Search NASA⌕ Search

SEARCH · Search NASA

Results for “Dimensionality Reduction”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,585 records · Page 88
All sourcesNASA NTRSDOE OSTI
NASA NTRS2006

Lightning Return-Stroke Current Waveforms Aloft, from Measured Field Change, Current, and Channel Geometry

Three-dimensional reconstructions of six rocket-triggered lightning channels are derived from stereo photographs. These reconstructed channels are used to infer the behavior of the current in return strokes above the ground from current waveforms measured at the channel base and electric-field-change waveforms measured at a range of 5.2 kilometers for 24 return strokes in these channels. Streak photographs of 14 of the same strokes are analyzed to determine the rise times, propagation speeds, and amplitudes of relative light intensity for comparison with the electrical inferences. Results include the following: 1) The fine structure of the field-change waveforms that were radiated by these subsequent return strokes can be explained, in large part, by channel geometry. 2) The average 10 - 90% rise time of the stroke current increased by about a factor of seven in our sample, from an observed 0.31 plus or minus 0.17 microseconds at the surface to an inferred 2.2 plus or minus 0.5 microcseconds at 1 kilometer path length above the surface. 3) The three-dimensional propagation speed of the current front averaged 1.80 plus or minus 0.24 X 10(exp 8) meters per second over channel lengths typically greater than 1 kilometer. 4) Assuming that the measured current was entirely due to the return stroke forced an unreasonably large and abrupt reduction in inferred current amplitude over the first few tens of meters above the surface, especially in cases when the leader was bright relative to its stroke. Therefore, a significant fraction of the current at the surface was probably due to the leader, at least in such cases. 5) Peak return-stroke currents decreased by approximately 37 plus or minus 12% from 100 meters to 1 kilometer of path length above the surface. Because of uncertainty about how to partition the measured current between leader and return stroke, we are unable to infer the variation of current amplitude near the ground.

Willett, J. C.↗
NASA NTRS2010

Characterizing Methods of Measuring Flow-Mediated Dilation in the Brachial Artery

Regulation of vascular tone is one of the many important functions of the vascular endothelium. Endothelial dysfunction is a critical early event in the pathogenesis of atherosclerosis and occurs in the absence of angiographic disease. Flow-Mediated Dilation (FMD) is a noninvasive technique commonly used to evaluate endothelium-dependent vasodilation in humans and gauge the health of the cardiovascular system. Reductions in brachial artery FMD have been strongly correlated with disease progression and are predictive of future cardiac events. The flow stimulus for brachial artery FMD occurs as a result of the increased shear stress following deflation of an occlusion cuff around the upper arm. Using 2-dimensional ultrasound, changes in arterial diameter up to 5-minutes following cuff deflation are calculated from baseline image measurements. Along with pulsed Doppler measures of flow velocity through the artery, flow-mediated, endothelium-dependent vasodilation can be assessed. There is debate among investigators, however, about the proper positioning of the occlusion cuff during FMD testing. It is thought that placement of the cuff around the upper arm may not accurately reflect the impact of nitric oxide, a critically important molecule released as a result of the increased shear stress created by the FMD technique. Data suggest that the production of other endogenous metabolites may also contribute to FMD-related changes when positioning the cuff around the upper arm. To overcome the potential influence of such molecules, researchers now suggest that the occlusion cuff be placed below the elbow allowing a more precise estimate of nitric oxide mediated dilation. The purpose of this study is to compare the differences in FMD between the two methodologies of occlusion cuff placement. In addition, this study will determine the method that is easier for ultrasound technicians to perform and will produce a low coefficient of variance between technicians. Ultimately, the results of this study will help in tracking any adverse cardiovascular effects of spaceflight.

Callender, Ariane R.↗
NASA NTRS2019

Does the Solar Wind Electric Field Control the Reconnection Rate at Earth's Subsolar Magnetopause?

We demonstrate that the reconnection rate at the subsolar magnetopause is stronglycontrolled by the solar wind electric field and depends weakly on the local properties of the dissipationregion. Our approach is to match the solar wind and magnetospheric states in an internal boundarylayer described by the Cassak and Shay (2007, https://doi.org/10.1063/1.2795630) expression fortwo-dimensional asymmetric reconnection. Faraday's law along the Sun-Earth line determinesthe variation of the solar wind electric field from the bow shock to the magnetopause. While themagnetospheric plasma exerts some control over the reconnection rate, magnetic flux pileup in the sheathpartially compensates for any local reduction in the reconnection rate. For a fixed magnetospheric state,the reconnection rate is shown to be directly proportional to the solar wind electric field, thus explainingwhy the solar wind electric field correlates well with geomagnetic indices.

