DOCSLIB.ORG

EA-CG: an Approximate Second-Order Method for Training Fully-Connected Neural Networks

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
EA-CG: an Approximate Second-Order Method for Training Fully-Connected Neural Networks The Thirty-Third AAAI Conference on Artificial Intelligence (AAAI-19) EA-CG: An Approximate Second-Order Method for Training Fully-Connected Neural Networks Sheng-Wei Chen Chun-Nan Chou Edward Y. Chang HTC Research & Healthcare HTC Research & Healthcare HTC Research & Healthcare sw [email protected] jason.cn [email protected] edward [email protected] Abstract the common practice is to use the Gauss-Newton matrix with a convex criterion function (Schraudolph 2002) or the Fisher For training fully-connected neural networks (FCNNs), we propose a practical approximate second-order method includ- matrix to measure curvature since both are guaranteed to be ing: 1) an approximation of the Hessian matrix and 2) a conju- positive semi-definite (PSD). gate gradient (CG) based method. Our proposed approximate Although these two kinds of matrices can alleviate the is- Hessian matrix is memory-efficient and can be applied to any sue of negative curvature, computing either the exact Gauss- FCNNs where the activation and criterion functions are twice Newton matrix or Fisher matrix even for a modestly-sized differentiable. We devise a CG-based method incorporating fully-connected neural network (FCNN) is intractable. In- one-rank approximation to derive Newton directions for train- tuitively, the analytic expression for the second derivative ing FCNNs, which significantly reduces both space and time requires O(N 2) computations if O(N) complexity is re- complexity. This CG-based method can be employed to solve any linear equation where the coefficient matrix is Kronecker- quired to compute the first derivative. Thus, several pioneer factored, symmetric and positive definite. Empirical studies works (LeCun et al. 1998; Amari, Park, and Fukumizu 2000; show the efficacy and efficiency of our proposed method. Schraudolph 2002) have used different methods to approx- imate either matrix. However, none of these methods have been shown to be computationally feasible and fundamen- Introduction tally more effective than first-order methods as reported Neural networks have been applied to solving problems in in (Martens 2010). Thus, there has been a growing trend several application domains such as computer vision (He towards conceiving more computationally feasible second- et al. 2016), natural language processing (Hochreiter and order methods for training FCNNs. Schmidhuber 1997), and disease diagnosis (Chang et al. We outline several notable works in chronological or- 2017). Training a neural network requires tuning its model der herein. Martens proposed a truncated-Newton method parameters using Backpropagation. Stochastic gradient de- for training deep auto-encoders. In this work, Martens used scent (SGD), Broyden-Fletcher-Goldfarb-Shanno and one- an R-operator (Pearlmutter 1994) to compute full Gauss- step secant are representative algorithms that have been em- Newton matrix-vector products and made good progress ployed for training in Backpropagation. within each update. Martens and Grosse developed a To date, SGD is widely used due to its low computational block-diagonal approximation to the Fisher matrix for FC- demand. SGD minimizes a function using the function’s first NNs, called Kronecker-factored Approximation Curvature derivative, and has been proven to be effective for train- (KFAC). They derived the update directions by exploiting ing large models. However, stochasticity in the gradient will the inverse property of Kronecker products. KFAC features slow down convergence for any gradient method such that that its cost in storing and inverting the devised approxima- none of them can be asymptotically faster than simple SGD tion does not depend on the amount of data used to estimate with Polyak averaging (Polyak and Juditsky 1992). Besides the Fisher matrix. The idea is further extended to convo- gradients, second-order methods utilize the curvature infor- lutional nets (Grosse and Martens 2016). Recently, Botev, mation of a loss function within the neighborhood of a given Ritter, and Barber presented a block-diagonal approxima- point to guide the update direction. Since each update be- tion to the Gauss-Newton matrix for FCNNs, referred to comes more precise, such methods can converge faster than as Kronecker-Factored Recursive Approximation (KFRA). first-order methods in terms of update iterations. They also utilized the inverse property of Kronecker prod- For solving a convex optimization problem, a second- ucts to derive the update directions. Similarly, Zhang et al. order method can always converge to the global minimum introduced a block-diagonal approximation of the Gauss- in much fewer steps than SGD. However, the problem of Newton matrix and used conjugate gradient (CG) method neural-network training can be non-convex, thereby suffer- to derive the update directions. ing