Corun: Concurrent Inference and Continuous Training at the Edge for Cost-Efficient AI-Based Mobile Image Sensing.

Intelligent mobile image sensing powered by deep learning analyzes images captured by cameras from mobile devices, such as smartphones or smartwatches. It supports numerous mobile applications, such as image classification, face recognition, and camera scene detection. Unfortunately, mobile devices often lack the resources necessary for deep learning, leading to increased inference latency and rapid battery consumption. Moreover, the inference accuracy may decline over time due to potential data drift. To address these issues, we introduce a new cost-efficient framework, called Corun, designed to simultaneously handle multiple inference queries and continual model retraining/fine-tuning of a pre-trained model on a single commodity GPU in an edge server to significantly improve the inference throughput, upholding the inference accuracy. The scheduling method of Corun undertakes offline profiling to find the maximum number of concurrent inferences that can be executed along with a retraining job on a single GPU without incurring an out-of-memory error or significantly increasing the latency. Our evaluation verifies the cost-effectiveness of Corun. The inference throughput provided by Corun scales with the number of concurrent inference queries. However, the latency of inference queries and the length of a retraining epoch increase at substantially lower rates. By concurrently processing multiple inference and retraining tasks on one GPU instead of using a separate GPU for each task, Corun could reduce the number of GPUs and cost required to deploy mobile image sensing applications based on deep learning at the edge.

Disentangling the effects of similarity, familiarity, and liking on social inference strategies.

People constantly make inferences about others' beliefs and preferences. People can draw on various sources of information to make these inferences, including stereotypes, self-knowledge, and target-specific knowledge. What leads people to use each of these sources of information over others? The current study examined factors that influence the use of these sources of information, focusing on three interpersonal dimensions - the extent to which people feel (a) familiar with, (b) similar to, or (c) like the target. In four studies (total N = 1136), participants inferred the beliefs and preferences of others - celebrities (Studies 1a-1b), constructed fictional targets (Study 2), and actual acquaintances (Study 3). Participants also rated familiarity with, similarity to, and liking of the target. Analyses assessed the use of each source of information by comparing inferences with information provided by those sources. Familiarity was associated with greater use of target-specific knowledge, while similarity and liking were associated with self-knowledge. Low similarity and high liking were associated with increased use of stereotypes. We discuss the implication of these findings and their applicability to unique cases, including inferences about celebrities, public figures, and positively stereotyped groups, in which familiarity, similarity, and liking do not perfectly align.

Estimating the effectiveness of marked sidewalks: An application of the spatial causality approach.

Various safety enhancements and policies have been proposed to enhance pedestrian safety and minimize vehicle-pedestrian accidents. A relatively recent approach involves marked sidewalks delineated by painted pathways, particularly in Asia's crowded urban centers, offering a cost-effective and space-efficient alternative to traditional paved sidewalks. While this measure has garnered interest, few studies have rigorously evaluated its effectiveness. Current before-after studies often use correlation-based approaches like regression, lacking effective consideration of causal relationships and confounding variables. Moreover, spatial heterogeneity in crash data is frequently overlooked during causal inference analyses, potentially leading to inaccurate estimations. This study introduces a geographically weighted difference-in-difference (GWDID) method to address these gaps and estimate the safety impact of marked sidewalks. This approach considers spatial heterogeneity within the dataset in the spatial causal inference framework, providing a more nuanced understanding of the intervention's effects. The simplicity of the modeling process makes it applicable to various study designs relying solely on pre- and post-exposure outcome measurements. Conventional DIDs and Spatial Lag-DID models were used for comparison. The dataset we utilized included a total of 13,641 pedestrian crashes across Taipei City, Taiwan. Then the crash point data was transformed into continuous probability values to determine the crash risk on each road segment using network kernel density estimation (NKDE). The treatment group comprised 1,407 road segments with marked sidewalks, while the control group comprised 3,097 segments with similar road widths. The pre-development program period was in 2017, and the post-development period was in 2020. Results showed that the GWDID model outperformed the spatial lag DID and traditional DID models. As a local causality model, it illustrated spatial heterogeneity in installing marked sidewalks. The program significantly reduced pedestrian crash risk in 43% of the total road segments in the treatment group. The coefficient distribution map revealed a range from -22.327 to 2.600, with over 95% of the area yielding negative values, indicating reduced crash risk after installing marked sidewalks. Notably, the impact of crash risk reduction increased from rural to urban areas, emphasizing the importance of considering spatial heterogeneity in transportation safety policy assessments.

