skip to main content


Title: BayGEN: A Bayesian Space‐Time Stochastic Weather Generator
Abstract

We present a Bayesian hierarchical space‐time stochastic weather generator (BayGEN) to generate daily precipitation and minimum and maximum temperatures. BayGEN employs a hierarchical framework with data, process, and parameter layers. In the data layer, precipitation occurrence at each site is modeled using probit regression using a spatially distributed latent Gaussian process; precipitation amounts are modeled as gamma random variables; and minimum and maximum temperatures are modeled as realizations from Gaussian processes. The latent Gaussian process that drives the precipitation occurrence process is modeled in the process layer. In the parameter layer, the model parameters of the data and process layers are modeled as spatially distributed Gaussian processes, consequently enabling the simulation of daily weather at arbitrary (unobserved) locations or on a regular grid. All model parameters are endowed with weakly informative prior distributions. The No‐U Turn sampler, an adaptive form of Hamiltonian Monte Carlo, is used to maximize the model likelihood function and obtain posterior samples of each parameter. Posterior samples of the model parameters propagate uncertainty to the weather simulations, an important feature that makes BayGEN unique compared to traditional weather generators. We demonstrate the utility of BayGEN with application to daily weather generation in a basin of the Argentine Pampas. Furthermore, we evaluate the implications of crop yield by driving a crop simulation model with weather simulations from BayGEN and an equivalent non‐Bayesian weather generator.

 
more » « less
Award ID(s):
1811294
NSF-PAR ID:
10453571
Author(s) / Creator(s):
 ;  ;  ;  ;  
Publisher / Repository:
DOI PREFIX: 10.1029
Date Published:
Journal Name:
Water Resources Research
Volume:
55
Issue:
4
ISSN:
0043-1397
Page Range / eLocation ID:
p. 2900-2915
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. This paper studies the fundamental problem of learning multi-layer generator models. The multi-layer generator model builds multiple layers of latent variables as a prior model on top of the generator, which benefits learning complex data distribution and hierarchical representations. However, such a prior model usually focuses on modeling inter-layer relations between latent variables by assuming non-informative (conditional) Gaussian distributions, which can be limited in model expressivity. To tackle this issue and learn more expressive prior models, we propose an energy-based model (EBM) on the joint latent space over all layers of latent variables with the multi-layer generator as its backbone. Such joint latent space EBM prior model captures the intra-layer contextual relations at each layer through layer-wise energy terms, and latent variables across different layers are jointly corrected. We develop a joint training scheme via maximum likelihood estimation (MLE), which involves Markov Chain Monte Carlo (MCMC) sampling for both prior and posterior distributions of the latent variables from different layers. To ensure efficient inference and learning, we further propose a variational training scheme where an inference model is used to amortize the costly posterior MCMC sampling. Our experiments demonstrate that the learned model can be expressive in generating high-quality images and capturing hierarchical features for better outlier detection. 
    more » « less
  2. Excessive phosphorus (P) applications to croplands can contribute to eutrophication of surface waters through surface runoff and subsurface (leaching) losses. We analyzed leaching losses of total dissolved P (TDP) from no-till corn, hybrid poplar (Populus nigra X P. maximowiczii), switchgrass (Panicum virgatum), miscanthus (Miscanthus giganteus), native grasses, and restored prairie, all planted in 2008 on former cropland in Michigan, USA. All crops except corn (13 kg P ha−1 year−1) were grown without P fertilization. Biomass was harvested at the end of each growing season except for poplar. Soil water at 1.2 m depth was sampled weekly to biweekly for TDP determination during March–November 2009–2016 using tension lysimeters. Soil test P (0–25 cm depth) was measured every autumn. Soil water TDP concentrations were usually below levels where eutrophication of surface waters is frequently observed (> 0.02 mg L−1) but often higher than in deep groundwater or nearby streams and lakes. Rates of P leaching, estimated from measured concentrations and modeled drainage, did not differ statistically among cropping systems across years; 7-year cropping system means ranged from 0.035 to 0.072 kg P ha−1 year−1 with large interannual variation. Leached P was positively related to STP, which decreased over the 7 years in all systems. These results indicate that both P-fertilized and unfertilized cropping systems may