from the issue of negative curvature. To avoid this issue, However, these prior works either impose constraints on Copyright c 2019, Association for the Advancement of Artificial their applicability or require considerable computation in Intelligence (www.aaai.org). All rights reserved. terms of memory or running time. On the one hand, sev- 3337 eral of these notable methods relying on the Gauss-Newton Suppose we have a minimization problem matrix face the essential limit that they cannot handle non- min f(θ); (1) convex criterion functions playing an important part in some θ problems. Use robust estimation in computer vision (Stew- art 1999) as an example. Some non-convex criterion func- where f is a convex and twice-differentiable function. Since the global minimum is at the point that the first derivative is tions such as Tukey’s biweight function are robustness to ∗ outliers and perform better than convex criterion functions zero, the solution θ can be derived from the equation (Belagiannis et al. 2015). On the other hand, in order to de- rf(θ∗) = 0: (2) rive the update directions, some of the methods utilizing the conventional CG method may take too much time, and the We can utilize a quadratic polynomial to approximate Prob- others exploiting the inverse property of Kronecker products lem 1 by conducting a Taylor expansion with a given point may require excessive memory space. θj. Then, the problem turns out to be To remedy the aforementioned issues, we propose a 1 block-diagonal approximation of the positive-curvature Hes- min f(θj + d) ≈ f(θj) + rf(θj)T d + dT r2f(θj)d; d 2 sian (PCH) matrix, which is memory-efficient. Our proposed PCH matrix can be applied to any FCNN where the acti- where r2f(θj) is the Hessian matrix of f at θj. Af- vation and criterion functions are twice differentiable. Par- ter applying the aforementioned approximation, we rewrite ticularly, our proposed PCH matrix can handle non-convex Eq. (2) as the linear equation criterion functions, which the Gauss-Newton methods can- j 2 j j not. Besides, we incorporate expectation approximation into rf(θ ) + r f(θ )d = 0: (3) the CG-based method, which is dubbed EA-CG, to derive Therefore, the Newton direction is obtained via update directions for training FCNNs in mini-batch setting. EA-CG significantly reduces the space and time complexity dj = −∇2f(θj)−1rf(θj); of the conventional CG method. Our experimental results ∗ show the efficacy and efficiency of our proposed method. and we can acquire θ by iteratively applying the update rule In this work, we focus on deriving a second-order method θj+1 = θj + ηdj; for training FCNNs since the shared weights of convolu- tional layers lead to the difficulties in factorizing its Hes- where η is the step size. sian. We defer tackling convolutional layers to our future For non-convex problems, the solution to Eq. (2) reflects work. We also focus on the classification problem and hence one of three possibilities: a local minimum θmin, a local do not consider auto-encoders. Our strategy is that once a maximum θmax or a saddle point θsaddle. Some previous simpler isolated problem can be effectively handled, we can works such as (Dauphin et al. 2014; Goldfarb et al. 2017; then extend our method to address more challenging issues. Mizutani and Dreyfus 2008) utilize negative curvature infor- In summary, the contributions of this paper are as follows: mation to converge to a local minimum. Before illustrating how these previous works tackle the issue of negative cur- 1. For curvature information, we propose the PCH matrix vature, we have to introduce a crucial concept that we can to improve the Gauss-Newton matrix for training FCNNs know the curvature information of f at a given point θ by with convex criterion functions and overcome the non- analyzing the Hessian matrix r2f(θ). On the one hand, the convex scenario. Hessian matrix of f at any θmin is positive semi-definite. 2. To derive the update directions, we devise effective EA- On the other hand, the Hessian matrices of f at any θmax CG method, which does converge faster in terms of wall and θsaddle are negative semi-definite and indefinite, respec- clock time and enjoys better testing accuracy than com- tively. After establishing the concept, we can use the follow- peting methods. Specially, the performance of EA-CG is ing property to understand how to utilize the negative curva- competitive with SGD. ture information to resolve the issue of negative curvature. Property 1. Let f be a non-convex and twice-differentiable Truncated-Newton Method on Non-Convex function. With a given point θj, we suppose that there ex- 2 j Problems ist some negative eigenvalues fλ1; : : : ; λsg for r f(θ ). Moreover, we take V = span(fv1;:::; vsg), which is the Newton’s method is one of the second-order minimization eigenspace corresponds to fλ1; : : : ; λsg. If we take methods, and is generally composed of two steps: 1) com- T 1 puting the Hessian matrix and 2) solving the system of lin- g(k) = f(θj) + rf(θj) v + vT r2f(θj)v; ear equations for update directions.
