-
Federated Bayesian Deep Learning: The Application of Statistical Aggregation Methods to Bayesian Models
Authors:
John Fischer,
Marko Orescanin,
Justin Loomis,
Patrick McClure
Abstract:
Federated learning (FL) is an approach to training machine learning models that takes advantage of multiple distributed datasets while maintaining data privacy and reducing communication costs associated with sharing local datasets. Aggregation strategies have been developed to pool or fuse the weights and biases of distributed deterministic models; however, modern deterministic deep learning (DL)…
▽ More
Federated learning (FL) is an approach to training machine learning models that takes advantage of multiple distributed datasets while maintaining data privacy and reducing communication costs associated with sharing local datasets. Aggregation strategies have been developed to pool or fuse the weights and biases of distributed deterministic models; however, modern deterministic deep learning (DL) models are often poorly calibrated and lack the ability to communicate a measure of epistemic uncertainty in prediction, which is desirable for remote sensing platforms and safety-critical applications. Conversely, Bayesian DL models are often well calibrated and capable of quantifying and communicating a measure of epistemic uncertainty along with a competitive prediction accuracy. Unfortunately, because the weights and biases in Bayesian DL models are defined by a probability distribution, simple application of the aggregation methods associated with FL schemes for deterministic models is either impossible or results in sub-optimal performance. In this work, we use independent and identically distributed (IID) and non-IID partitions of the CIFAR-10 dataset and a fully variational ResNet-20 architecture to analyze six different aggregation strategies for Bayesian DL models. Additionally, we analyze the traditional federated averaging approach applied to an approximate Bayesian Monte Carlo dropout model as a lightweight alternative to more complex variational inference methods in FL. We show that aggregation strategy is a key hyperparameter in the design of a Bayesian FL system with downstream effects on accuracy, calibration, uncertainty quantification, training stability, and client compute requirements.
△ Less
Submitted 4 April, 2024; v1 submitted 22 March, 2024;
originally announced March 2024.
-
Nanostructured Polymer Films with Metal-like Thermal Conductivity
Authors:
Yanfei Xu,
Daniel Kraemer,
Bai Song,
Zhang Jiang,
Jiawei Zhou,
James Loomis,
Jianjian Wang,
Mingda Li,
Hadi Ghasemi,
Xiaopeng Huang,
Xiaobo Li,
Gang Chen
Abstract:
Thermally conductive polymers are of fundamental interest and can also be exploited in thermal management applications. Recent studies have shown stretched polymers can achieve high thermal conductivity. However, the transport mechanisms of heat in thermally conductive polymers have yet to be elucidated. Here we report a method for scalable fabrication of polyethylene films with a high thermal con…
▽ More
Thermally conductive polymers are of fundamental interest and can also be exploited in thermal management applications. Recent studies have shown stretched polymers can achieve high thermal conductivity. However, the transport mechanisms of heat in thermally conductive polymers have yet to be elucidated. Here we report a method for scalable fabrication of polyethylene films with a high thermal conductivity of 62 W/m-K. The achieved thermal conductivity is over two orders-of-magnitude greater than that of typical polymers (~0.1 W/m-K), and exceeds those of many metals and ceramics used as traditional heat conductors. Careful structural studies are carried out and reveal that the film consists of nanofibers with crystalline and amorphous regions. Contrary to conventional wisdom, we reveal the importance of the amorphous morphology in achieving such high thermal conductivity, rather than simply from enhancements in the degree of crystallinity and crystallite alignment. The amorphous phase reaches a remarkably high thermal conductivity of ~16 W/m-K. Even still, we identify that the presence of this amorphous phase is the dominant factor as the film thermal conductivity is still much lower than the predicted values for bulk single-crystal polyethylene (237 K/m-K). This work lays the foundation for the rational design and synthesis of thermally conductive polymers, and opens up new opportunities for advanced heat management, particularly when flexible, lightweight, chemically inert and electrically insulating thermal conductors are desired.
△ Less
Submitted 21 August, 2017;
originally announced August 2017.
