-
Federated Evaluation and Tuning for On-Device Personalization: System Design & Applications
Authors:
Matthias Paulik,
Matt Seigel,
Henry Mason,
Dominic Telaar,
Joris Kluivers,
Rogier van Dalen,
Chi Wai Lau,
Luke Carlson,
Filip Granqvist,
Chris Vandevelde,
Sudeep Agarwal,
Julien Freudiger,
Andrew Byde,
Abhishek Bhowmick,
Gaurav Kapoor,
Si Beaumont,
Áine Cahill,
Dominic Hughes,
Omid Javidbakht,
Fei Dong,
Rehan Rishi,
Stanley Hung
Abstract:
We describe the design of our federated task processing system. Originally, the system was created to support two specific federated tasks: evaluation and tuning of on-device ML systems, primarily for the purpose of personalizing these systems. In recent years, support for an additional federated task has been added: federated learning (FL) of deep neural networks. To our knowledge, only one other…
▽ More
We describe the design of our federated task processing system. Originally, the system was created to support two specific federated tasks: evaluation and tuning of on-device ML systems, primarily for the purpose of personalizing these systems. In recent years, support for an additional federated task has been added: federated learning (FL) of deep neural networks. To our knowledge, only one other system has been described in literature that supports FL at scale. We include comparisons to that system to help discuss design decisions and attached trade-offs. Finally, we describe two specific large scale personalization use cases in detail to showcase the applicability of federated tuning to on-device personalization and to highlight application specific solutions.
△ Less
Submitted 16 February, 2021;
originally announced February 2021.
-
Optimization of Solidification in Die Casting using Numerical Simulations and Machine Learning
Authors:
Shantanu Shahane,
Narayana Aluru,
Placid Ferreira,
Shiv G Kapoor,
Surya Pratap Vanka
Abstract:
In this paper, we demonstrate the combination of machine learning and three dimensional numerical simulations for multi-objective optimization of low pressure die casting. The cooling of molten metal inside the mold is achieved typically by passing water through the cooling lines in the die. Depending on the cooling line location, coolant flow rate and die geometry, nonuniform temperatures are imp…
▽ More
In this paper, we demonstrate the combination of machine learning and three dimensional numerical simulations for multi-objective optimization of low pressure die casting. The cooling of molten metal inside the mold is achieved typically by passing water through the cooling lines in the die. Depending on the cooling line location, coolant flow rate and die geometry, nonuniform temperatures are imposed on the molten metal at the mold wall. This boundary condition along with the initial molten metal temperature affect the product quality quantified in terms of micro-structure parameters and yield strength. A finite volume based numerical solver is used to determine the temperature-time history and correlate the inputs to outputs. The objective of this research is to develop and demonstrate a procedure to obtain the initial and wall temperatures so as to optimize the product quality. The non-dominated sorting genetic algorithm (NSGA-II) is used for multi-objective optimization in this work. The number of function evaluations required for NSGA-II can be of the order of millions and hence, the finite volume solver cannot be used directly for optimization. Therefore, a multilayer perceptron feed-forward neural network is first trained using the results from the numerical solution of the fluid flow and energy equations and is subsequently used as a surrogate model. As an assessment, simplified versions of the actual problem are designed to first verify results of the genetic algorithm. An innovative local sensitivity based approach is then used to rank the final Pareto optimal solutions and select a single best design.
△ Less
Submitted 3 January, 2020; v1 submitted 8 January, 2019;
originally announced January 2019.
