In defense of the black box Holm, Elizabeth A
Science (American Association for the Advancement of Science),
04/2019, Letnik:
364, Številka:
6435
Journal Article
Recenzirano
Black box algorithms can be useful in science and engineering
The science fiction writer Douglas Adams imagined the greatest computer ever built, Deep Thought, programmed to answer the deepest ...question ever asked: the Great Question of Life, the Universe, and Everything. After 7.5 million years of processing, Deep Thought revealed its answer: Forty-two (
1
). As artificial intelligence (AI) systems enter every sector of human endeavor—including science, engineering, and health—humanity is confronted by the same conundrum that Adams encapsulated so succinctly: What good is knowing the answer when it is unclear why it is the answer? What good is a black box?
Overweight and obesity result from an imbalance between caloric intake and energy expenditure, including expenditure from spontaneous physical activity (SPA). Changes in SPA and resulting changes in ...non-exercise activity thermogenesis (NEAT) likely interact with diet to influence risk for obesity. However, previous research on the relationship between diet, physical activity, and energy expenditure has been mixed. The neuropeptide orexin is a driver of SPA, and orexin neuron activity can be manipulated using DREADDs (Designer Receptors Exclusively Activated by Designer Drugs). We hypothesized that HFD decreases SPA and NEAT, and that DREADD-mediated activation of orexin neuron signaling would abolish this decrease and produce an increase in NEAT instead. To test these ideas, we characterized behaviors to determine the extent to which access to a high-fat diet (HFD) influences the proportion and probability of engaging in food intake and activity. We then measured NEAT following access to HFD and following a DREADD intervention targeting orexin neurons. Two cohorts of orexin-cre male mice were injected with an excitatory DREADD virus into the caudal hypothalamus, where orexin neurons are concentrated. Mice were then housed in continuous metabolic phenotyping cages (Sable Promethion). Food intake, indirect calorimetry, and SPA were automatically measured every second. For cohort 1 (n=8), animals were given access to chow, then switched to HFD. For cohort 2 (n=4/group), half of the animals were given access to HFD, the other access to chow. Then, among animals on HFD, orexin neurons were activated following injections of clozapine n-oxide (CNO). Mice on HFD spent significantly less time eating (p<0.01) and more time inactive compared to mice on chow (p<0.01). Following a meal, mice on HFD were significantly more likely to engage in periods of inactivity compared to those on chow (p<0.05). NEAT was decreased in animals on HFD, and was increased to the NEAT level of control animals following activation of orexin neurons with DREADDs. Food intake (kilocalories) was not significantly different between mice on chow and HFD, yet mice on chow expended more energy per unit of SPA, relative to that in mice consuming HFD. These results suggest that HFD consumption reduces SPA and NEAT, and increases inactivity following a meal. Together, the data suggest a change in the efficiency of energy expenditure based upon diet, such that SPA during HFD burns fewer calories compared to SPA on a standard chow diet.
•A high fat diet decreases spontaneous physical activity and time spent eating.•A high fat diet encourages longer bouts of inactivity, particularly after a meal.•Energy expenditure from SPA (NEAT) is decreased in animals on a high fat diet.•Orexin activation increases previously low levels of NEAT induced by high fat diet.•A high fat diet increases energy efficiency during physical activity.
Display omitted
•A computer vision approach is used to compute microstructural fingerprints.•Using these fingerprints, microstructures can be classified with >80% accuracy.•Microstructural ...fingerprints form the basis for a visual image search engine.
The ‘bag of visual features’ image representation was applied to create generic microstructural signatures that can be used to automatically find relationships in large and diverse microstructural image data sets. Using this representation, a support vector machine (SVM) was trained to classify microstructures into one of seven groups with greater than 80% accuracy over 5-fold cross validation. In addition, the bag of visual features was implemented as the basis for a visual search engine that determines the best matches for a query image in a database of microstructures. These novel applications demonstrate the potential and the limitations of computer vision concepts in microstructural science.
The energies of a set of 388 distinct grain boundaries have been calculated based on embedded-atom method interatomic potentials for Ni and Al. The boundaries considered are a complete catalog of the ...coincident site lattice boundaries constructible in a computational cell of a prescribed size. Correlations of the boundary energy with other boundary properties (disorientation angle,
Σ value, excess boundary volume and proximity of boundary normals to 〈1
1
1〉) are examined. None of the usual geometric properties associated with grain boundary energy are useful predictors for this data set. The data set is incorporated as
supplementary material to facilitate the search for more complex correlations. The energies of corresponding boundaries in Ni and Al are found to differ by approximately a scaling factor related to the Voigt average shear modulus or C
44. Crystallographically close boundaries have similar energies; hence a table of grain boundary energies could be used for interpolation.