GeneWorker: An end-to-end robotic reinforcement learning approach with collaborative generator and worker networks.

Reinforcement learning aided by the skill conception exhibits potent capabilities in guiding autonomous agents toward acquiring meaningful behaviors. However, in the current landscape of reinforcement learning, a skill is often merely a rudimentary abstraction of a sequence of primitive actions, serving as a component of the input to policy networks with fixed network parameters. This rigid methodology presents obstacles when attempting to integrate with burgeoning techniques such as meta-learning and large language models. To address this issue, we introduce a unique neural skill representation that abstracts the activation of neurons in each neural layer. Based on this, a novel end-to-end robotic reinforcement learning algorithm is proposed, in which two sub-networks, i.e., generator and worker networks, implement collaborative inferences via neural skills. Specifically, the generator produces a series of multi-spatial neural skills, providing efficient guidance for subsequent decision-making; by integrating these skills, the worker can determine its own network weights and biases to cope with various environmental conditions. Therefore, actions can be sampled with flexibly changeable network parameters through the collaboration between generator and worker networks. The experiments demonstrate that GeneWorker can achieve a mean success rate of over 90.67% on continuous robotic tasks and outperforms previous state-of-the-art methods by a minimum of 54% on the pick-and-place task.

Machine Learning to Predict Pregnancy in Dairy Cows: An Approach Integrating Automated Activity Monitoring and On-Farm Data.

Automated activity monitoring (AAM) systems are critical in the dairy industry for detecting estrus and optimizing the timing of artificial insemination (AI), thus enhancing pregnancy success rates in cows. This study developed a predictive model to improve pregnancy success by integrating AAM data with cow-specific and environmental factors. Utilizing data from 1,054 cows, this study compared the pregnancy outcomes between two AI timings-8 or 10 h post-AAM alarm. Variables such as age, parity, body condition, locomotion, and vaginal discharge scores, peripartum diseases, the breeding program, the bull used for AI, milk production at the time of AI, and environmental conditions (season, relative humidity, and temperature-humidity index) were considered alongside the AAM data on rumination, activity, and estrus intensity. Six predictive models were assessed to determine their efficacy in predicting pregnancy success: logistic regression, Bagged AdaBoost algorithm, linear discriminant, random forest, support vector machine, and Bagged Classification Tree. Integrating the on-farm data with AAM significantly enhanced the pregnancy prediction accuracy at AI compared to using AAM data alone. The random forest models showed a superior performance, with the highest Kappa statistic and lowest false positive rates. The linear discriminant and logistic regression models demonstrated the best accuracy, minimal false negatives, and the highest area under the curve. These findings suggest that combining on-farm and AAM data can significantly improve reproductive management in the dairy industry.

Experimental evidence for a semantic typology of emoji: Inferences of co-, pro-, and post-text emoji.

Emoji symbols are widely used in online communication, particularly in instant messaging and on social media platforms. Existing research draws comparisons between the functions of emoji and those of gestures, with recent work extending a proposed typology of gestures to emoji, arguing that different emoji types can be distinguished by their placement within the modified text and by their semantic contribution (the linguistic inferences that they give rise to). In this paper, we present four experiments designed to test the predictions of this extended typology, the results of which suggest that emoji symbols indeed trigger the hypothesised linguistic inferences. The findings provide support for a semantic typology of emoji and contribute further evidence of the parallels between gesture and emoji.

Mitogenomic phylogeny, biogeography, and cryptic divergence of the genus Silurus (Siluriformes: Siluridae).