-
Protection Against Reconstruction and Its Applications in Private Federated Learning
Authors:
Abhishek Bhowmick,
John Duchi,
Julien Freudiger,
Gaurav Kapoor,
Ryan Rogers
Abstract:
In large-scale statistical learning, data collection and model fitting are moving increasingly toward peripheral devices---phones, watches, fitness trackers---away from centralized data collection. Concomitant with this rise in decentralized data are increasing challenges of maintaining privacy while allowing enough information to fit accurate, useful statistical models. This motivates local notio…
▽ More
In large-scale statistical learning, data collection and model fitting are moving increasingly toward peripheral devices---phones, watches, fitness trackers---away from centralized data collection. Concomitant with this rise in decentralized data are increasing challenges of maintaining privacy while allowing enough information to fit accurate, useful statistical models. This motivates local notions of privacy---most significantly, local differential privacy, which provides strong protections against sensitive data disclosures---where data is obfuscated before a statistician or learner can even observe it, providing strong protections to individuals' data. Yet local privacy as traditionally employed may prove too stringent for practical use, especially in modern high-dimensional statistical and machine learning problems. Consequently, we revisit the types of disclosures and adversaries against which we provide protections, considering adversaries with limited prior information and ensuring that with high probability, ensuring they cannot reconstruct an individual's data within useful tolerances. By reconceptualizing these protections, we allow more useful data release---large privacy parameters in local differential privacy---and we design new (minimax) optimal locally differentially private mechanisms for statistical learning problems for \emph{all} privacy levels. We thus present practicable approaches to large-scale locally private model training that were previously impossible, showing theoretically and empirically that we can fit large-scale image classification and language models with little degradation in utility.
△ Less
Submitted 3 June, 2019; v1 submitted 3 December, 2018;
originally announced December 2018.
-
Finite Volume Simulation Framework for Die Casting with Uncertainty Quantification
Authors:
Shantanu Shahane,
Narayana Aluru,
Placid Ferreira,
Shiv G Kapoor,
Surya Pratap Vanka
Abstract:
The present paper describes the development of a novel and comprehensive computational framework to simulate solidification problems in materials processing, specifically casting processes. Heat transfer, solidification and fluid flow due to natural convection are modeled. Empirical relations are used to estimate the microstructure parameters and mechanical properties. The fractional step algorith…
▽ More
The present paper describes the development of a novel and comprehensive computational framework to simulate solidification problems in materials processing, specifically casting processes. Heat transfer, solidification and fluid flow due to natural convection are modeled. Empirical relations are used to estimate the microstructure parameters and mechanical properties. The fractional step algorithm is modified to deal with the numerical aspects of solidification by suitably altering the coefficients in the discretized equation to simulate selectively only in the liquid and mushy zones. This brings significant computational speed up as the simulation proceeds. Complex domains are represented by unstructured hexahedral elements. The algebraic multigrid method, blended with a Krylov subspace solver is used to accelerate convergence. State of the art uncertainty quantification technique is included in the framework to incorporate the effects of stochastic variations in the input parameters. Rigorous validation is presented using published experimental results of a solidification problem.
△ Less
Submitted 19 October, 2018;
originally announced October 2018.
-
Means Moments and Newton's Inequalities
Authors:
R. Sharma,
A. Sharma,
R. Saini,
G. Kapoor
Abstract:
It is shown that Newton's inequalities and the related Maclaurin's inequalities provide several refinements of the fundamental Arithmetic mean - Geometric mean - Harmonic mean inequality in terms of the means and variance of positive real numbers. We also obtain some inequalities involving third and fourth central moments of real numbers.
It is shown that Newton's inequalities and the related Maclaurin's inequalities provide several refinements of the fundamental Arithmetic mean - Geometric mean - Harmonic mean inequality in terms of the means and variance of positive real numbers. We also obtain some inequalities involving third and fourth central moments of real numbers.
△ Less
Submitted 15 February, 2017;
originally announced February 2017.
-
Complementary upper bounds for fourth central moment with extensions and applications
Authors:
R. Sharma,
R. Kumar,
R. Saini,
G. Kapoor
Abstract:
We prove some inequalities involving fourth central moment of a random variable that takes values in a given finite interval. Both discrete and continuous cases are considered. Bounds for the spread are obtained when a given nxn complex matrix has real eigenvalues. Likewise, we discuss bounds for the spans of polynomial equations.