Dorelli, John C.↗
DOE OSTI2014

Twisted 3D N=4 supersymmetric YM on deformed A{sub 3}{sup *} lattice

We study a class of twisted 3D N=4 supersymmetric Yang-Mills (SYM) theory on particular 3-dimensional lattice L{sub 3D} formally denoted as L{sub 3D}{sup su{sub 3}×u{sub 1}} and given by non-trivial fibration L{sub 1D}{sup u{sub 1}}×L{sub 2D}{sup su{sub 3}} with base L{sub 2D}{sup su{sub 3}}=A{sub 2}{sup *}, the weight lattice of SU(3). We first, develop the twisted 3D N=4 SYM in continuum by using superspace method where the scalar supercharge Q is manifestly exhibited. Then, we show how to engineer the 3D lattice L{sub 3D}{sup su{sub 3}×u{sub 1}} that host this theory. After that we build the lattice action S{sub latt} invariant under the following three points: (i) U(N) gauge invariance, (ii) BRST symmetry, (iii) the S{sub 3} point group symmetry of L{sub 3D}{sup su{sub 3}×u{sub 1}}. Other features such as reduction to twisted 2D supersymmetry with 8 supercharges living on L{sub 2D}≡L{sub 2D}{sup su{sub 2}×u{sub 1}}, the extension to twisted maximal 5D SYM with 16 supercharges on lattice L{sub 5D}≡L{sub 5D}{sup su{sub 4}×u{sub 1}} as well as the relation with known results are also given.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2015

Two-dimensional semiconductor HfSe{sub 2} and MoSe{sub 2}/HfSe{sub 2} van der Waals heterostructures by molecular beam epitaxy

Using molecular beam epitaxy, atomically thin 2D semiconductor HfSe{sub 2} and MoSe{sub 2}/HfSe{sub 2} van der Waals heterostructures are grown on AlN(0001)/Si(111) substrates. Details of the electronic band structure of HfSe{sub 2} are imaged by in-situ angle resolved photoelectron spectroscopy indicating a high quality epitaxial layer. High-resolution surface tunneling microscopy supported by first principles calculations provides evidence of an ordered Se adlayer, which may be responsible for a reduction of the measured workfunction of HfSe{sub 2} compared to theoretical predictions. The latter reduction minimizes the workfunction difference between the HfSe{sub 2} and MoSe{sub 2} layers resulting in a small valence band offset of only 0.13 eV at the MoSe{sub 2}/HfSe{sub 2} heterointerface and a weak type II band alignment.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2017

Two-Dimensional J eff = 1 / 2 Antiferromagnetic Insulator Unraveled from Interlayer Exchange Coupling in Artificial Perovskite Iridate Superlattices

We report an experimental investigation of the two-dimensional J eff =1/2 antiferromagnetic Mott insulator by varying the interlayer exchange coupling in [(SrIrO 3 ) 1 , (SrTiO 3 ) m ] (m=1, 2 and 3) superlattices. Although all samples exhibited an insulating ground state with long-range magnetic order, temperature-dependent resistivity measurements showed a stronger insulating behavior in the m = 2 and m = 3 samples than the m = 1 sample which displayed a clear kink at the magnetic transition. This difference indicates that the blocking effect of the excessive SrTiO 3 layer enhances the effective electron-electron correlation and strengthens the Mott phase. The significant reduction of the Néel temperature from 150 K for m = 1 to 40 K for m = 2 demonstrates that the long-range order stability in the former is boosted by a substantial interlayer exchange coupling. Resonant x-ray magnetic scattering revealed that the interlayer exchange coupling has a switchable sign, depending on the SrTiO 3 layer number m, for maintaining canting-induced weak ferromagnetism. In conclusion, the nearly unaltered transition temperature between the m = 2 and the m = 3 demonstrated that we have realized a two-dimensional antiferromagnet at finite temperatures with diminishing interlayer exchange coupling.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
DOE OSTI2017