Load more

Recommended publications
  • Training Autoencoders by Alternating Minimization
    Under review as a conference paper at ICLR 2018 TRAINING AUTOENCODERS BY ALTERNATING MINI- MIZATION Anonymous authors Paper under double-blind review ABSTRACT We present DANTE, a novel method for training neural networks, in particular autoencoders, using the alternating minimization principle. DANTE provides a distinct perspective in lieu of traditional gradient-based backpropagation techniques commonly used to train deep networks. It utilizes an adaptation of quasi-convex optimization techniques to cast autoencoder training as a bi-quasi-convex optimiza- tion problem. We show that for autoencoder configurations with both differentiable (e.g. sigmoid) and non-differentiable (e.g. ReLU) activation functions, we can perform the alternations very effectively. DANTE effortlessly extends to networks with multiple hidden layers and varying network configurations. In experiments on standard datasets, autoencoders trained using the proposed method were found to be very promising and competitive to traditional backpropagation techniques, both in terms of quality of solution, as well as training speed. 1 INTRODUCTION For much of the recent march of deep learning, gradient-based backpropagation methods, e.g. Stochastic Gradient Descent (SGD) and its variants, have been the mainstay of practitioners. The use of these methods, especially on vast amounts of data, has led to unprecedented progress in several areas of artificial intelligence. On one hand, the intense focus on these techniques has led to an intimate understanding of hardware requirements and code optimizations needed to execute these routines on large datasets in a scalable manner. Today, myriad off-the-shelf and highly optimized packages exist that can churn reasonably large datasets on GPU architectures with relatively mild human involvement and little bootstrap effort.
    [Show full text]
  • Deep Learning Based Computer Generated Face Identification Using
    applied sciences Article Deep Learning Based Computer Generated Face Identification Using Convolutional Neural Network L. Minh Dang 1, Syed Ibrahim Hassan 1, Suhyeon Im 1, Jaecheol Lee 2, Sujin Lee 1 and Hyeonjoon Moon 1,* 1 Department of Computer Science and Engineering, Sejong University, Seoul 143-747, Korea; [email protected] (L.M.D.); [email protected] (S.I.H.); [email protected] (S.I.); [email protected] (S.L.) 2 Department of Information Communication Engineering, Sungkyul University, Seoul 143-747, Korea; [email protected] * Correspondence: [email protected] Received: 30 October 2018; Accepted: 10 December 2018; Published: 13 December 2018 Abstract: Generative adversarial networks (GANs) describe an emerging generative model which has made impressive progress in the last few years in generating photorealistic facial images. As the result, it has become more and more difficult to differentiate between computer-generated and real face images, even with the human’s eyes. If the generated images are used with the intent to mislead and deceive readers, it would probably cause severe ethical, moral, and legal issues. Moreover, it is challenging to collect a dataset for computer-generated face identification that is large enough for research purposes because the number of realistic computer-generated images is still limited and scattered on the internet. Thus, a development of a novel decision support system for analyzing and detecting computer-generated face images generated by the GAN network is crucial. In this paper, we propose a customized convolutional neural network, namely CGFace, which is specifically designed for the computer-generated face detection task by customizing the number of convolutional layers, so it performs well in detecting computer-generated face images.