We prove some inequalities involving fourth central moment of a random variable that takes values in a given finite interval. Both discrete and continuous cases are considered. Bounds for the spread are obtained when a given nxn complex matrix has real eigenvalues. Likewise, we discuss bounds for the spans of polynomial equations.
△ Less
Submitted 12 March, 2015;
originally announced March 2015.
-
Chaos in Dynamics of a Family of Transcendental Meromorphic Functions
Authors:
M. Sajid,
G. P. Kapoor
Abstract:
The characterization and properties of Julia sets of one parameter family of transcendental meromorphic functions $ζ_λ(z)=λ\frac{z}{z+1} e^{-z}$, $λ>0$, $z\in \mathbb{C}$ is investigated in the present paper. It is found that bifurcations in the dynamics of $ζ_λ(x)$, $x\in {\mathbb{R}}\setminus \{-1\}$, occur at several parameter values and the dynamics of the family becomes chaotic when the param…
▽ More
The characterization and properties of Julia sets of one parameter family of transcendental meromorphic functions $ζ_λ(z)=λ\frac{z}{z+1} e^{-z}$, $λ>0$, $z\in \mathbb{C}$ is investigated in the present paper. It is found that bifurcations in the dynamics of $ζ_λ(x)$, $x\in {\mathbb{R}}\setminus \{-1\}$, occur at several parameter values and the dynamics of the family becomes chaotic when the parameter $λ$ crosses certain values. The Lyapunov exponent of $ζ_λ(x)$ for certain values of the parameter $λ$ is computed for quantifying the chaos in its dynamics. The characterization of the Julia set of the function $ζ_λ(z)$ as complement of the basin of attraction of an attracting real fixed point of $ζ_λ(z)$ is found here and is applied to computationally simulate the images of the Julia sets of $ζ_λ(z)$. Further, it is established that the Julia set of $ζ_λ(z)$ for $λ>(\sqrt{2}+1) e^{\sqrt{2}}$ contains the complement of attracting periodic orbits of $ζ_λ(x)$. Finally, the results on the dynamics of functions $λ\tan z$, $λ\in {\mathbb{\hat{C}}}\setminus\{0\}$, $E_λ(z) = λ\frac{e^{z} -1}{z}$, $λ> 0$ and $f_λ=λf(z)$, $λ>0$, where $f(z)$ has certain properties, are compared with the results found in the present paper.
△ Less
Submitted 7 September, 2014;
originally announced September 2014.
-
Lidstone Fractal Interpolation and Error Analysis
Authors:
G. P. Kapoor,
M. Sahoo
Abstract:
In the present paper, the notion of Lidstone Fractal Interpolation Function ($Lidstone \ FIF$) is introduced to interpolate and approximate data generating functions that arise from real life objects and outcomes of several scientific experiments. A Lidstone FIF extends the classical Lidstone Interpolation Function which is generally found not to be satisfactory in interpolation and approximation…
▽ More
In the present paper, the notion of Lidstone Fractal Interpolation Function ($Lidstone \ FIF$) is introduced to interpolate and approximate data generating functions that arise from real life objects and outcomes of several scientific experiments. A Lidstone FIF extends the classical Lidstone Interpolation Function which is generally found not to be satisfactory in interpolation and approximation of such functions. For a data $\{(x_n,y_{n,2k}); n=0,1,\ldots,N \ \text{and} \ k=0,1,\ldots,p\}$ with $N,p\in\mathbb{N}$, the existence of Lidstone FIF is proved in the present work and a computational method for its construction is developed. The constructed Lidstone FIF is a $C^{2p}[x_0,x_N]$ fractal function $\ell_α$ satisfying $\ell_α^{(2k)}(x_n)=y_{n,2k}$, $n=0,1,\ldots,N$,\ $k=0,1,\ldots,p$. Our error estimates establish that the order of $L^\infty$-error in approximation of a data generating function in $C^{2p}[x_0,x_N]$ by Lidstone FIF is of the order $N^{-2p}$, while $L^\infty$-error in approximation of $2k$-order derivative of the data generating function by corresponding order derivative of Lidstone FIF is of the order $N^{-(2p-2k)}$. The results found in the present work are illustrated for computational constructions of a Lidstone FIF and its derivatives with an example of a data generating function.