The genus Silurus, an important group of catfish, exhibits heterogeneous distribution in Eurasian freshwater systems. This group includes economically important and endangered species, thereby attracting considerable scientific interest. Despite this interest, the lack of a comprehensive phylogenetic framework impedes our understanding of the mechanisms underlying the extensive diversity found within this genus. Herein, we analyzed 89 newly sequenced and 20 previously published mitochondrial genomes (mitogenomes) from 13 morphological species to reconstruct the phylogenetic relationships, biogeographic history, and species diversity of Silurus. Our phylogenetic reconstructions identified eight clades, supported by both maximum-likelihood and Bayesian inference. Sequence-based species delimitation analyses yielded multiple molecular operational taxonomic units (MOTUs) in several taxa, including the Silurus asotus complex (four MOTUs) and Silurus microdorsalis (two MOTUs), suggesting that species diversity is underestimated in the genus. A reconstructed time-calibrated tree of Silurus species provided an age estimate of the most recent common ancestor of approximately 37.61 million years ago (Ma), with divergences among clades within the genus occurring between 11.56 Ma and 29.44 Ma, and divergences among MOTUs within species occurring between 3.71 Ma and 11.56 Ma. Biogeographic reconstructions suggested that the ancestral area for the genus likely encompassed China and the Korean Peninsula, with multiple inferred dispersal events to Europe and Central and Western Asia between 21.78 Ma and 26.67 Ma and to Japan between 2.51 Ma and 18.42 Ma. Key factors such as the Eocene-Oligocene extinction event, onset and intensification of the monsoon system, and glacial cycles associated with sea-level fluctuations have likely played significant roles in shaping the evolutionary history of the genus Silurus.

Illusory inferences in conditional expressions.

A robber points a gun at a cashier and says: "Only one of these two options is true: If you conceal the combination to the safe, then I kill you; otherwise, if you don´t conceal the combination to the safe, then I kill you." Hearing this statement, most people conclude that, in either case, "I kill you." This is an illusory response, in fact; the valid conclusion states "I don´t kill you." The research reported here studied the roles that different expressions of conditionals ("if-then," "only if," and "if and only if") play in the illusory response. Three experiments show that participants inferred the conclusion "I kill you" from the conditional "if-then" and "I may or may not kill you" from the conditional "only if," while selecting both options with similar frequency for the biconditional "if and only if." These results shed light on the main theories of deductive reasoning.

Heuristic strategy of intuitive statistical inferences in 7- to 10-year-old children.

Intuitive statistical inferences refer to making inferences about uncertain events based on limited probabilistic information, which is crucial for both human and non-human species' survival and reproduction. Previous research found that 7- and 8-year-old children failed in intuitive statistical inference tasks after heuristic strategies had been controlled. However, few studies systematically explored children's heuristic strategies of intuitive statistical inferences and their potential numerical underpinnings. In the current research, Experiment 1 (N = 81) examined 7- to 10-year-olds' use of different types of heuristic strategies; results revealed that children relied more on focusing on the absolute number strategy. Experiment 2 (N = 99) and Experiment 3 (N = 94) added continuous-format stimuli to examine whether 7- and 8-year-olds could make genuine intuitive statistical inferences instead of heuristics. Results revealed that both 7- and 8-year-olds and 9- and 10-year-olds performed better in intuitive statistical inference tasks with continuous-format stimuli, even after focusing on the absolute number strategy had been controlled. The results across the three experiments preliminarily hinted that the ratio processing system might rely on the approximate number system. Future research could clarify what specific numerical processing mechanism may be used and how it might support children's statistical intuitions.

Capturing the unpredictability of stem cells.

A new mathematical model that can be applied to both single-cell and bulk DNA sequencing data sheds light on the processes governing population dynamics in stem cells.

Measures of genetic diversification in somatic tissues at bulk and single-cell resolution.

Intra-tissue genetic heterogeneity is universal to both healthy and cancerous tissues. It emerges from the stochastic accumulation of somatic mutations throughout development and homeostasis. By combining population genetics theory and genomic information, genetic heterogeneity can be exploited to infer tissue organization and dynamics in vivo. However, many basic quantities, for example the dynamics of tissue-specific stem cells remain difficult to quantify precisely. Here, we show that single-cell and bulk sequencing data inform on different aspects of the underlying stochastic processes. Bulk-derived variant allele frequency spectra (VAF) show transitions from growing to constant stem cell populations with age in samples of healthy esophagus epithelium. Single-cell mutational burden distributions allow a sample size independent measure of mutation and proliferation rates. Mutation rates in adult hematopietic stem cells are higher compared to inferences during development, suggesting additional proliferation-independent effects. Furthermore, single-cell derived VAF spectra contain information on the number of tissue-specific stem cells. In hematopiesis, we find approximately 2 × 105 HSCs, if all stem cells divide symmetrically. However, the single-cell mutational burden distribution is over-dispersed compared to a model of Poisson distributed random mutations. A time-associated model of mutation accumulation with a constant rate alone cannot generate such a pattern. At least one additional source of stochasticity would be needed. Possible candidates for these processes may be occasional bursts of stem cell divisions, potentially in response to injury, or non-constant mutation rates either through environmental exposures or cell-intrinsic variation.