Anisotropic features of two-dimensional hydrogen atom in magnetic field

The aim of this study is the numerical research of anisotropic characteristics of a two-dimensional (2D) hydrogen atom induced by a magnetic field. The ground state energy (GSE) of the 2D hydrogen atom and the corresponding wavefunction have been numerically calculated in the Born–Oppenheimer approximation and taking into account the finite proton mass. The nonlinear dependence of the GSE on angle α between the magnetic field vector and the normal to the electron motion plane has been found in a wide range of the magnetic field. The effect of a significant reduction of the GSE (up to 1.9-fold) is observed with increasing the angle α up to 90°. The agreement with experimental data has been demonstrated. The dependences of the GSE of a 2D exciton in GaAs/Al{sub 0.33}Ga{sub 0.67}As have been determined for various tilt angles and magnetic fields.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗
DOE OSTI2015

The effect of charged quantum dots on the mobility of a two-dimensional electron gas: How important is the Coulomb scattering?

We have investigated the influence of a layer of charged self-assembled quantum dots (QDs) on the mobility of a nearby two-dimensional electron gas (2DEG). Time-resolved transconductance spectroscopy was used to separate the two contributions of the change in mobility, which are: (i) The electrons in the QDs act as Coulomb scatterers for the electrons in the 2DEG. (ii) The screening ability and, hence, the mobility of the 2DEG decreases when the charge carrier density is reduced by the charged QDs, i.e., the mobility itself depends on the charge carrier concentration. Surprisingly, we find a negligible influence of the Coulomb scattering on the mobility for a 2DEG, separated by a 30 nm tunneling barrier to the layer of QDs. This means that the mobility change is completely caused by depletion, i.e., reduction of the charge carrier density in the 2DEG, which indirectly influences the mobility.

75 CONDENSED MATTER PHYSICS, SUPERCONDUCTIVITY AND↗
NASA NTRS1984

Loss reduction in axial-flow compressors through low-speed model testing

A systematic procedure for reducing losses in axial-flow compressors is presented. In this procedure, a large, low-speed, aerodynamic model of a high-speed core compressor is designed and fabricated based on aerodynamic similarity principles. This model is then tested at low speed where high-loss regions associated with three-dimensional endwall boundary layers flow separation, leakage, and secondary flows can be located, detailed measurements made, and loss mechanisms determined with much greater accuracy and much lower cost and risk than is possible in small, high-speed compressors. Design modifications are made by using custom-tailored airfoils and vector diagrams, airfoil endbends, and modified wall geometries in the high-loss regions. The design improvements resulting in reduced loss or increased stall margin are then scaled to high speed. This paper describes the procedure and presents experimental results to show that in some cases endwall loss has been reduced by as much as 10 percent, flow separation has been reduced or eliminated, and stall margin has been substantially improved by using these techniques.

Wisler, D. C.↗
DOE OSTI2015

A Multicomponent Blend as a Diesel Fuel Surrogate for Compression Ignition Engine Applications

A mixture of n-dodecane and m-xylene is investigated as a diesel fuel surrogate for compression ignition engine applications. Compared to neat n-dodecane, this binary mixture is more representative of diesel fuel because it contains an alkyl-benzene which represents an important chemical class present in diesel fuels. A detailed multi-component mechanism for n-dodecane and m-xylene was developed by combining a previously developed n-dodecane mechanism with a recently developed mechanism for xylenes. The xylene mechanism is shown to reproduce experimental ignition data from a rapid compression machine and shock tube, speciation data from the jet stirred reactor and flame speed data. This combined mechanism was validated by comparing predictions from the model with experimental data for ignition in shock tubes and for reactivity in a flow reactor. The combined mechanism, consisting of 2885 species and 11754 reactions, was reduced to a skeletal mechanism consisting 163 species and 887 reactions for 3D diesel engine simulations. The mechanism reduction was performed using directed relation graph (DRG) with expert knowledge (DRG-X) and DRG-aided sensitivity analysis (DRGASA) at a fixed fuel composition of 77% of n-dodecane and 23% m-xylene by volume. The sample space for the reduction covered pressure of 1 – 80 bar, equivalence ratio of 0.5 – 2.0, and initial temperature of 700 – 1600 K for ignition. The skeletal mechanism was compared with the detailed mechanism for ignition and flow reactor predictions. Finally, the skeletal mechanism was validated against a spray flame dataset under diesel engine conditions documented on the Engine Combustion Network (ECN) website. These multidimensional simulations were performed using a Representative Interactive Flame (RIF) turbulent combustion model. In conclusion, encouraging results were obtained compared to the experiments with regards to the predictions of ignition delay and lift-off length at different ambient temperatures.

