HatchEnsemble: an efficient and practical uncertainty quantification method for deep neural networks

Quantifying predictive uncertainty in deep neural networks is a challenging and yet unsolved problem. Existing quantification approaches can be categorized into two lines. Bayesian methods provide a complete uncertainty quantification theory but are often not scalable to large-scale models. Along another line, non-Bayesian methods have good scalability and can quantify uncertainty with high quality. The most remarkable idea in this line is Deep Ensemble, but it is limited in practice due to its expensive computational cost. Thus, we propose HatchEnsemble to improve the efficiency and practicality of Deep Ensemble. The main idea is to use function-preserving transformations, ensuring HatchNets to inherit the knowledge learned by a single model called SeedNet. This process is called hatching, and HatchNet can be obtained by continuously widening the SeedNet. Based on our method, two different hatches are proposed, respectively, for ensembling the same and different architecture networks. To ensure the diversity of models, we also add random noises to parameters during hatching. Experiments on both clean and corrupted datasets show that HatchEnsemble can give a competitive prediction performance and better-calibrated uncertainty quantification in a shorter time compared with baselines.

Product Development Model for Tempe Chocolate Chips Based on Costumer Preferences in Banjarnegara, Central Java, Indonesia

This study aims to determine the attributes and categories that are important in product development and its model to obtain appropriate design recommendations based on consumer preferences using Quantification Theory Type 1 (QTT1). The number of respondents involved was 250 respondents in Banjarnegara Regency who were selected on purposive sampling. Respondents were asked to rate the level of preference for 14 samples of chocolate tempe chips. The results of the identification of attributes and categories that affect preferences obtained seven attributes and 19 categories, namely shape attributes (round, square, triangle, and oval), thickness (thin and thick), size (large and medium), distinctive aroma (tempe, chocolate, and fruits), flavors (chocolate and variants), surface layers (glossy, sprinkles, sprinkles & chocolate, visible tempe surface & sprinkles), as well as coatings (chocolate compound coating and chocolate coating & sprinkles). Based on the data analysis, some recommendations for product development of chocolate tempe chips are: round, 0.15 mm thick, 8 cm wide, flavored and chocolatey, glossy coating with chocolate coating and sprinkles. Keywords: costumer’s preferences, product development model, tempe chocolate chips

Job satisfaction of female radiological technologists: Factor analysis using Quantification Theory Type II

Background: Recently, there has been an increased demand for female radiological technologists (hereinafter referred to as female technologists) due to the spread of mammography, but there have been no previous studies on the job satisfaction of female technologists. In this study, we attempted to extract factors that affect the job satisfaction of female technologists to identify factors in the working environment that make it easy for female technologists to work. Methods: In this study, the results of a questionnaire survey of 287 female technologists in Hokkaido, conducted by the Hokkaido Radiological Technologist Organization “Furattokai”, were analyzed using Quantification Theory Type II. The factors considered were years of experience, number of duty shifts, number of technologists, male-female gender gap, number of job changes, sexual harassment, overtime, and paid holiday usage rate. Results: The results revealed that in addition to work environments where female technologists can work for a long time, the work content and relationships with male technologists, including sexual harassment, affected the job satisfaction of female technologists.Conclusions:This study extracted factors that affected the job satisfaction of female technologists as basic data for creating a working environment that makes it easier for female technologists to continue working for a long period of time, as the demand for female technologists is expected to increase in the future. The factors that strongly correlated with job satisfaction were not specific to women, as they are also common to men, including “years of experience” and “number of shifts on duty”.

Univariate and Multivariate Analyses of Influencing Factors on Methane Adsorption Capacity of Semianthracite

Methane adsorption isotherm experiments on semianthracite (2.00-2.33% R o , max ) collected from the Xin’an coal mine, Henan Province, China, were conducted to investigate the effects of pore structure, coal quality, coal maceral, and coal rank on methane adsorption capacity with applications of univariate and multivariate analyses. Methane adsorption capacity varies significantly from 12.03 to 28.40 cm3/g. In univariate analysis, methane adsorption capacity has a strong positive correlation with pore specific surface area, weak positive correlations with pore volume and ash content, and weak negative correlations with moisture content and inertinite content. No correlation is observed between methane adsorption capacity and coal rank. In multivariate analysis, the mathematical model of methane adsorption capacity affected by the combined individual variables is established based on quantification theory I. There are similarities and differences between the two analyses. The similarities are that pore specific surface area has the greatest contribution to methane adsorption capacity, while coal rank has the least contribution. The differences are reflected in two aspects. Firstly, the other influencing factors contribute differently to methane adsorption capacity. Secondly, the positive or negative correlations of some influencing factors present the opposite. The mathematic model synthetically covers the combined effects of the influencing factors, which is more representative in evaluating methane adsorption capacity.