leach legacy P from past cropland management. Experimental details The Biofuel Cropping System Experiment (BCSE) is located at the W.K. Kellogg Biological Station (KBS) (42.3956° N, 85.3749° W; elevation 288 m asl) in southwestern Michigan, USA. This site is a part of the Great Lakes Bioenergy Research Center (www.glbrc.org) and is a Long-term Ecological Research site (www.lter.kbs.msu.edu). Soils are mesic Typic Hapludalfs developed on glacial outwash54 with high sand content (76% in the upper 150 cm) intermixed with silt-rich loess in the upper 50 cm55. The water table lies approximately 12–14 m below the surface. The climate is humid temperate with a mean annual air temperature of 9.1 °C and annual precipitation of 1005 mm, 511 mm of which falls between May and September (1981–2010)56,57. The BCSE was established as a randomized complete block design in 2008 on preexisting farmland. Prior to BCSE establishment, the field was used for grain crop and alfalfa (Medicago sativa L.) production for several decades. Between 2003 and 2007, the field received a total of ~ 300 kg P ha−1 as manure, and the southern half, which contains one of four replicate plots, received an additional 206 kg P ha−1 as inorganic fertilizer. The experimental design consists of five randomized blocks each containing one replicate plot (28 by 40 m) of 10 cropping systems (treatments) (Supplementary Fig. S1; also see Sanford et al.58). Block 5 is not included in the present study. Details on experimental design and site history are provided in Robertson and Hamilton57 and Gelfand et al.59. Leaching of P is analyzed in six of the cropping systems: (i) continuous no-till corn, (ii) switchgrass, (iii) miscanthus, (iv) a mixture of five species of native grasses, (v) a restored native prairie containing 18 plant species (Supplementary Table S1), and (vi) hybrid poplar. Agronomic management Phenological cameras and field observations indicated that the perennial herbaceous crops emerged each year between mid-April and mid-May. Corn was planted each year in early May. Herbaceous crops were harvested at the end of each growing season with the timing depending on weather: between October and November for corn and between November and December for herbaceous perennial crops. Corn stover was harvested shortly after corn grain, leaving approximately 10 cm height of stubble above the ground. The poplar was harvested only once, as the culmination of a 6-year rotation, in the winter of 2013–2014. Leaf emergence and senescence based on daily phenological images indicated the beginning and end of the poplar growing season, respectively, in each year. Application of inorganic fertilizers to the different crops followed a management approach typical for the region (Table 1). Corn was fertilized with 13 kg P ha−1 year−1 as starter fertilizer (N-P-K of 19-17-0) at the time of planting and an additional 33 kg P ha−1 year−1 was added as superphosphate in spring 2015. Corn also received N fertilizer around the time of planting and in mid-June at typical rates for the region (Table 1). No P fertilizer was applied to the perennial grassland or poplar systems (Table 1). All perennial grasses (except restored prairie) were provided 56 kg N ha−1 year−1 of N fertilizer in early summer between 2010 and 2016; an additional 77 kg N ha−1 was applied to miscanthus in 2009. Poplar was fertilized once with 157 kg N ha−1 in 2010 after the canopy had closed. Sampling of subsurface soil water and soil for P determination Subsurface soil water samples were collected beneath the root zone (1.2 m depth) using samplers installed at approximately 20 cm into the unconsolidated sand of 2Bt2 and 2E/Bt horizons (soils at the site are described in Crum and Collins54). Soil water was collected from two kinds of samplers: Prenart samplers constructed of Teflon and silica (http://www.prenart.dk/soil-water-samplers/) in replicate blocks 1 and 2 and Eijkelkamp ceramic samplers (http://www.eijkelkamp.com) in blocks 3 and 4 (Supplementary Fig. S1). The samplers were installed in 2008 at an angle using a hydraulic corer, with the sampling tubes buried underground within the plots and the sampler located about 9 m from the plot edge. There were no consistent differences in TDP concentrations between the two sampler types. Beginning in the 2009 growing season, subsurface soil water was sampled at weekly to biweekly intervals during non-frozen periods (April–November) by applying 50 kPa of vacuum to each sampler for 24 h, during which the extracted water was collected in glass bottles. Samples were filtered using different filter types (all 0.45 µm pore size) depending on the volume of leachate collected: 33-mm dia. cellulose acetate membrane filters when volumes were less than 50 mL; and 47-mm dia. Supor 450 polyethersulfone membrane filters for larger volumes. Total dissolved phosphorus (TDP) in water samples was analyzed by persulfate digestion of filtered