30 DIRECT ENERGY CONVERSION↗
DOE OSTI2016

Sensitivity of Turbine-Height Wind Speeds to Parameters in Planetary Boundary-Layer and Surface-Layer Schemes in the Weather Research and Forecasting Model

We evaluate the sensitivity of simulated turbine-height winds to 26 parameters applied in a planetary boundary layer (PBL) scheme and a surface layer scheme of the Weather Research and Forecasting (WRF) model over an area of complex terrain during the Columbia Basin Wind Energy Study. An efficient sampling algorithm and a generalized linear model are used to explore the multiple-dimensional parameter space and quantify the parametric sensitivity of modeled turbine-height winds. The results indicate that most of the variability in the ensemble simulations is contributed by parameters related to the dissipation of the turbulence kinetic energy (TKE), Prandtl number, turbulence length scales, surface roughness, and the von Kármán constant. The relative contributions of individual parameters are found to be dependent on both the terrain slope and atmospheric stability. The parameter associated with the TKE dissipation rate is found to be the most important one, and a larger dissipation rate can produce larger hub-height winds. A larger Prandtl number results in weaker nighttime winds. Increasing surface roughness reduces the frequencies of both extremely weak and strong winds, implying a reduction in the variability of the wind speed. All of the above parameters can significantly affect the vertical profiles of wind speed, the altitude of the low-level jet and the magnitude of the wind shear strength. The wind direction is found to be modulated by the same subset of influential parameters. Remainder of abstract is in attachment.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2019

Atomic-Level Insight into the Postsynthesis Band Gap Engineering of a Lewis Base Polymer Using Lewis Acid Tris(pentafluorophenyl)borane

Here, we investigate the binding properties of the Lewis acid tris(pentafluorophenyl)borane with a Lewis base semiconducting polymer, PFPT, and the subsequent mechanism of band gap reduction. Experiments and quantum chemical calculations confirm that the formation of a Lewis acid adduct is energetically favorable (ΔG° < -0.2 eV), with preferential binding at the pyridyl nitrogen in the polymer backbone over other Lewis base sites. Upon adduct formation, ultraviolet photoelectron spectroscopy indicates only a slight decrease in the HOMO energy, implying that a larger reduction in the LUMO energy is primarily responsible for the observed optical band gap narrowing (ΔE opt = 0.3 eV). Herein, we also provide the first spatially resolved picture of how Lewis acid adducts form in heterogeneous, disordered polymer/tris(pentafluorophenyl)borane thin films via one- (1D) and two-dimensional (2D) solid-state nuclear magnetic resonance. Notably, solid-state 1D 11 B, 13 C{ 1 H}, and 13 C{ 19 F} cross-polarization magic-angle spinning (CP-MAS) NMR and 2D 1 H{ 19 F} and 1 H{ 1 H} correlation NMR analyses establish that BCF molecules are intercalated between branched C 16 H 33 side chains with the boron atom facing toward the pyridyl nitrogen atoms of PFPT.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2017

Graphene-Directed Formation of a Nitrogen-Doped Porous Carbon Sheet with High Catalytic Performance for the Oxygen Reduction Reaction

In this paper, a nitrogen (N)-doped porous carbon sheet is prepared by in situ polymerization of pyrrole on both sides of graphene oxide, following which the polypyrrole layers are then transformed to the N-doped porous carbon layers during the following carbonization, and a sandwich structure is formed. Such a sheet-like structure possesses a high specific surface area and, more importantly, guarantees the sufficient utilization of the N-doping active porous sites. The internal graphene layer acts as an excellent electron pathway, and meanwhile, the external thin and porous carbon layer helps to decrease the ion diffusion resistance during electrochemical reactions. As a result, this sandwich structure exhibits prominent catalytic activity toward the oxygen reduction reaction in alkaline media, as evidenced by a more positive onset potential, a larger diffusion-limited current, better durability and poison-tolerance than commercial Pt/C. Finally, this study shows a novel method of using graphene to template the traditional porous carbon into a two-dimensional, thin, and porous carbon sheet, which greatly increases the specific surface area and boosts the utilization of inner active sites with suppressed mass diffusion resistance.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2026