The difficulty of detecting inbreeding depression and its effect on conservation decisions.

Statistical inferences about inbreeding depression are often derived from analyses with low power and a high risk of failing to detect inbreeding depression. That risk is widely appreciated by scientists familiar with the relevant statistical and genetical theory, but may be overlooked and underappreciated by decision-makers. Consequently, there is value in demonstrating this risk using a real example. We use data from the wolf population on Isle Royale to demonstrate the difficulty of making reliable statistical inferences about inbreeding depression. This wolf population is known-by other methods-to have gone effectively extinct due to deleterious genetic processes associated with inbreeding. Beyond that demonstration, we use two case-studies-wolves on Isle Royale and vaquita (porpoises) from the Gulf of California, Mexico-to show how statistical inferences about inbreeding depression can affect conservation decisions. According to most decision theory, decisions depend importantly on: 1) probabilities that certain states exist (e.g. inbreeding depression is present) and 2) the utility assigned to various outcomes (e.g. the value of acting to mitigate inbreeding when it is present). The probabilities are provided by statistical inference; whereas utilities are almost entirely determined by normative values and judgements. Our analysis suggests that decisions to mitigate inbreeding depression are often driven more by utilities (normative values) than probabilities (statistical inferences). As such, advocates for mitigating inbreeding depression will benefit from better communicating to decision-makers the value of populations persisting and the extent to which decisions should depend on normative values.

Evaluating low-cost monitoring designs for PM2.5 exposure assessment with a spatiotemporal modeling approach.

Determining the most feasible and cost-effective approaches to improving PM2.5 exposure assessment with low-cost monitors (LCMs) can considerably enhance the quality of its epidemiological inferences. We investigated features of fixed-site LCM designs that most impact PM2.5 exposure estimates to be used in long-term epidemiological inference for the Adult Changes in Thought Air Pollution (ACT-AP) study. We used ACT-AP collected and calibrated LCM PM2.5 measurements at the two-week level from April 2017 to September 2020 (N of monitors [measurements] = 82 [502]). We also acquired reference-grade PM2.5 measurements from January 2010 to September 2020 (N = 78 [6186]). We used a spatiotemporal modeling approach to predict PM2.5 exposures with either all LCM measurements or varying subsets with reduced temporal or spatial coverage. We evaluated the models based on a combination of cross-validation and external validation at locations of LCMs included in the models (N = 82), and also based on an independent external validation with a set of LCMs not used for the modeling (N = 30). We found that the model's performance declined substantially when LCM measurements were entirely excluded (spatiotemporal validation R2 [RMSE] = 0.69 [1.2 µg/m3]) compared to the model with all LCM measurements (0.84 [0.9 µg/m3]). Temporally, using the farthest apart measurements (i.e., the first and last) from each LCM resulted in the closest model's performance (0.79 [1.0 µg/m3]) to the model with all LCM data. The models with only the first or last measurement had decreased performance (0.77 [1.1 µg/m3]). Spatially, the model's performance decreased linearly to 0.74 (1.1 µg/m3) when only 10% of LCMs were included. Our analysis also showed that LCMs located in densely populated, road-proximate areas improved the model more than those placed in moderately populated, road-distant areas.

Demographic responses of oceanic island birds to local and regional ecological disruptions revealed by whole-genome sequencing.