samples to convert all phosphorus forms to soluble reactive phosphorus, followed by colorimetric analysis by long-pathlength spectrophotometry (UV-1800 Shimadzu, Japan) using the molybdate blue method60, for which the method detection limit was ~ 0.005 mg P L−1. Between 2009 and 2016, soil samples (0–25 cm depth) were collected each autumn from all plots for determination of soil test P (STP) by the Bray-1 method61, using as an extractant a dilute hydrochloric acid and ammonium fluoride solution, as is recommended for neutral to slightly acidic soils. The measured STP concentration in mg P kg−1 was converted to kg P ha−1 based on soil sampling depth and soil bulk density (mean, 1.5 g cm−3). Sampling of water samples from lakes, streams and wells for P determination In addition to chemistry of soil and subsurface soil water in the BCSE, waters from lakes, streams, and residential water supply wells were also sampled during 2009–2016 for TDP analysis using Supor 450 membrane filters and the same analytical method as for soil water. These water bodies are within 15 km of the study site, within a landscape mosaic of row crops, grasslands, deciduous forest, and wetlands, with some residential development (Supplementary Fig. S2, Supplementary Table S2). Details of land use and cover change in the vicinity of KBS are given in Hamilton et al.48, and patterns in nutrient concentrations in local surface waters are further discussed in Hamilton62. Leaching estimates, modeled drainage, and data analysis Leaching was estimated at daily time steps and summarized as total leaching on a crop-year basis, defined from the date of planting or leaf emergence in a given year to the day prior to planting or emergence in the following year. TDP concentrations (mg L−1) of subsurface soil water were linearly interpolated between sampling dates during non-freezing periods (April–November) and over non-sampling periods (December–March) based on the preceding November and subsequent April samples. Daily rates of TDP leaching (kg ha−1) were calculated by multiplying concentration (mg L−1) by drainage rates (m3 ha−1 day−1) modeled by the Systems Approach for Land Use Sustainability (SALUS) model, a crop growth model that is well calibrated for KBS soil and environmental conditions. SALUS simulates yield and environmental outcomes in response to weather, soil, management (planting dates, plant population, irrigation, N fertilizer application, and tillage), and genetics63. The SALUS water balance sub-model simulates surface runoff, saturated and unsaturated water flow, drainage, root water uptake, and evapotranspiration during growing and non-growing seasons63. The SALUS model has been used in studies of evapotranspiration48,51,64 and nutrient leaching20,65,66,67 from KBS soils, and its predictions of growing-season evapotranspiration are consistent with independent measurements based on growing-season soil water drawdown53 and evapotranspiration measured by eddy covariance68. Phosphorus leaching was assumed insignificant on days when SALUS predicted no drainage. Volume-weighted mean TDP concentrations in leachate for each crop-year and for the entire 7-year study period were calculated as the total dissolved P leaching flux (kg ha−1) divided by the total drainage (m3 ha−1). One-way ANOVA with time (crop-year) as the fixed factor was conducted to compare total annual drainage rates, P leaching rates, volume-weighted mean TDP concentrations, and maximum aboveground biomass among the cropping systems over all seven crop-years as well as with TDP concentrations from local lakes, streams, and groundwater wells. When a significant (α = 0.05) difference was detected among the groups, we used the Tukey honest significant difference (HSD) post-hoc test to make pairwise comparisons among the groups. In the case of maximum aboveground biomass, we used the Tukey–Kramer method to make pairwise comparisons among the groups because the absence of poplar data after the 2013 harvest resulted in unequal sample sizes. We also used the Tukey–Kramer method to compare the frequency distributions of TDP concentrations in all of the soil leachate samples with concentrations in lakes, streams, and groundwater wells, since each sample category had very different numbers of measurements. Individual spreadsheets in “data table_leaching_dissolved organic carbon and nitrogen.xls” 1.    annual precip_drainage 2.    biomass_corn, perennial grasses 3.    biomass_poplar 4.    annual N leaching _vol-wtd conc 5.    Summary_N leached 6.    annual DOC leachin_vol-wtd conc 7.    growing season length 8.    correlation_nh4 VS no3 9.    