Non-equilibrium reducing flame aerosol process to create supported high-entropy alloy nanoparticles

High-entropy alloy (HEA) nanomaterials provide opportunities and property combinations for energy and electronic applications, but their practical synthesis faces challenges of elemental immiscibility, metal reducibility, and particle aggregation during their synthesis. Herein, we report a broadly applicable non-equilibrium, scalable, and in-situ reducing flame aerosol process for synthesis of supported HEA nanoparticles. This versatile process can directly load a high concentration of 2 ~ 4 nm HEA nanoparticles on various 1- to 3-dimensional supports. Notably, simultaneous formation of HEA nanoparticles and a mesoporous silica support was successfully realized in a single step. Exploration of this process demonstrates the role of kinetics and entropy on decreasing alloy particle size and altering the reducibility of elements. We propose an entropy-induced reduction mechanism to incorporate oxidizable elements into HEAs, which extends the compositional space of HEA nanoparticles. As a representative catalytic application, we present a RuPdOsIrPt/graphene electrocatalyst with high activity and stability for hydrogen oxidation reaction. Our findings open horizons for high-performance HEA design and applications in diverse fields such as catalysis, electrochemistry, and sensing.

organic↗
DOE OSTI2016

Nonlinear equation for Farley–Buneman waves in multispecies plasma

The nonlinear equation describing the Farley–Buneman (FB) waves in multispecies collisional plasmas is derived by employing the multiple-scale reduction analysis. It is shown that the presence of several ion species with different collisionalities and different ion masses removes the degeneracy of the nonlinear equation and generates the nonlinear terms resulting in wave steepening and wave breaking. This effect may be responsible for formation of one-dimensional coherent FB waves of a finite amplitude.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗
DOE OSTI2026

GP Cosmology Surrogate v1.0

GP Cosmology Surrogate is a Python library for building and training a generalized multi-output Gaussian process (GP) framework of @takhtaganov2021cosmic. In this approach, the surrogate is constructed sequentially, guided by a Bayesian optimization acquisition function that targets reduction of emulation error in the regions most consistent with the observational data. This adaptive design concentrates computational resources where they have the greatest impact on inference accuracy. The library supports efficient training for separable GP kernels, which allows the use of Kronecker algebra to handle high-dimensional input spaces and large numbers of correlated outputs. This makes it well suited for applications such as modeling cosmological power spectra, large-scale physical simulations, and multi-output hyperparameter tuning. By combining scalable multi-output GP modeling with data-driven adaptive sampling, GPsurrogate enables parameter inference and optimization with substantially fewer simulations than conventional space-filling designs.

Lukic, Zarija [Lawrence Berkeley National Laborato↗
DOE OSTI2018

High Efficiency, Inexpensive Thin Film III-V Photovoltaics using Single-Crystalline-Like, Flexible Substrates