Disentangling the effects of ecological disruptions operating at different spatial and temporal scales in shaping past species' demography is particularly important in the current context of rapid environmental changes driven by both local and regional factors. We argue that volcanic oceanic islands provide useful settings to study the influence of past ecological disruptions operating at local and regional scales on population demographic histories. We investigate potential drivers of past population dynamics for three closely related species of passerine birds from two volcanic oceanic islands, Reunion and Mauritius (Mascarene archipelago), with distinct volcanic history. Using ABC and PSMC inferences from complete genomes, we reconstructed the demographic history of the Reunion Grey White-eye (Zosterops borbonicus (Pennant, 1781)), the Reunion Olive White-eye (Z. olivaceus (Linnaeus, 1766)) and the Mauritius Grey White-eye (Z. mauritianus (Gmelin, 1789)) and searched for possible causes underlying similarities or differences between species living on the same or different islands. Both demographic inferences strongly support ancient and long-term expansions in all species. They also reveal different trajectories between species inhabiting different islands, but consistent demographic trajectories in species or populations from the same island. Species from Reunion appear to have experienced synchronous reductions in population size during the Last Glacial Maximum, a trend not seen in the Mauritian species. Overall, this study suggests that local events may have played a role in shaping population trajectories of these island species. It also highlights the potential of our conceptual framework to disentangle the effects of local and regional drivers on past species' demography and long-term population processes.

Profiling the barriers to the spreading of news using news headlines.

News headlines can be a good data source for detecting the barriers to the spreading of news in news media, which can be useful in many real-world applications. In this study, we utilize semantic knowledge through the inference-based model COMET and the sentiments of news headlines for barrier classification. We consider five barriers, including cultural, economic, political, linguistic, and geographical and different types of news headlines, including health, sports, science, recreation, games, homes, society, shopping, computers, and business. To that end, we collect and label the news headlines automatically for the barriers using the metadata of news publishers. Then, we utilize the extracted common-sense inferences and sentiments as features to detect the barriers to the spreading of news. We compare our approach to the classical text classification methods, deep learning, and transformer-based methods. The results show that (1) the inference-based semantic knowledge provides distinguishable inferences across the 10 categories that can increase the effectiveness and enhance the speed of the classification model; (2) the news of positive sentiments cross the political barrier, whereas the news of negative sentiments cross the cultural, economic, linguistic, and geographical barriers; (3) the proposed approach using inferences-based semantic knowledge and sentiment improves performance compared with using only headlines in barrier classification. The average F1-score for 4 out of 5 barriers has significantly improved as follows: for cultural barriers from 0.41 to 0.47, for economic barriers from 0.39 to 0.55, for political barriers from 0.59 to 0.70 and for geographical barriers from 0.59 to 0.76.

Social perception inferences of computer-generated faces: an Asian Indian and United States cultural comparison.

Results from research with computer-generated faces have demonstrated that participants are able to make different trait inferences to different generated faces. However, only a few studies using computer-generated faces with cross-cultural samples have been done. This study compared the facial trait inference results from India and the United States, using three validated neutral expression computer-generated faces from the University of Chicago Perception and Judgment Lab database as facial stimuli. The three faces varied in perceived threat. Participants were asked about the attractiveness, pleasing-ness (to look at), honesty, and potential threat in each of the three faces. Results indicated that participants from both cultural samples made the same inferences to the three faces; participants rated the attractiveness, pleasing-ness, and honesty highest in the low threat face and lowest in the high threat face. Indian participants perceive the high threat face to be less threatening than the United States participants. Participants were also asked about the emotional expression on each of the faces, even though the faces were presumably neutral. United States participants were significantly more likely to indicate that the faces in all three threat conditions were emotionally neutral, compared to Indian participants, reflecting a cultural In-group bias, in which members of a culture are more accurately able to identify expressions on faces from their own culture.

Effects of marine sediment as agricultural substrate on soil microbial diversity: an amplicon sequencing study.