correlations_don VS no3_doc VS don Each spreadsheet is described below along with an explanation of variates. Note that ‘nan’ indicate data are missing or not available. First row indicates header; second row indicates units 1. Spreadsheet: annual precip_drainage Description: Precipitation measured from nearby Kellogg Biological Station (KBS) Long Term Ecological Research (LTER) Weather station, over 2009-2016 study period. Data shown in Figure 1; original data source for precipitation (https://lter.kbs.msu.edu/datatables/7). Drainage estimated from SALUS crop model. Note that drainage is percolation out of the root zone (0-125 cm). Annual precipitation and drainage values shown here are calculated for growing and non-growing crop periods. Variate    Description year    year of the observation crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” precip_G    precipitation during growing period (milliMeter) precip_NG    precipitation during non-growing period (milliMeter) drainage_G    drainage during growing period (milliMeter) drainage_NG    drainage during non-growing period (milliMeter)      2. Spreadsheet: biomass_corn, perennial grasses Description: Maximum aboveground biomass measurements from corn, switchgrass, miscanthus, native grass and restored prairie plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2009-2015. Data shown in Figure 2.   Variate    Description year    year of the observation date    day of the observation (mm/dd/yyyy) crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” replicate    each crop has four replicated plots, R1, R2, R3 and R4 station    stations (S1, S2 and S3) of samplings within the plot. For more details, refer to link (https://data.sustainability.glbrc.org/protocols/156) species    plant species that are rooted within the quadrat during the time of maximum biomass harvest. See protocol for more information, refer to link (http://lter.kbs.msu.edu/datatables/36) For maize biomass, grain and whole biomass reported in the paper (weed biomass or surface litter are excluded). Surface litter biomass not included in any crops; weed biomass not included in switchgrass and miscanthus, but included in grass mixture and prairie. fraction    Fraction of biomass biomass_plot    biomass per plot on dry-weight basis (Grams_Per_SquareMeter) biomass_ha    biomass (megaGrams_Per_Hectare) by multiplying column biomass per plot with 0.01 3. Spreadsheet: biomass_poplar Description: Maximum aboveground biomass measurements from poplar plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2009-2015. Data shown in Figure 2. Note that poplar biomass was estimated from crop growth curves until the poplar was harvested in the winter of 2013-14. Variate    Description year    year of the observation method    methods of poplar biomass sampling date    day of the observation (mm/dd/yyyy) replicate    each crop has four replicated plots, R1, R2, R3 and R4 diameter_at_ground    poplar diameter (milliMeter) at the ground diameter_at_15cm    poplar diameter (milliMeter) at 15 cm height biomass_tree    biomass per plot (Grams_Per_Tree) biomass_ha    biomass (megaGrams_Per_Hectare) by multiplying biomass per tree with 0.01 4. Spreadsheet: annual N leaching_vol-wtd conc Description: Annual leaching rate (kiloGrams_N_Per_Hectare) and volume-weighted mean N concentrations (milliGrams_N_Per_Liter) of nitrate (no3) and dissolved organic nitrogen (don) in the leachate samples collected from corn, switchgrass, miscanthus, native grass, restored prairie and poplar plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2009-2016. Data for nitrogen leached and volume-wtd mean N concentration shown in Figure 3a and Figure 3b, respectively. Note that ammonium (nh4) concentration were much lower and often undetectable (<0.07 milliGrams_N_Per_Liter). Also note that in 2009 and 2010 crop-years, data from some replicates are missing.    Variate    Description crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” crop-year    year of the observation replicate    each crop has four replicated plots, R1, R2, R3 and R4 no3 leached    annual leaching rates of nitrate (kiloGrams_N_Per_Hectare) don leached    annual leaching rates of don (kiloGrams_N_Per_Hectare) vol-wtd no3 conc.    Volume-weighted mean no3 concentration (milliGrams_N_Per_Liter) vol-wtd don conc.    Volume-weighted mean don concentration (milliGrams_N_Per_Liter) 5. Spreadsheet: summary_N leached Description: Summary of total amount and forms of N leached (kiloGrams_N_Per_Hectare) and the percent of applied N lost to leaching over the seven years for corn, switchgrass, miscanthus, native grass, restored prairie and poplar plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2009-2016. Data for nitrogen amount leached shown in Figure 4a and percent of applied N lost shown in Figure 4b. Note the fraction of unleached N includes in harvest, accumulation in root biomass, soil organic matter or gaseous N emissions were not measured in the study. Variate    Description crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” no3 leached    annual leaching rates of nitrate (kiloGrams_N_Per_Hectare) don leached    annual leaching rates of don (kiloGrams_N_Per_Hectare) N unleached    N unleached (kiloGrams_N_Per_Hectare) in other sources are not studied % of N applied N lost to leaching    % of N applied N lost to leaching 6. Spreadsheet: annual DOC leachin_vol-wtd conc Description: Annual leaching rate (kiloGrams_Per_Hectare) and volume-weighted mean N concentrations (milliGrams_Per_Liter) of dissolved organic carbon (DOC) in the leachate samples collected from corn, switchgrass, miscanthus, native grass, restored prairie and poplar plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2009-2016. Data for DOC leached and volume-wtd mean DOC concentration shown in Figure 5a and Figure 5b, respectively. Note that in 2009 and 2010 crop-years, water samples were not available for DOC measurements.     