Website measures of engagement captured from millions of users, such as in-page scrolling and viewport position, can provide deeper understanding of attention than possible with simpler measures, such as dwell time. Using data from 1.2M news reading sessions, we examine and evaluate three increasingly sophisticated models of sub-document attention computed from viewport time, the time a page component is visible on the user display. Our modeling incorporates prior eye-tracking knowledge about onscreen reading, and we validate it by showing how, when used to estimate user reading rate, it aligns with known empirical measures. We then show how our models reveal an interaction between article topic and attention to page elements. Our approach supports refined large-scale measurement of user engagement at a level previously available only from lab-based eye-tracking studies.

These days, anyone with a laptop, Python, and a free afternoon can build an intelligent machine. In fact, if you have a bunch of data, go ahead and throw it into a deep neural net: your worries will be over! Machine learning will save us all! Unfortunately, those of us who spend our days fighting with these systems know that this is far from the truth. As anything created by humans, these algorithms have limitations, and there is no algorithmic substitute for careful thought. The gory details of techniques like Support Vector Machines, Neural Networks, and K-means clustering can be found in innumerable blog posts, texts, and online courses; we won’t discuss the details here. Instead, this talk will focus on the conceptual and mathematical framework under which all machine learning algorithms fall. We’ll briefly discuss the workings of a few of the most popular modern techniques (#DeepLearning, anyone?) and how they fit into this framework. More importantly, we’ll discuss some often-overlooked considerations that go into the design of these systems: human biases, transparency, and maintainability. Ultimately, we hope to provide a solid understanding of what these algorithms can do, what they can’t do, and when they should or should not be used.

The discovery of shared behavioral processes across phyla is a significant step in the establishment of a comparative study of behavior. We use immobility as an origin and reference for the measurement of fly locomotor behavior; speed, walking direction and trunk orientation as the degrees of freedom shaping this behavior; and cocaine as the parameter inducing progressive transitions in and out of immobility. We characterize and quantify the generative rules that shape Drosophila locomotor behavior, bringing about a gradual buildup of kinematic degrees of freedom during the transition from immobility to normal behavior, and the opposite narrowing down into immobility. Transitions into immobility unfold via sequential enhancement and then elimination of translation, curvature and finally rotation. Transitions out of immobility unfold by progressive addition of these degrees of freedom in the opposite order. The same generative rules have been found in vertebrate locomotor behavior in several contexts (pharmacological manipulations, ontogeny, social interactions) involving transitions in-and-out of immobility. Recent claims for deep homology between arthropod central complex and vertebrate basal ganglia provide an opportunity to examine whether the rules we report also share common descent. Our approach prompts the discovery of behavioral homologies, contributing to the elusive problem of behavioral evolution.

Unattended and continuously running environmental noise monitoring systems can capture an intractable amount of data. The signals captured can include a multitude of sources (e.g., wind noise and anthropogenic noise sources) in addition to the environmental noise sources of interest (e.g., aircraft, vehicles, trains, and military weapons). In this presentation, we explore the use of machine-learning methods to effectively isolate and identify environmental noise sources captured on such a noise monitoring system. Specifically, we consider the use of both unsupervised (e.g., principle components analysis, clustering methods, and deep belief networks) and supervised (e.g., logistic regression, support vector machines, and neural networks) pattern-learning methods to derive the features of interest and classify the signals based on the obtained features. The generalization performance of each method is assessed using a dataset of over 120,000 human classified signals, and the strengths and weaknesses of each approach are discussed.

The discovery of shared behavioral processes across phyla is an essential step in the establishment of a comparative study of behavior. We use immobility as an origin and reference for the measurement of locomotor behavior; speed, direction of walking and direction of facing as the three degrees of freedom shaping fly locomotor behavior; and cocaine as the parameter inducing a progressive transition in and out of immobility. In this way we expose and quantify the generative rules that shape part of fruit fly locomotor behavior, bringing about a gradual buildup of freedom during the transition from immobility to normal behavior and a precisely opposite narrowing down during the transition into immobility. During buildup the fly exhibits enhancement and then reduction to normal values of movement along each degree of freedom: first, body rotation in the horizontal plane, then path curvature and then speed of translation. Transition into immobility unfolds by narrowing down of the repertoire in the opposite sequential order, showing reciprocal relations during both buildup and narrowing down. The same generative rules apply to vertebrate locomotor behavior in a variety of contexts involving transition out and into immobility. Recent claims for deep homology between the arthropod central complex and the vertebrate basal ganglia provide an opportunity to examine whether the generative rules we discovered also share common descent. Neurochemical processes mediating the buildup of locomotor behavior in vertebrates could guide the search for equivalent processes in arthropods. The measurement methodology we use prompts the discovery of candidate behavioral homologies. Significance Statement Do flies and mice share the same behavior? By defining immobility as an intrinsic reference point for locomotor behavior we show that the rules that generate the transition from immobility to full blown normal behavior, and from full blown behavior to immobility are shared by fruit flies and mice. These rules constitute a much desired aim of evolutionary biology: the discovery of behavioral homologies across distant phyla. The methodology we use facilitates the discovery of cross-phyletic behavioral homologies, shedding light on the problem of the evolution of behavior.