BACKGROUND: The soil microbiota has a direct impact on plant development and other metabolic systems, such as the degradation of organic matter and the availability of microelements and metabolites. In the context of agricultural soils, microbial activity is crucial for maintaining soil health and productivity. Thus, the present study aimed to identify, characterize, and quantify the microbial communities of four types of substrates with varying proportions of marine port sediment used for cultivating lemons. By investigating microbial diversity and relative abundance, the work aimed to highlight the importance of soil microbial communities in agriculture when alternative culture media was used. RESULTS: The composition and structure of the sampled microbial communities were assessed through the amplification and sequencing of the V3-V4 variable regions of the 16 S rRNA gene The results revealed a diverse microbial community composition in all substrate samples, with a total of 41 phyla, 113 classes, 266 orders, 405 families, 715 genera, and 1513 species identified. Among these, Proteobacteria, Bacteroidota, Planctomycetota, Patescibacteria, Chloroflexi, Actinobacteriota, Acidobacteriota, Verrucomicrobiota, and Gemmatimonadota accounted for over 90% of the bacterial reads, indicating their dominance in the substrates. CONCLUSIONS: The impact of the substrate origin on the diversity and relative abundace of the microbiota was confirmed. The higher content of beneficial bacterial communities for plant development identified in peat could explain why is considered an ideal agricultural substrate. Development of "beneficial for plants" bacterial communities in alternative agricultural substrates, regardless of the edaphic characteristics, opens the possibility of studying the forced and specific inoculation of these culture media aiming to be agriculturally ideals.

High Performance on a Pragmatic Task May Not Be the Result of Successful Reasoning: On the Importance of Eliciting Participants' Reasoning Strategies.

Formal probabilistic models, such as the Rational Speech Act model, are widely used for formalizing the reasoning involved in various pragmatic phenomena, and when a model achieves good fit to experimental data, that is interpreted as evidence that the model successfully captures some of the underlying processes. Yet how can we be sure that participants' performance on the task is the result of successful reasoning and not of some feature of experimental setup? In this study, we carefully manipulate the properties of the stimuli that have been used in several pragmatics studies and elicit participants' reasoning strategies. We show that certain biases in experimental design inflate participants' performance on the task. We then repeat the experiment with a new version of stimuli which is less susceptible to the identified biases, obtaining a somewhat smaller effect size and more reliable estimates of individual-level performance.

The organizational principles of impression formation.

Trait impressions about others are a fundamental tool to navigate the rich social environment and yet a unitary model of its organizational principles is still lacking. The statistical properties of impression formation observed in previous studies are akin to processes that govern information encoding and storage in memory, suggesting similar cognitive and computational mechanisms. Here, in 2,780 participants, impression formation has been formalized with a computational model representing three organizational principles of memory (temporal, semantic and valence-related). The model specifically captured two main patterns of impression formation: (1) a negative valence effect that makes negative impressions loom longer than positive ones; (2) an interaction effect between the temporal and valence content that endorses more negative impressions when negative information is met first. This work shows that mechanisms of information encoding, storage and retrieval interact in ways that explain biased impression formation about social partners, thereby providing quantitative evidence for those mechanisms in individuals' impressions of others' social qualities. We discuss the implications of these results for social impressions in different, real-world contexts, and suggest how the proposed model might be extended to capture other kinds of effects, from negativity bias and pessimism to social discrimination.

Estimating Systemic Cognitive States from a Mixture of Physiological and Brain Signals.

As human-machine teams are being considered for a variety of mixed-initiative tasks, detecting and being responsive to human cognitive states, in particular systematic cognitive states, is among the most critical capabilities for artificial systems to ensure smooth interactions with humans and high overall team performance. Various human physiological parameters, such as heart rate, respiration rate, blood pressure, and skin conductance, as well as brain activity inferred from functional near-infrared spectroscopy or electroencephalogram, have been linked to different systemic cognitive states, such as workload, distraction, or mind-wandering among others. Whether these multimodal signals are indeed sufficient to isolate such cognitive states across individuals performing tasks or whether additional contextual information (e.g., about the task state or the task environment) is required for making appropriate inferences remains an important open problem. In this paper, we introduce an experimental and machine learning framework for investigating these questions and focus specifically on using physiological and neurophysiological measurements to learn classifiers associated with systemic cognitive states like cognitive load, distraction, sense of urgency, mind wandering, and interference. Specifically, we describe a multitasking interactive experimental setting used to obtain a comprehensive multimodal data set which provided the foundation for a first evaluation of various standard state-of-the-art machine learning techniques with respect to their effectiveness in inferring systemic cognitive states. While the classification success of these standard methods based on just the physiological and neurophysiological signals across subjects was modest, which is to be expected given the complexity of the classification problem and the possibility that higher accuracy rates might not in general be achievable, the results nevertheless can serve as a baseline for evaluating future efforts to improve classification, especially methods that take contextual aspects such as task and environmental states into account.