Variate    Description crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” crop-year    year of the observation replicate    each crop has four replicated plots, R1, R2, R3 and R4 doc leached    annual leaching rates of nitrate (kiloGrams_Per_Hectare) vol-wtd doc conc.    volume-weighted mean doc concentration (milliGrams_Per_Liter) 7. Spreadsheet: growing season length Description: Growing season length (days) of corn, switchgrass, miscanthus, native grass, restored prairie and poplar plots in the Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2009-2015. Date shown in Figure S2. Note that growing season is from the date of planting or emergence to the date of harvest (or leaf senescence in case of poplar).   Variate    Description crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” year    year of the observation growing season length    growing season length (days) 8. Spreadsheet: correlation_nh4 VS no3 Description: Correlation of ammonium (nh4+) and nitrate (no3-) concentrations (milliGrams_N_Per_Liter) in the leachate samples from corn, switchgrass, miscanthus, native grass, restored prairie and poplar plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2013-2015. Data shown in Figure S3. Note that nh4+ concentration in the leachates was very low compared to no3- and don concentration and often undetectable in three crop-years (2013-2015) when measurements are available. Variate    Description crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” date    date of the observation (mm/dd/yyyy) replicate    each crop has four replicated plots, R1, R2, R3 and R4 nh4 conc    nh4 concentration (milliGrams_N_Per_Liter) no3 conc    no3 concentration (milliGrams_N_Per_Liter)   9. Spreadsheet: correlations_don VS no3_doc VS don Description: Correlations of don and nitrate concentrations (milliGrams_N_Per_Liter); and doc (milliGrams_Per_Liter) and don concentrations (milliGrams_N_Per_Liter) in the leachate samples of corn, switchgrass, miscanthus, native grass, restored prairie and poplar plots in Great Lakes Bioenergy Research Center (GLBRC) Biomass Cropping System Experiment (BCSE) during 2013-2015. Data of correlation of don and nitrate concentrations shown in Figure S4 a and doc and don concentrations shown in Figure S4 b. Variate    Description crop    “corn” “switchgrass” “miscanthus” “nativegrass” “restored prairie” “poplar” year    year of the observation don    don concentration (milliGrams_N_Per_Liter) no3     no3 concentration (milliGrams_N_Per_Liter) doc    doc concentration (milliGrams_Per_Liter) 
    more » « less
  3. Abstract

    We develop a space‐time Bayesian hierarchical modeling (BHM) framework for two flood risk attributes—seasonal daily maximum flow and the number of events that exceed a threshold during a season (NEETM)—at a suite of gauge locations on a river network. The model uses generalized extreme value (GEV) and Poisson distributions as marginals for these flood attributes with non‐stationary parameters. The rate parameters of the Poisson distribution and location, scale, and shape parameters of the GEV are modeled as linear functions of suitable covariates. Gaussian copulas are applied to capture the spatial dependence. The best covariates are selected using the Watanabe‐Akaike information criterion (WAIC). The modeling framework results in the posterior distribution of the flood attributes at all the gauges and various lead times. We demonstrate the utility of this modeling framework to forecast the flood risk attributes during the summer peak monsoon season (July‐August) at five gauges in the Narmada River basin (NRB) of West‐Central India for several lead times (0–3 months). As potential covariates, we consider climate indices such as El Niño–Southern Oscillation (ENSO), the Indian Ocean Dipole (IOD), and the Pacific Warm Pool Region (PWPR) from antecedent seasons, which have shown strong teleconnections with the Indian monsoon. We also include new indices related to the East Pacific and West Indian Ocean regions depending on the lead times. We show useful long lead skill from this modeling approach which has a strong potential to enable robust risk‐based flood mitigation and adaptation strategies 3 months before flood occurrences.