We apply a deep convolutional neural network segmentation model to enable novel automated microstructure segmentation applications for complex microstructures typically evaluated manually and ...subjectively. We explore two microstructure segmentation tasks in an openly available ultrahigh carbon steel microstructure dataset: segmenting cementite particles in the spheroidized matrix, and segmenting larger fields of view featuring grain boundary carbide, spheroidized particle matrix, particle-free grain boundary denuded zone, and Widmanstätten cementite. We also demonstrate how to combine these data-driven microstructure segmentation models to obtain empirical cementite particle size and denuded zone width distributions from more complex micrographs containing multiple microconstituents. The full annotated dataset is available on materialsdata.nist.gov.
The fields of machining learning and artificial intelligence are rapidly expanding, impacting nearly every technological aspect of society. Many thousands of published manuscripts report advances ...over the last 5 years or less. Yet materials and structures engineering practitioners are slow to engage with these advancements. Perhaps the recent advances that are driving other technical fields are not sufficiently distinguished from long-known informatics methods for materials, thereby masking their likely impact to the materials, processes, and structures engineering (MPSE). Alternatively, the diverse nature and limited availability of relevant materials data pose obstacles to machine-learning implementation. The glimpse captured in this overview is intended to draw focus to selected distinguishing advances, and to show that there are opportunities for these new technologies to have transformational impacts on MPSE. Further, there are opportunities for the MPSE fields to contribute understanding to the emerging machine-learning tools from a physics basis. We suggest that there is an immediate need to expand the use of these new tools throughout MPSE, and to begin the transformation of engineering education that is necessary for ongoing adoption of the methods.
The absolute grain boundary mobility of 388 nickel grain boundaries was calculated using a synthetic driving force molecular dynamics method; complete results appear in the Supplementary materials. ...Over 25% of the boundaries, including most of the non-Σ3 highest mobility boundaries, moved by a coupled shear mechanism. The range of non-shearing boundary mobilities is from 40 to 400
m/s
GPa, except for Σ3 incoherent twins which have mobilities of 200–2000
m/s
GPa. Some boundaries, including all the 〈1
1
1〉 twist boundaries, are immobile within the resolution of the simulation. Boundary mobility is not correlated with scalar parameters such as disorientation angle, Σ value, excess volume or boundary energy. Boundaries less than 15° from each other in five-dimensional crystallographic space tend to have similar mobilities. Some boundaries move via a non-activated motion mechanism, which greatly increases low-temperature mobility. Thermal roughening of grain boundaries is widely observed, with estimated roughening temperatures substantially among boundaries.
The thermodynamic equilibrium state of crystalline materials is a single crystal; however, polycrystalline grain growth almost always stops before this state is reached. Although typically attributed ...to solute drag, grain-growth stagnation occurs, even in high-purity materials. Recent studies indicate that grain boundaries undergo thermal roughening associated with an abrupt mobility change, so that at typical annealing temperatures, polycrystals will contain both smooth (slow) and rough (fast) boundaries. Mesoscale grain-growth models, validated by large-scale polycrystalline molecular dynamics simulations, show that even small fractions of smooth, slow boundaries can stop grain growth. We conclude that grain-boundary roughening provides an alternate stagnation mechanism that applies even to high-purity materials.
Microstructural characterization and analysis is the foundation of microstructural science, connecting materials structure to composition, process history, and properties. Microstructural ...quantification traditionally involves a human deciding what to measure and then devising a method for doing so. However, recent advances in computer vision (CV) and machine learning (ML) offer new approaches for extracting information from microstructural images. This overview surveys CV methods for numerically encoding the visual information contained in a microstructural image using either feature-based representations or convolutional neural network (CNN) layers, which then provides input to supervised or unsupervised ML algorithms that find associations and trends in the high-dimensional image representation. CV/ML systems for microstructural characterization and analysis span the taxonomy of image analysis tasks, including image classification, semantic segmentation, object detection, and instance segmentation. These tools enable new approaches to microstructural analysis, including the development of new, rich visual metrics and the discovery of processing-microstructure-property relationships.