and HIV-1 would have a marginal effect on interception of WP donations.

How can we efficiently propagate uncertainty in a latent state representation with recurrent neural networks? This paper introduces stochastic recurrent neural networks which glue a deterministic recurrent neural network and a state space model together to form a stochastic and sequential neural generative model. The clear separation of deterministic and stochastic layers allows a structured variational inference network to track the factorization of the model's posterior distribution. By retaining both the nonlinear recursive structure of a recurrent neural network and averaging over the uncertainty in a latent path, like a state space model, we improve the state of the art results on the Blizzard and TIMIT speech modeling data sets by a large margin, while achieving comparable performances to competing methods on polyphonic music modeling.

Abstract Since Hoggart and Mendoza's article on ‘African immigrant workers in Spanish agriculture’ in Sociologia Ruralis in 1999 there has been a proliferation of interest in labour migration to/in rural Europe. It is now clear that the rural realm has been, and is being, transformed by immigration, and that low‐wage migrant workers in the food production industry are playing a particularly prominent role in this transformation. This article takes stock of the literature and identifies seven key issues associated with low‐wage labour migration, contemporary food production, and rural change. Most notably, since the 1990s, there has been growing demand for migrants in the segmented, and sometimes exploitative, labour markets of the European food production industries. This demand has been met across a variety of contexts, with states and labour market intermediaries playing a largely supportive role. However, migrants’ integration into rural communities has often been problematic, with the emphasis being on the need for, rather than needs of, low‐wage migrant workers.

Abstract This paper presents an observational and experimental study of the basking behaviour and heat exchange rate of the montane lizard Lacerta monticola. The results obtained by these procedures were coupled in order to understand behavioural mechanims promoting effective thermoregulation at high altitudes. Heating rate was higher when body size was smaller, and substrate temperature and sun rays incidence angle were higher. The lizards cooled faster when body size and substrate temperature were lower, and when the body temperature of the lizard going into shadow was higher. Time exposed to sun and mean duration of basking periods were longer early in the morning, while bask frequency increased through the morning. Our results suggest that time devoted to basking is mainly obtained by regulating bask duration. Lizards obtained the necessary time for heating by means of long basking periods. Mean travel distance per minute and distance to the nearest refuge increased from early morning to midday. These behavioural variables were tightly correlated with the expected heating rate of individuals. Body size affects thermoregulatory behaviour as well as locomotor activity. Juvenile lizards, with small body mass and high surface‐to‐volume ratios, were subjected to faster heating and cooling rates, basked more frequently than adults (but during shorter periods), and devoted more time to locomotion than adults. The thermoregulatory behaviour of L. monticola is the result of the combination of shuttling heliothermy by basking and the exploitation of thermal opportunities offered by patches in shade through thermal exchange with the substrate.

Abstract A recently presented model of tongue projection dynamics is used to generate a series of predictions concerning morphologies to be expected under selection for increased distance of projection, increased speed of projection, and increased directional versatility. A general understanding of biomechanical events and the model are used as points of departure for making specific predictions concerning details of structure in skeletal, muscular and connective tissue components of the tongue and associated structures. Comparative methods are used to examine these predictions in the genera of plethodontid salamanders. These salamanders are known to project their tongues to different degrees, and this knowledge is used to test the hypotheses concerning morphological specialization. Three distinct groups of plethodontid salamanders have evolved specializations for long distance projection, and these genera differ from one another in important ways in respect to specific character complexes. For example, the tropical genera and Hydromantes use CBII as the major force transmission element in the skeleton, while Eurycea and its allies use CBI in this role. Hydromantes differs from both in having a uniquely proportioned and structured hyobranchial skeleton and associated musculature. Less extreme specializations for tongue projection are found in different combinations in three other groups. Finally, two distinct groups of generalized species having only limited tongue projection capabilities are recognized, each having a unique complex of inter‐related features. Each of these eight groups is recognized and characterized as a functional mode, and hypotheses concerning the biomechanical meaning of the character complexes of each are formulated.