     
    more » « less
  4. Abstract

    Regional, automated meteorological networks, such as the Oklahoma Mesonet can potentially provide high quality forcing data for generating gridded surfaces, but proven methods of interpolating weather variables between the station locations are needed. We compared two interpolation methods, ordinary kriging (OK) and empirical Bayesian kriging (EBK), with and without using long‐term climate imprints (CI), for creating spatially continuous, daily weather datasets. Daily meteorological variables (maximum and minimum temperature, solar radiation, and precipitation) from the Oklahoma Mesonet for the period 1997–2014 were interpolated using geoprocessing tools in ArcGIS. Cross‐validation was used for evaluation of interpolation methods, with 90% of sites chosen randomly for the training set and the remaining 10% left for validation. For all interpolation approaches, cross‐validation showed coefficient of determination (R2) values of .99 and .98 for daily maximum and minimum air temperatures, with mean absolute error (MAE) ranging from ±0.45–0.50 °C for maximum temperature and ±0.77–0.80 °C for minimum temperature. Likewise, for daily solar radiation,R2values of .94 and .93 showed overall good prediction accuracy with MAE values 1.00 and 1.01 MJ m–2 d–1for EBK and OK, respectively. However, for rainfall, all methods yieldedR2values ≤.67, suggesting a need for more effective interpolation method. Based on its lower computational time and lower input data requirement, OK appears preferable to the other approaches tested here to provide the daily weather data for gridded models in Oklahoma and other regions with similar monitoring networks.

     
    more » « less
  5. Batch Normalization (BN) is essential to effectively train state-of-the-art deep Convolutional Neural Networks (CNN). It normalizes the layer outputs during training using the statistics of each mini-batch. BN accelerates training procedure by allowing to safely utilize large learning rates and alleviates the need for careful initialization of the parameters. In this work, we study BN from the viewpoint of Fisher kernels that arise from generative probability models. We show that assuming samples within a mini-batch are from the same probability density function, then BN is identical to the Fisher vector of a Gaussian distribution. That means batch normalizing transform can be explained in terms of kernels that naturally emerge from the probability density function that models the generative process of the underlying data distribution. Consequently, it promises higher discrimination power for the batch-normalized mini-batch. However, given the rectifying non-linearities employed in CNN architectures, distribution of the layer outputs show an asymmetric characteristic. Therefore, in order for BN to fully benefit from the aforementioned properties, we propose approximating underlying data distribution not with one, but a mixture of Gaussian densities. Deriving Fisher vector for a Gaussian Mixture Model (GMM), reveals that batch normalization can be improved by independently normalizing with respect to the statistics of disentangled sub-populations. We refer to our proposed soft piecewise version of batch normalization as Mixture Normalization (MN). Through extensive set of experiments on CIFAR-10 and CIFAR-100, using both a 5-layers deep CNN and modern Inception-V3 architecture, we show that mixture normalization reduces required number of gradient updates to reach the maximum test accuracy of the batch normalized model by ∼31%-47% across a variety of training scenarios. Replacing even a few BN modules with MN in the 48-layers deep Inception-V3 architecture is sufficient to not only obtain considerable training acceleration but also better final test accuracy. We show that similar observations are valid for 40 and 100-layers deep DenseNet architectures as well. We complement our study by evaluating the application of mixture normalization to the Generative Adversarial Networks (GANs), where "mode collapse" hinders the training process. We solely replace a few batch normalization layers in the generator with our proposed mixture normalization. Our experiments using Deep Convolutional GAN (DCGAN) on CIFAR-10 show that mixture normalized DCGAN not only provides an acceleration of ∼58% but also reaches lower (better) "Fréchet Inception Distance" (FID) of 33.35 compared to 37.56 of its batch normalized counterpart. 
    more » « less