△ Less
Submitted 9 July, 2014;
originally announced July 2014.
-
Convergence of Cubic Spline Super Fractal Interpolation Functions
Authors:
G. P. Kapoor,
Srijanani Anurag Prasad
Abstract:
In the present work, the notion of Cubic Spline Super Fractal Interpolation Function (SFIF) is introduced to simulate an object that depicts one structure embedded into another and its approximation properties are investigated. It is shown that, for an equidistant partition points of [x_0,x_N], the interpolating Cubic Spline (SFIF) and their derivatives converge respectively to the data generating…
▽ More
In the present work, the notion of Cubic Spline Super Fractal Interpolation Function (SFIF) is introduced to simulate an object that depicts one structure embedded into another and its approximation properties are investigated. It is shown that, for an equidistant partition points of [x_0,x_N], the interpolating Cubic Spline (SFIF) and their derivatives converge respectively to the data generating function and its derivatives at the rate of h^(2-j+e) (0<e<1), j=0,1,2 as the norm h of the partition of [x_0,x_N] approaches zero. The convergence results for Cubic Spline (SFIF) found here show that any desired accuracy can be achieved in the approximation of a regular data generating function and its derivatives by a Cubic Spline (SFIF) and its corresponding derivatives.
△ Less
Submitted 19 January, 2012;
originally announced January 2012.
-
Multiresolution Analysis Based on Coalescence Hidden-variable FIF
Authors:
G. P. Kapoor,
Srijanani Anurag Prasad
Abstract:
In the present paper, multiresolution analysis arising from Coalescence Hidden-variable Fractal Interpolation Functions (CHFIFs) is accomplished. The availability of a larger set of free variables and constrained variables with CHFIF in multiresolution analysis based on CHFIFs provides more control in reconstruction of functions in L2(\mathbb{R})than that provided by multiresolution analysis based…
▽ More
In the present paper, multiresolution analysis arising from Coalescence Hidden-variable Fractal Interpolation Functions (CHFIFs) is accomplished. The availability of a larger set of free variables and constrained variables with CHFIF in multiresolution analysis based on CHFIFs provides more control in reconstruction of functions in L2(\mathbb{R})than that provided by multiresolution analysis based only on Affine Fractal Interpolation Functions (AFIFs). In our approach, the vector space of CHFIFs is introduced, its dimension is determined and Riesz bases of vector subspaces Vk, k \in \mathbb{Z}, consisting of certain CHFIFs in L2(\mathbb{R}) \cap C0(\mathbb{R}) are constructed. As a special case, for the vector space of CHFIFs of dimension 4, orthogonal bases for the vector subspaces Vk, k \in \mathbb{Z}, are explicitly constructed and, using these bases, compactly supported continuous orthonormal wavelets are generated.
△ Less
Submitted 17 January, 2012;
originally announced January 2012.
-
Super Fractal Interpolation Functions
Authors:
G. P. Kapoor,
Srijanani Anurag Prasad
Abstract:
In the present work, the notion of Super Fractal Interpolation Function (SFIF) is introduced for finer simulation of the objects of the nature or outcomes of scientific experiments that reveal one or more structures embedded in to another. In the construction of SFIF, an IFS is chosen from a pool of several IFS at each level of iteration leading to implementation of the desired randomness and vari…
▽ More
In the present work, the notion of Super Fractal Interpolation Function (SFIF) is introduced for finer simulation of the objects of the nature or outcomes of scientific experiments that reveal one or more structures embedded in to another. In the construction of SFIF, an IFS is chosen from a pool of several IFS at each level of iteration leading to implementation of the desired randomness and variability in fractal interpolation of the given data. Further, an expository description of our investigations on the integral, the smoothness and determination of conditions for existence of derivatives of a SFIF is given in the present work.