We use immobility as an origin and reference for the measurement of locomotor behavior; speed, the direction of walking and the direction of facing as the three degrees of freedom shaping fly locomotor behavior, and cocaine as the parameter inducing a progressive transition in and out of immobility. In this way we expose and quantify the generative rules that shape fruit fly locomotor behavior, which consist of a gradual narrowing down of the fly's locomotor freedom of movement during the transition into immobility and a precisely opposite expansion of freedom during the transition from immobility to normal behavior. The same generative rules of narrowing down and expansion apply to vertebrate behavior in a variety of contexts, Recent claims for deep homology between the vertebrate basal ganglia and the arthropod central complex, and neurochemical processes explaining the expansion of locomotor behavior in vertebrates could guide the search for equivalent neurochemical processes that mediate locomotor narrowing down and expansion in arthropods. We argue that a methodology for isolating relevant measures and quantifying generative rules having a potential for discovering candidate behavioral homologies is already available and we specify some of its essential features.

In previous studies it has been shown that in vertebrates the ground plan of locomotor behavior consists of a transition from extensive mobility to immobility through an increase and then a gradual decrease in forward translation, accompanied by a concurrent gradual increase in whole body rotation around the hindquarters (shutdown of behavior). The opposite transition from immobility to extensive mobility involves an increase and then decrease in rotation around the hindquarters, accompanied by a concurrent gradual increase in forward translation (warm-up) [1,2,3].

Gene interactions are acknowledged to be a likely source of missing heritability in large‐scale genetic studies of complex neurological phenotypes. However, involvement of rare variants, de novo mutations, genetic lesions that are not easily detected with commonly used methods and epigenetic factors also are possible explanations. We used a laboratory evolution study to investigate the modulatory effects of background genetic variation on the phenotypic effect size of a null mutation with known impact on olfactory learning. To accomplish this, we first established a population that contained variation at just 23 loci and used selection to evolve suppression of the learning defect seen with null mutations in the rutabaga adenylyl cyclase. We thus biased the system to favor relatively simplified outcomes by choosing a Mendelian trait and by restricting the genetic variation segregating in the population. This experimental design also assures that the causal effects are among the known 23 segregating loci. We observe a robust response to selection that requires the presence of the 23 variants. Analyses of the underlying genotypes showed that interactions between more than two loci are likely to be involved in explaining the selection response, with implications for the missing heritability problem.

Evaluating the effectiveness of outdoor sound propagation assessment schemes using experimentally collected data is a complex task. Specifically, when the acoustic data collected assume non-Gaussian distributions with differing sample sizes per assigned class, traditional parametric statistical techniques may not be used. As an alternative, this research introduces an original cost function to evaluate and compare various methods aimed at defining acoustic propagation classes based on meteorological measurements. Class assignments in each scheme are based on the atmospheric stability, the strength of the vertical effective sound speed gradient, and the vertical effective sound speed profile. The acoustic data included in this analysis were generated by a high-energy impulsive source and gathered at source-to-receiver distances of 1, 2, 4, and 8 km. The results indicate that the assessment scheme based on the strength of the effective sound speed gradient most effectively classifies the peak level sound propagation sampled in temperate and desert climate conditions at these distances.