    [Show full text]
  • Matrix Calculus
    Appendix D Matrix Calculus From too much study, and from extreme passion, cometh madnesse. Isaac Newton [205, §5] − D.1 Gradient, Directional derivative, Taylor series D.1.1 Gradients Gradient of a differentiable real function f(x) : RK R with respect to its vector argument is defined uniquely in terms of partial derivatives→ ∂f(x) ∂x1 ∂f(x) , ∂x2 RK f(x) . (2053) ∇ . ∈ . ∂f(x) ∂xK while the second-order gradient of the twice differentiable real function with respect to its vector argument is traditionally called the Hessian; 2 2 2 ∂ f(x) ∂ f(x) ∂ f(x) 2 ∂x1 ∂x1∂x2 ··· ∂x1∂xK 2 2 2 ∂ f(x) ∂ f(x) ∂ f(x) 2 2 K f(x) , ∂x2∂x1 ∂x2 ··· ∂x2∂xK S (2054) ∇ . ∈ . .. 2 2 2 ∂ f(x) ∂ f(x) ∂ f(x) 2 ∂xK ∂x1 ∂xK ∂x2 ∂x ··· K interpreted ∂f(x) ∂f(x) 2 ∂ ∂ 2 ∂ f(x) ∂x1 ∂x2 ∂ f(x) = = = (2055) ∂x1∂x2 ³∂x2 ´ ³∂x1 ´ ∂x2∂x1 Dattorro, Convex Optimization Euclidean Distance Geometry, Mεβoo, 2005, v2020.02.29. 599 600 APPENDIX D. MATRIX CALCULUS The gradient of vector-valued function v(x) : R RN on real domain is a row vector → v(x) , ∂v1(x) ∂v2(x) ∂vN (x) RN (2056) ∇ ∂x ∂x ··· ∂x ∈ h i while the second-order gradient is 2 2 2 2 , ∂ v1(x) ∂ v2(x) ∂ vN (x) RN v(x) 2 2 2 (2057) ∇ ∂x ∂x ··· ∂x ∈ h i Gradient of vector-valued function h(x) : RK RN on vector domain is → ∂h1(x) ∂h2(x) ∂hN (x) ∂x1 ∂x1 ··· ∂x1 ∂h1(x) ∂h2(x) ∂hN (x) h(x) , ∂x2 ∂x2 ··· ∂x2 ∇ .
    [Show full text]
  • Differentiability in Several Variables: Summary of Basic Concepts
    DIFFERENTIABILITY IN SEVERAL VARIABLES: SUMMARY OF BASIC CONCEPTS 3 3 1. Partial derivatives If f : R ! R is an arbitrary function and a = (x0; y0; z0) 2 R , then @f f(a + tj) ¡ f(a) f(x ; y + t; z ) ¡ f(x ; y ; z ) (1) (a) := lim = lim 0 0 0 0 0 0 @y t!0 t t!0 t etc.. provided the limit exists. 2 @f f(t;0)¡f(0;0) Example 1. for f : R ! R, @x (0; 0) = limt!0 t . Example 2. Let f : R2 ! R given by ( 2 x y ; (x; y) 6= (0; 0) f(x; y) = x2+y2 0; (x; y) = (0; 0) Then: @f f(t;0)¡f(0;0) 0 ² @x (0; 0) = limt!0 t = limt!0 t = 0 @f f(0;t)¡f(0;0) ² @y (0; 0) = limt!0 t = 0 Note: away from (0; 0), where f is the quotient of differentiable functions (with non-zero denominator) one can apply the usual rules of derivation: @f 2xy(x2 + y2) ¡ 2x3y (x; y) = ; for (x; y) 6= (0; 0) @x (x2 + y2)2 2 2 2. Directional derivatives. If f : R ! R is a map, a = (x0; y0) 2 R and v = ®i + ¯j is a vector in R2, then by definition f(a + tv) ¡ f(a) f(x0 + t®; y0 + t¯) ¡ f(x0; y0) (2) @vf(a) := lim = lim t!0 t t!0 t Example 3. Let f the function from the Example 2 above. Then for v = ®i + ¯j a unit vector (i.e.