△ Less
Submitted 17 January, 2012;
originally announced January 2012.
-
Stability of Coalescence Hidden variable Fractal Interpolation Surfaces
Authors:
G. P. Kapoor,
Srijanani Anurag Prasad
Abstract:
In the present paper, the stability of Coalescence Hidden variable Fractal Interpolation Surfaces(CHFIS) is established. The estimates on error in approximation of the data generating function by CHFIS are found when there is a perturbation in independent, dependent and hidden variables. It is proved that any small perturbation in any of the variables of generalized interpolation data results in…
▽ More
In the present paper, the stability of Coalescence Hidden variable Fractal Interpolation Surfaces(CHFIS) is established. The estimates on error in approximation of the data generating function by CHFIS are found when there is a perturbation in independent, dependent and hidden variables. It is proved that any small perturbation in any of the variables of generalized interpolation data results in only small perturbation of CHFIS. Our results are likely to be useful in investigations of texture of surfaces arising from the simulation of surfaces of rocks, sea surfaces, clouds and similar natural objects wherein the generating function depends on more than one variable.
△ Less
Submitted 3 September, 2009;
originally announced September 2009.
-
Coefficient Estimates for Inverses of Starlike Functions of Positive Order
Authors:
G. P. Kapoor,
A. K. Mishra
Abstract:
In the present paper, the coefficient estimates are found for the class $\mathcal S^{*-1}(α)$ consisting of inverses of functions in the class of univalent starlike functions of order $α$ in $\mathcal D=\{z\in\mathbb C:|z|<1\}$. These estimates extend the work of {\it Krzyz, Libera and Zlotkiewicz [Ann. Univ. Marie Curie-Sklodowska, 33(1979), 103-109]} who found sharp estimates on only first two…
▽ More
In the present paper, the coefficient estimates are found for the class $\mathcal S^{*-1}(α)$ consisting of inverses of functions in the class of univalent starlike functions of order $α$ in $\mathcal D=\{z\in\mathbb C:|z|<1\}$. These estimates extend the work of {\it Krzyz, Libera and Zlotkiewicz [Ann. Univ. Marie Curie-Sklodowska, 33(1979), 103-109]} who found sharp estimates on only first two coefficients for the functions in the class $\mathcal S^{*-1}(α)$. The coefficient estimates are also found for the class $\sum^{*-1}(α)$, consisting of inverses of functions in the class $\sum^*(α)$ of univalent starlike functions of order $α$ in $\mathcal V=\{z\in\mathbb C:1<|z|<\infty\}$. The open problem of finding sharp coefficient estimates for functions in the class $\sum^*(α)$ stands completely settled in the present work by our method developed here.
△ Less
Submitted 3 November, 2005;
originally announced November 2005.
-
Coalescence Hidden Variable Fractal Interpolation Functions and its Smoothness Analysis
Authors:
A. K. B. Chand,
G. P. Kapoor
Abstract:
We construct a coalescence hidden variable fractal interpolation function (CHFIF) through a non-diagonal iterated function system(IFS). Such a FIF may be self-affine or non-self-affine depending on the parameters of the defining non-diagonal IFS. The smoothness analysis of the CHFIF has been carried out by using the operator approximation technique. The deterministic construction of functions ha…
▽ More
We construct a coalescence hidden variable fractal interpolation function (CHFIF) through a non-diagonal iterated function system(IFS). Such a FIF may be self-affine or non-self-affine depending on the parameters of the defining non-diagonal IFS. The smoothness analysis of the CHFIF has been carried out by using the operator approximation technique. The deterministic construction of functions having order of modulus continuity O (|t|^{delta} (log|t|)^m), (m a non-negative integer and 0 < delta le 1) is possible through our CHFIF. The bounds of fractal dimension of CHFIFs are obtained first in certain critical cases and then, using estimation of these bounds, the bounds of fractal dimension of any FIF are found.
△ Less
Submitted 3 November, 2005;
originally announced November 2005.