In this study we characterize the coordination between the direction a fruit-fly walks and the direction it faces, as well as offer a methodology for isolating and validating key variables with which we phenotype fly locomotor behavior. Our fundamental finding is that the angular interval between the direction a fly walks and the direction it faces is actively managed in intact animals and modulated in a patterned way with drugs. This interval is small in intact flies, larger with alcohol and much larger with cocaine. The dynamics of this interval generates six coordinative modes that flow smoothly into each other. Under alcohol and much more so under cocaine, straight path modes dwindle and modes involving rotation proliferate. To obtain these results we perform high content analysis of video-tracked open field locomotor behavior. Presently there is a gap between the quality of descriptions of insect behaviors that unfold in circumscribed situations, and descriptions that unfold in extended time and space. While the first describe the coordination between low-level kinematic variables, the second quantify cumulative measures and subjectively defined behavior patterns. Here we reduce this gap by phenotyping extended locomotor behavior in terms of the coordination between low-level kinematic variables, which we quantify, combining into a single field two disparate fields, that of high content phenotyping and that of locomotor coordination. This will allow the study of the genes/brain/locomotor coordination interface in genetically engineered and pharmacologically manipulated animal models of human diseases.

Statistical evidence for various models relating day-night sound level (DNL) to community noise annoyance is assessed with the Akaike information criterion. In particular, community-specific adjustments such as the community tolerance level (CTL, the DNL at which 50% of survey respondents are highly annoyed) and community tolerance spread (CTS, the difference between the DNL at which 90% and 10% are highly annoyed) are considered. The results strongly support models characterizing annoyance on a community-by-community basis, rather than with complete pooling and analysis of all available surveys. The most likely model was found to be a 2-parameter logistic model, with CTL and CTS fit independently to survey data from each community.

Outdoor sound propagation is highly dependent upon meteorological conditions. While this, of course, is a trivial statement, predicting sound levels based on meteorology is not. This is especially true for signals that propagate many kilometers, as is the case for those generated by high-energy impulsive sources such as explosions and heavy weaponry; waves have ample opportunity for refraction by and scattering from local atmospheric features along the entire propagation path. The range of received blast levels at distances greater than 2 km can span nearly 50 dB, depending on weather conditions. Using a statistical learning method known as a Random Forest, we demonstrate the prediction of levels from simple meteorological measurements in the face of this extreme variability. With simple, spatially sparse meteorological data, the model can predict levels to within 3 dB at 2 km and 5 dB at 15 km. The results presented here suggest that as more data are acquired through continuous noise monitoring programs, physics-blind, data-driven statistical models have the potential to supplant computationally intensive propagation models for noise prediction. Caveats and cautions when using these types of machine learning methods will also be discussed.

Blast noise from military installations often has a negative impact on the quality of life of residents living in nearby communities. This, in turn, negatively impacts the military's testing & training capabilities due to restrictions, curfews, or range closures enacted to address noise complaints. In order to more directly manage noise around military installations, accurate noise monitoring around bases has become a necessity. Although most noise monitors are simple sound level meters, more recent ones are capable of discerning blasts from ambient noise with some success. Investigators at the University of Pittsburgh (Pitt) developed a more advanced noise classifier that can discern between wind, aircraft, and blast noise, while simultaneously lowering the measurement threshold. Here, more recent work between Pitt and the US Army Engineer Research and Development Center will be presented from the development of a more advanced classifier that identifies additional classes of noise such as machine gun fire, vehicles, and electronic noise. Additional signal metrics were explored given the increased complexity of the classifier. By broadening the types of noise the system can accurately classify and increasing the number of metrics, a new system was developed with increased blast noise accuracy, decreased number of missed events, and significantly fewer false positives.

Temperature and wind effects on outdoor sound propagation have been well-studied in numerous theoretical investigations, leading to a number of commonly held beliefs about how weather affects outdoor propagation. For example, louder sound levels are expected when downwind propagation and temperature inversions are present, whereas lower levels are expected when upwind propagation and temperature lapses exist. However, the validity of such relationships has not been rigorously tested in the field for various terrains at long distances from an outdoor source. One is justified in questioning the validity of these relationships due to the difficulty of adequately modeling the dynamic atmosphere along the entire propagation path. This study examines some commonly held notions of outdoor sound propagation by experimentally investigating the effects of temperature gradients and wind speed/direction on sound propagation at long distances from a typical impulsive source. Temperature and wind conditions measured near the source and along the line of sound propagation are correlated with the received sound pressure levels recorded at distances up to 15 km from the source. The variability of peak sound pressure levels occurring under similar wind and temperature conditions is assessed, demonstrating that much complexity underlies these common aphorisms.