    [Show full text]
  • Lipschitz Recurrent Neural Networks
    Published as a conference paper at ICLR 2021 LIPSCHITZ RECURRENT NEURAL NETWORKS N. Benjamin Erichson Omri Azencot Alejandro Queiruga ICSI and UC Berkeley Ben-Gurion University Google Research [email protected] [email protected] [email protected] Liam Hodgkinson Michael W. Mahoney ICSI and UC Berkeley ICSI and UC Berkeley [email protected] [email protected] ABSTRACT Viewing recurrent neural networks (RNNs) as continuous-time dynamical sys- tems, we propose a recurrent unit that describes the hidden state’s evolution with two parts: a well-understood linear component plus a Lipschitz nonlinearity. This particular functional form facilitates stability analysis of the long-term behavior of the recurrent unit using tools from nonlinear systems theory. In turn, this en- ables architectural design decisions before experimentation. Sufficient conditions for global stability of the recurrent unit are obtained, motivating a novel scheme for constructing hidden-to-hidden matrices. Our experiments demonstrate that the Lipschitz RNN can outperform existing recurrent units on a range of bench- mark tasks, including computer vision, language modeling and speech prediction tasks. Finally, through Hessian-based analysis we demonstrate that our Lipschitz recurrent unit is more robust with respect to input and parameter perturbations as compared to other continuous-time RNNs. 1 INTRODUCTION Many interesting problems exhibit temporal structures that can be modeled with recurrent neural networks (RNNs), including problems in robotics, system identification, natural language process- ing, and machine learning control. In contrast to feed-forward neural networks, RNNs consist of one or more recurrent units that are designed to have dynamical (recurrent) properties, thereby enabling them to acquire some form of internal memory.
    [Show full text]
  • Neural Networks
    School of Computer Science 10-601B Introduction to Machine Learning Neural Networks Readings: Matt Gormley Bishop Ch. 5 Lecture 15 Murphy Ch. 16.5, Ch. 28 October 19, 2016 Mitchell Ch. 4 1 Reminders 2 Outline • Logistic Regression (Recap) • Neural Networks • Backpropagation 3 RECALL: LOGISTIC REGRESSION 4 Using gradient ascent for linear Recall… classifiers Key idea behind today’s lecture: 1. Define a linear classifier (logistic regression) 2. Define an objective function (likelihood) 3. Optimize it with gradient descent to learn parameters 4. Predict the class with highest probability under the model 5 Using gradient ascent for linear Recall… classifiers This decision function isn’t Use a differentiable function differentiable: instead: 1 T pθ(y =1t)= h(t)=sign(θ t) | 1+2tT( θT t) − sign(x) 1 logistic(u) ≡ −u 1+ e 6 Using gradient ascent for linear Recall… classifiers This decision function isn’t Use a differentiable function differentiable: instead: 1 T pθ(y =1t)= h(t)=sign(θ t) | 1+2tT( θT t) − sign(x) 1 logistic(u) ≡ −u 1+ e 7 Recall… Logistic Regression Data: Inputs are continuous vectors of length K. Outputs are discrete. (i) (i) N K = t ,y where t R and y 0, 1 D { }i=1 ∈ ∈ { } Model: Logistic function applied to dot product of parameters with input vector. 1 pθ(y =1t)= | 1+2tT( θT t) − Learning: finds the parameters that minimize some objective function. θ∗ = argmin J(θ) θ Prediction: Output is the most probable class. yˆ = `;Kt pθ(y t) y 0,1 | ∈{ } 8 NEURAL NETWORKS 9 Learning highly non-linear functions f: X à Y l f might be non-linear
    [Show full text]
  • Jointly Clustering with K-Means and Learning Representations
    Deep k-Means: Jointly clustering with k-Means and learning representations Maziar Moradi Fard Univ. Grenoble Alpes, CNRS, Grenoble INP – LIG [email protected] Thibaut Thonet Univ. Grenoble Alpes, CNRS, Grenoble INP – LIG [email protected] Eric Gaussier Univ. Grenoble Alpes, CNRS, Grenoble INP – LIG, Skopai [email protected] Abstract We study in this paper the problem of jointly clustering and learning representations. As several previous studies have shown, learning representations that are both faithful to the data to be clustered and adapted to the clustering algorithm can lead to better clustering performance, all the more so that the two tasks are performed jointly. We propose here such an approach for k-Means clustering based on a continuous reparametrization of the objective function that leads to a truly joint solution. The behavior of our approach is illustrated on various datasets showing its efficacy in learning representations for objects while clustering them. 1 Introduction Clustering is a long-standing problem in the machine learning and data mining fields, and thus accordingly fostered abundant research. Traditional clustering methods, e.g., k-Means [23] and Gaussian Mixture Models (GMMs) [6], fully rely on the original data representations and may then be ineffective when the data points (e.g., images and text documents) live in a high-dimensional space – a problem commonly known as the curse of dimensionality. Significant progress has been arXiv:1806.10069v2 [cs.LG] 12 Dec 2018 made in the last decade or so to learn better, low-dimensional data representations [13].