An innovative, non-mainstream approach has been developed to fabricate inexpensive GaAs solar cells directly on epi-ready flexible metal substrates. Using templates made by ion beam assisted deposition (IBAD) and epitaxial germanium films made by Plasma Enhanced Chemical Vapor Deposition (PECVD), epitaxial GaAs films with high opto-electronic and structural quality have been grown on metal substrates using metal organic chemical vapor deposition (MOCVD). Crystallographic out-of-plane and in-plane texture spread less than 1 degree full-width-at-half-maximum (FWHM) was achieved in undoped, n-type and p-type GaAs films as well as in AlGaAs films on metal substrates. Electron mobility of >1300 cm 2 /V-s was attained in GaAs:Si n-type films. Also, hole mobility of >100 cm 2 /V-s was measured in GaAs:Zn films. Consistent doping was obtained in GaAs:Si (10 15 -10 18 cm -3 ), and GaAs:Zn films (10 17 -10 19 cm -3 ), necessary for the active device layers of a GaAs solar cell. Defect reduction techniques (thermal cycling, superlattice structures) were implemented to reduce defect density in GaAs films from 1 × 10 9 cm -2 to 9 × 10 7 cm -2 . Lifetime of 2.3 ns was obtained in AlGaAs/GaAs double-heterostructure device. Solar cells were fabricated based on designs using 1-dimensional and 2-dimensional simulations tailored to the specific properties of the GaAs films on metal substrates. All device fabrication processes (lithography, anti-reflection coating (ARC), device isolation and mesa formation and ohmic metal electrode formation) suitable for III-V solar cell fabrication on metal substrates were fully developed. A bi-layered ARC using ZnS and MgF 2 was used to minimize the reflection over wide range of spectrum. When this ARC was applied after removal of the parasitic absorption layer, the conversion efficiencies were improved up to 11.1% with V oc = ~0.621 V, J sc = ~24.5 mA/cm 2 , and FF = ~0.735. Yet another accomplishment in this project was the development of all electrically-conducting architecture for GaAs solar cells on metal substrates. Biaxial-texture was successfully achieved in all electrically-conducting buffer layer on metal substrate in structure based on IBAD TiN instead of IBAD MgO. The all electrically-conducting Ge/NiSi 2 /TiN/IBAD-TiN/Hastelloy substrate was used to fabricate back contacts for vertically conducting solar cell structures on metal substrates. A metallic resistivity as low as 4.17 μΩ•cm was obtained. A significant improvement of the texture of the Ge template was achieved using a Ag buffer layer between the TiN films. Epitaxial GaAs was successfully grown on the all-electrically-conducting buffer architecture. A Liquid Phase Epitaxy (LPE) method to drastically increase grain size of epitaxial Ge while maintaining the single-crystalline-like texture. Grain size up to 35 µm was achieved in LPE Ge grown on sputtered Ge with a grain size of ~ 1 µm on IBAD template on metal substrate. Epitaxial GaAs was grown on LPE Ge, with an out-of-plane texture of 0.34° compared to 0.98° in GaAs grown on sputter deposited Ge. A model was developed to simulate the effect of grain boundaries as a function of grain size of GaAs films on metal substrates. This model showed that even the very low angle grain boundaries (<1º misorientation) of our GaAs films led to decrease in Voc. Hydrogen plasma passivation treatment is being developed to improve open circuit voltage and the conversion efficiency of GaAs solar cells on flexible substrate.

14 SOLAR ENERGY↗
DOE OSTI2023

Extending Parsimonious Bayesian Inference

Parsimonious Bayesian inference is a theoretical framework for efficient data assimilation that seeks to balance increased consistency between predictions and training data against corresponding increases in model complexity. Within this framework, over-training is understood as optimization that encodes excessive information within model parameters while only achieving small improvements between predictions and training data. This project aims to develop practical methods of limiting excess model information during optimization. One key observation is that practical heuristics for parsimonious learning in high-dimensions must balance expressivity, i.e. the ability of the model to capture diverse predictions with only a few non-zero parameters, against discoverability, i.e. the ability to train the model with gradient-based optimization and drive parameters to low information states. As such, we developed logical activation functions that are able to adaptively approximate arbitrary truth tables that define Boolean logic operations within a probabilistic framework. These functions have demonstrated the ability to learn exclusive disjunction (XOR) and conditioned disjunction (if [condition] then [result_if_true] else [result_if_false]) within a single layer of a neural network. To efficiently exploit these activation functions to drive parsimonious learning required several other advances within the domain of variational inference. The most efficient form of complexity suppression is structured sparsification, driving most model parameters to zero while achieving the structural coherence among nonzeros needed for bandwidth reduction. Such models are not only far more efficient at suppressing information-theoretic complexity, they also reduce the other forms of complexity (computations, communication, storage, and the number of dependencies needed to evaluate predictions). Aiming to support enhanced sparsification, this project examined new approaches to high-dimensional variational inference that allow us to calibrate and control parameter uncertainty during optimization. By identifying which parameters can sustain sparsifying perturbations with little impact on prediction quality, we can develop better pruning strategies by framing them as approximate Bayesian inference. These advances also open paths to mitigate concerns with deploying advanced learning methods in resource-constrained environments, such as running models on power-limited or communication-limited devices.

97 MATHEMATICS AND COMPUTING↗