Variability in received sound levels were investigated at distances ranging from 4 m to 16 km from a typical blast source in two locations with different climates and terrain. Four experiments were conducted, two in a temperate climate with a hilly terrain and two in a desert climate with a flat terrain, under a variety of meteorological conditions. Sound levels were recorded in three different directions around the source during the summer and winter seasons in each location. Testing occurred over the course of several days for each experiment during all 24 h of the day, and meteorological data were gathered throughout each experiment. The peak levels (L(Pk)), C-weighted sound exposure levels (CSEL), and spectral characteristics of the received sound pressure levels were analyzed. The results show high variability in L(Pk) and CSEL at distances beyond 2 km from the source for each experiment, which was not clearly explained by the time of day the blasts occurred. Also, as expected, higher frequency energy is attenuated more drastically than the lower frequency energy as the distance from the source increases. These data serve as a reference for long-distance blast sound propagation.

In studying the gene/brain/behavior relationship, detailed quantitative descriptions of the behavior of the animal's behavior is required. Especially important are descriptions of behavior in terms of the dynamics of the coordination between the direction of progression of the animal and the animal’s body orientation, for these provide fundamental insight into motor control. For mouse locomotor behavior these measures have been welldeveloped, yet such descriptions are still sparse for Drosophila, the model system of choice for behavior geneticists. Here we present a dynamic description of the coordination between these two variables, as they change throughout different temporal segments of the motion. Methodology/Principal Findings Using intrinsic statistical and geometrical properties of a fly’s movement, we have uncovered six fundamental modes of motion related to translation and rotation of a fly’s body. Assignment of the frame-by-frame instantaneous movement into one of the six modes, followed by clustering and algorithmic classification, allows us to quantify the behavior in terms of the proportions and dynamic sequencing of each mode. The analysis uncovered that for normal flies, the angle between the direction the animal walks and the direction it faces is quite small. Under the influence of alcohol and cocaine, however, the angular interval between these two variables increases. Conclusions/Significance For several decades the representation of movement, indispensable for studying the interface between genes brain and behavior, suffered from the use of ad hoc building blocks such as “behavior patterns” or “response categories” assumed to be performed unblock by the whole organism, one at a time. The ethological and psychological schools using such representations, were based on expert (indeed, subjective) decisions. In the vast majority these alleged building blocks could not be shown to have a physiological reality in the brain; the school using them has been described as the “school of immaculate perception”. These ad hoc units were used as “black boxes” and their variable kinematic content has been disregarded. Disregarding the coordination between translation and rotation, for example, sidestepped the problem of coordination, which is at the heart of the brain/behavior interface. The huge progress in tracking and storing technology allowed the recording of continuous kinematic variables such as the location and orientation of the organism, thus bringing about a most significant change, allowing, for the first time, the study of coordination between these variables in long stretches of behavior lasting hours or even days. This time scale is necessary for the study of e.g., drug effects on behavior, or for the phenotyping of genetic mutants exploratory and locomotor behavior. The pioneers utilizing this technology use, however, the

A common set of signal features measurable by a basic sound level meter are analyzed, and the quality of information carried in subsets of these features are examined for their ability to discriminate military blast and non-blast sounds. The analysis is based on over 120 000 human classified signals compiled from seven different datasets. The study implements linear and Gaussian radial basis function (RBF) support vector machines (SVM) to classify blast sounds. Using the orthogonal centroid dimension reduction technique, intuition is developed about the distribution of blast and non-blast feature vectors in high dimensional space. Recursive feature elimination (SVM-RFE) is then used to eliminate features containing redundant information and rank features according to their ability to separate blasts from non-blasts. Finally, the accuracy of the linear and RBF SVM classifiers is listed for each of the experiments in the dataset, and the weights are given for the linear SVM classifier.

It is common for residents living on and around military installations to be exposed to a significant amount of high amplitude impulsive noise, primarily from large weaponry and other blast noise producing sources. Yet in comparison to transportation noise, there have been relatively few studies of how communities and individuals respond to this type of noise. This presentation will report the latest findings from recent human response to blast noise studies conducted at three military installations. Across all sites, blast noise has been found to be the most annoying noise source, despite the fact that a large percentage of respondents reported that their neighborhood was a good or excellent place to live. It has also been found that each community and individual has a unique tolerance to blast noise. Furthermore, individuals use a different and finite portion of the response scale, suggesting that the current methodology of fixating on the percent of the population that is highly annoyed may inadvertently be discarding useful response information. Comparisons between respondents living on- and off-post within and between study sites will be made, with special emphasis placed on differences between the community tolerance level and the community tolerance spread for each site.