    [Show full text]
  • Differentiability of the Convolution
    Differentiability of the convolution. Pablo Jimenez´ Rodr´ıguez Universidad de Valladolid March, 8th, 2019 Pablo Jim´enezRodr´ıguez Differentiability of the convolution Definition Let L1[−1; 1] be the set of the 2-periodic functions f so that R 1 −1 jf (t)j dt < 1: Given f ; g 2 L1[−1; 1], the convolution of f and g is the function Z 1 f ∗ g(x) = f (t)g(x − t) dt −1 Pablo Jim´enezRodr´ıguez Differentiability of the convolution Definition Let L1[−1; 1] be the set of the 2-periodic functions f so that R 1 −1 jf (t)j dt < 1: Given f ; g 2 L1[−1; 1], the convolution of f and g is the function Z 1 f ∗ g(x) = f (t)g(x − t) dt −1 Pablo Jim´enezRodr´ıguez Differentiability of the convolution Definition Let L1[−1; 1] be the set of the 2-periodic functions f so that R 1 −1 jf (t)j dt < 1: Given f ; g 2 L1[−1; 1], the convolution of f and g is the function Z 1 f ∗ g(x) = f (t)g(x − t) dt −1 Pablo Jim´enezRodr´ıguez Differentiability of the convolution Definition Let L1[−1; 1] be the set of the 2-periodic functions f so that R 1 −1 jf (t)j dt < 1: Given f ; g 2 L1[−1; 1], the convolution of f and g is the function Z 1 f ∗ g(x) = f (t)g(x − t) dt −1 Pablo Jim´enezRodr´ıguez Differentiability of the convolution If f is Lipschitz, then f ∗ g is Lipschitz.
    [Show full text]
  • The Simple Essence of Automatic Differentiation
    The Simple Essence of Automatic Differentiation CONAL ELLIOTT, Target, USA Automatic differentiation (AD) in reverse mode (RAD) is a central component of deep learning and other uses of large-scale optimization. Commonly used RAD algorithms such as backpropagation, however, are complex and stateful, hindering deep understanding, improvement, and parallel execution. This paper develops a simple, generalized AD algorithm calculated from a simple, natural specification. The general algorithm is then specialized by varying the representation of derivatives. In particular, applying well-known constructions to a naive representation yields two RAD algorithms that are far simpler than previously known. In contrast to commonly used RAD implementations, the algorithms defined here involve no graphs, tapes, variables, partial derivatives, or mutation. They are inherently parallel-friendly, correct by construction, and usable directly from an existing programming language with no need for new data types or programming style, thanks to use of an AD-agnostic compiler plugin. CCS Concepts: · Mathematics of computing → Differential calculus; · Theory of computation → Program reasoning; Program specifications; Additional Key Words and Phrases: automatic differentiation, program calculation, category theory ACM Reference Format: Conal Elliott. 2018. The Simple Essence of Automatic Differentiation. Proc. ACM Program. Lang. 2, ICFP, Article 70 (September 2018), 29 pages. https://doi.org/10.1145/3236765 1 INTRODUCTION Accurate, efficient, and reliable computation of derivatives has become increasingly important over the last several years, thanks in large part to the successful use of backpropagation in machine learning, including multi-layer neural networks, also known as łdeep learningž [Goodfellow et al. 2016; Lecun et al. 2015]. Backpropagation is a specialization and independent invention of the reverse mode of automatic differentiation (AD) and is used to tune a parametric model to closely match observed data, using gradient descent (or stochastic gradient descent).