-
Hybrid optical-thermal devices and materials for light manipulation and radiative cooling
Authors:
Svetlana V. Boriskina,
Jonathan K. Tong,
Wei-Chun Hsu,
Lee Weinstein,
Xiaopeng Huang,
James Loomis,
Yanfei Xu,
Gang Chen
Abstract:
We report on optical design and applications of hybrid meso-scale devices and materials that combine optical and thermal management functionalities owing to their tailored resonant interaction with light in visible and infrared frequency bands. We outline a general approach to designing such materials, and discuss two specific applications in detail. One example is a hybrid optical-thermal antenna…
▽ More
We report on optical design and applications of hybrid meso-scale devices and materials that combine optical and thermal management functionalities owing to their tailored resonant interaction with light in visible and infrared frequency bands. We outline a general approach to designing such materials, and discuss two specific applications in detail. One example is a hybrid optical-thermal antenna with sub-wavelength light focusing, which simultaneously enables intensity enhancement at the operating wavelength in the visible and reduction of the operating temperature. The enhancement is achieved via light recycling in the form of whispering-gallery modes trapped in an optical microcavity, while cooling functionality is realized via a combination of reduced optical absorption and radiative cooling. The other example is a fabric that is opaque in the visible range yet highly transparent in the infrared, which allows the human body to efficiently shed energy in the form of thermal emission. Such fabrics can find numerous applications for personal thermal management and for buildings energy efficiency improvement.
△ Less
Submitted 8 September, 2015;
originally announced September 2015.
-
Infrared-Transparent Visible-Opaque Fabrics for Wearable Personal Thermal Management
Authors:
Jonathan K. Tong,
Xiaopeng Huang,
Svetlana V. Boriskina,
James Loomis,
Yanfei Xu,
Gang Chen
Abstract:
Personal cooling technologies locally control the temperature of an individual rather than a large space, thus providing personal thermal comfort while supplementing cooling loads in thermally regulated environments. This can lead to significant energy and cost savings. In this study, a new approach to personal cooling was developed using an infrared-transparent visible-opaque fabric (ITVOF), whic…
▽ More
Personal cooling technologies locally control the temperature of an individual rather than a large space, thus providing personal thermal comfort while supplementing cooling loads in thermally regulated environments. This can lead to significant energy and cost savings. In this study, a new approach to personal cooling was developed using an infrared-transparent visible-opaque fabric (ITVOF), which provides passive cooling via the transmission of thermal radiation emitted by the human body directly to the environment. Here, we present a conceptual framework to thermally and optically design an ITVOF. Using a heat transfer model, the fabric was found to require a minimum infrared (IR) transmittance of 0.644 and a maximum IR reflectance of 0.2 to ensure thermal comfort at ambient temperatures as high as 26.1oC (79oF). To meet these requirements, an ITVOF design was developed using synthetic polymer fibers with an intrinsically low IR absorptance. These fibers were then structured to minimize IR reflection via weak Rayleigh scattering while maintaining visible opaqueness via strong Mie scattering. For a fabric composed of parallel-aligned polyethylene fibers, numerical finite element simulations predict 1 μm diameter fibers bundled into 30 μm yarns can achieve a total hemispherical IR transmittance of 0.972, which is nearly perfectly transparent to mid- and far-IR radiation. The visible wavelength properties of the ITVOF are comparable to conventional textiles ensuring opaqueness to the human eye. By providing personal cooling in a form amenable to everyday use, ITVOF-based clothing offers a simple, low-cost solution to reduce energy consumption in HVAC systems.
△ Less
Submitted 15 July, 2015;
originally announced July 2015.
-
Linewidth of single photon transitions in Mn$_{12}$-acetate
Authors:
Beth Parks,
Joseph Loomis,
Evan Rumberger,
David N. Hendrickson,
George Christou
Abstract:
We use time-domain terahertz spectroscopy to measure the position and linewidth of single photon transitions in Mn$_{12}$-acetate. This linewidth is compared to the linewidth measured in tunneling experiments. We conclude that local magnetic fields (due to dipole or hyperfine interactions) cannot be responsible for the observed linewidth, and suggest that the linewidth is due to variations in th…
▽ More
We use time-domain terahertz spectroscopy to measure the position and linewidth of single photon transitions in Mn$_{12}$-acetate. This linewidth is compared to the linewidth measured in tunneling experiments. We conclude that local magnetic fields (due to dipole or hyperfine interactions) cannot be responsible for the observed linewidth, and suggest that the linewidth is due to variations in the anisotropy constants for different clusters. We also calculate a lower limit on the dipole field distribution that would be expected due to random orientations of clusters and find that collective effects must narrow this distribution in tunneling measurements.
△ Less
Submitted 31 October, 2001;
originally announced October 2001.