Research on application method of uncertainty quantification technology in equipment test identification

This paper introduces the concepts of equipment test qualification and uncertainty quantification, and the analysis framework and process of equipment test uncertainty quantification. It analyzes the data uncertainty, model uncertainty and environmental uncertainty, and studies the corresponding uncertainty quantification theory to provide technical reference for the application of uncertainty quantification technology in the field of test identification.

Site Classification of Eucalyptus urophylla × Eucalyptus grandis Plantations in China

Background and Objectives: It is important to match species needs with site conditions for sustainable forestry. In Eucalyptus urophylla × Eucalyptus grandis plantations in southern Yunnan, China, species-site mismatches have led to inappropriate expansion and management, which has degraded forests and decreased efficiency in plantation production. Further research is needed to understand the relationship between tree growth and site productivity. We empirically explored site features and classified site types within these plantations in southern Yunnan. Our objective was to develop a theoretical basis for improving site selection for afforestation, and to establish intensive management in that region. Materials and Methods: 130 standard plots were set up in 1−15-year-old eucalyptus plantations in Pu’er and Lincang. We used quantification theory to examine the relationship between dominant tree growth traits and site factors. Hierarchical cluster analysis and canonical correlation analysis were applied to classify sites and evaluate the growth potential of E. urophylla × E. grandis plantations, respectively. Results: The multiple correlation coefficient between eight site factors (altitude, slope, slope position, aspect, soil depth, texture, bulk density, and litter thickness) and the quantitative growth of the dominant tree was 0.834 (p < 0.05). Slope position, altitude, and soil depth were the main factors contributing to the variation in stand growth. Plantation growth was best on lower slopes at relatively low altitude, where thick and weathered red soil layers existed. Conversely, the poorest plantations were located on upper slopes at higher altitude, with a thin semi-weathered purple soil layer. The soil factors total nitrogen (N) and potassium (K), trace boron (B), copper (Cu), and zinc (Zn) content, available phosphorous (P), and organic matter content in the soil influenced plantation growth. Conclusions. The addition of N, P, and K fertilizer as well as trace elements such as B, Cu, and Zn can promote the productivity of these plantations.

Epistemic Modals and Alternative Possibilities

Indicative judgments pertain to what is true. Epistemic modal judgments pertain to what must or might be true relative to a body of information. A standard view is that epistemic modals implicitly quantify over alternative possibilities, or ways things could turn out. On this view, a proposition must be true just in case it is true in all the possibilities consistent with the available information, and a proposition might be true just in case it is true in at least one possibility consistent with the available information. I report three experiments testing this view of epistemic modals. The results show that although modal judgments are sensitive to information about alternative possibilities, the standard quantification theory mischaracterizes the ordinary meaning of modals. I then report two more experiments testing the hypothesis that epistemic modals express willingness to attribute knowledge based on the available information. The results support this hypothesis. The results also show that the difference between “inside” and “outside” probabilistic information, familiar from the judgment and decision-making literature, affects epistemic modal judgments.

Verifiability of genus-level classification under quantification and parsimony theories: a case study of follicucullid radiolarians

The classical taxonomy of fossil invertebrates is based on subjective judgments of morphology, which can cause confusion, because there are no codified standards for the classification of genera. Here, we explore the validity of the genus taxonomy of 75 species and morphospecies of the Follicucullidae, a late Paleozoic family of radiolarians, using a new method, Hayashi's quantification theory II (HQT-II), a general multivariate statistical method for categorical datasets relevant to discriminant analysis. We identify a scheme of 10 genera rather than the currently accepted 3 genera (Follicucullus, Ishigaconus, and Parafollicucullus). As HQT-II cannot incorporate stratigraphic data, a phylogenetic tree of Follicucullidae was reconstructed for 38 species using maximum parsimony. Six lineages emerged, roughly in concordance with the results of HQT-II. Combined with parsimony ancestral state reconstruction, the ancestral group of this family is Haplodiacanthus. Five other groups were discriminated, the Parafollicucullus, Curvalbaillella, Pseudoalbaillella, Longtanella, and Follicucullus–Cariver lineages. The morphological evolution of these lineages comprises a minimum essential list of eight states of four traits. HQT-II is a novel discriminant analytical multivariate method that may be of value in other taxonomic problems of paleobiology.