    [Show full text]
  • Expressivity of Deep Neural Networks∗
    Expressivity of Deep Neural Networks∗ Ingo G¨uhring yz Mones Raslan yz Gitta Kutyniokzx 10/07/2020 Abstract In this review paper, we give a comprehensive overview of the large variety of approximation results for neural networks. Approximation rates for classical function spaces as well as benefits of deep neural networks over shallow ones for specifically structured function classes are discussed. While the main body of existing results is for general feedforward architectures, we also review approximation results for convolutional, residual and recurrent neural networks. Keywords: approximation, expressivity, deep vs. shallow, function classes 1 Introduction While many aspects of the success of deep learning still lack a comprehensive mathematical explanation, the approximation properties1 of neural networks have been studied since around 1960 and are relatively well understood. Statistical learning theory formalizes the problem of approximating - in this context also called learning - a function from a finite set of samples. Next to statistical and algorithmic considerations, approximation theory plays a major role for the analysis of statistical learning problems. We will clarify this in the following by introducing some fundamental notions.2 Assume that X is an input space and Y a target space, L : Y × Y ! [0; 1] is a loss function and P(X ;Y) a (usually unknown) probability distribution on some σ-algebra of X × Y. We then aim at finding a minimizer of the risk functional3 Z X R : Y ! [0; 1]; f 7! L (f(x); y) dP(X ;Y)(x; y); X ×Y X induced by L and P(X ;Y) (where Y denotes the set of all functions from X to Y).
    [Show full text]
  • Differentiable Programming for Piecewise Polynomial
    Under review as a conference paper at ICLR 2021 DIFFERENTIABLE PROGRAMMING FOR PIECEWISE POLYNOMIAL FUNCTIONS Anonymous authors Paper under double-blind review ABSTRACT The paradigm of differentiable programming has significantly enhanced the scope of machine learning via the judicious use of gradient-based optimization. How- ever, standard differentiable programming methods (such as autodiff) typically require that the machine learning models be differentiable, limiting their applica- bility. We introduce a new, principled approach to extend gradient-based optimiza- tion to piecewise smooth models, such as k-histograms, splines, and segmentation maps. We derive an accurate form of the weak Jacobian of such functions and show that it exhibits a block-sparse structure that can be computed implicitly and efficiently. We show that using the redesigned Jacobian leads to improved per- formance in applications such as denoising with piecewise polynomial regression models, data-free generative model training, and image segmentation. 1 INTRODUCTION Motivation: Differentiable programming has been proposed as a paradigm shift in algorithm design. The main idea is to leverage gradient-based optimization to optimize the parameters of the algo- rithm, allowing for end-to-end trainable systems (such as deep neural networks) to exploit structure and patterns in data and achieve better performance. This approach has found use in a large va- riety of applications such as scientific computing (Innes, 2020; Innes et al., 2019; Schafer et al., 2020), image processing (Li et al., 2018b), physics engines (Degrave et al., 2017), computational simulations (Alnæs et al., 2015), and graphics (Li et al., 2018a; Chen et al., 2019). One way to leverage differential programming modules is to encode additional structural priors as “layers” in a larger machine learning model.
    [Show full text]
  • GAN and What It Is?
    Basics: Why GAN and what it is? The Optimization Problem of GANs How to Train GANs: An Odyssey How to Evaluate GANs? An Application Tour of GANs Summary and Open Challenges Outline Basics: Why GAN and what it is? Generative vs. Discriminative Models • Given a distribution of inputs X and labels Y • Discriminative: model the conditional distribution P(Y | X). • Generative networks model the joint distribution P(X , Y). • If the model understands the joint distribution P(X ,Y), then it • can calculate P(Y | X) using Bayes rule • can perform other tasks like P(X | Y), i.e., generating data from the label (called “conditional generation” in GANs) • understands the distribution better than a discriminative model, a scientific philosophy called “analysis by synthesis” “The analysis of the patterns generated by the world in any modality, with all their naturally occurring complexity and ambiguity, with the goal of reconstructing the processes, objects and events that produced them” - David Mumford Generative vs. Discriminative Models • Even if you only have X, you can still build a generative model • … making generative modeling amendable to unsupervised/semi-supervised representation learning • Not every problem is discriminative, but all problems could be generative! • However, generative modeling is harder! • Map from X to Y is typically many to one • Map from Y to X is typically one to many • Dimensionality of X typically >> dimensionality of Y • Hence compared to estimating P(Y | X), the estimation of P(X, Y) usually runs into the “curse of dimensionality” Setup of Generative Models Autoregressive Models RNNs for Autoregressive Language Modeling PixelRNN (van der Oord et al.
    [Show full text]