### 计算机代写|机器学习代写machine learning代考|COMP30027

statistics-lab™ 为您的留学生涯保驾护航 在代写机器学习 machine learning方面已经树立了自己的口碑, 保证靠谱, 高质且原创的统计Statistics代写服务。我们的专家在代写机器学习 machine learning代写方面经验极为丰富，各种代写机器学习 machine learning相关的作业也就用不着说。

• Statistical Inference 统计推断
• Statistical Computing 统计计算
• (Generalized) Linear Models 广义线性模型
• Statistical Machine Learning 统计机器学习
• Longitudinal Data Analysis 纵向数据分析
• Foundations of Data Science 数据科学基础

## 计算机代写|机器学习代写machine learning代考|Random Neural Networks

Although much less popular than modern deep neural networks, neural networks with random fixed weights are simpler to analyze. Such networks have frequently arisen in the past decades as an appropriate solution to handle the possibly restricted number of training data, to reduce the computational and memory complexity and, from another viewpoint, can be seen as efficient random feature extractors. These neural networks in fact find their roots in Rosenblatt’s perceptron [Rosenblatt, 1958] and have then been many times revisited, rediscovered, and analyzed in a number of works, both in their feedforward [Schmidt et al., 1992] and recurrent [Gelenbe, 1993] versions. The simplest modern versions of these random networks are the so-called extreme learning machine [Huang et al., 2012] for the feedforward case, which one may seem as a mere linear regression method on nonlinear random features, and the echo state network [Jaeger, 2001] for the recurrent case. Also see Scardapane and Wang [2017] for a more exhaustive overview of randomness in neural networks.

It is also to be noted that deep neural networks are initialized at random and that random operations (such as random node deletions or voluntarily not-learning a large proportion of randomly initialized neural network weights, that is, random dropout) are common and efficient in neural network learning [Srivastava et al., 2014, Frankle and Carbin, 2019]. We may also point the recent endeavor toward neural network “learning without backpropagation,” which, inspired by biological neural networks (which naturally do not operate backpropagation learning), proposes learning mechanisms with fixed random backward weights and asymmetric forward learning procedures [Lillicrap et al., 2016, Nøkland, 2016, Baldi et al., 2018, Frenkel et al., 2019, Han et al., 2019]. As such, the study of random neural network structures may be instrumental to future improved understanding and designs of advanced neural network structures.

As shall be seen subsequently, the simple models of random neural networks are to a large extent connected to kernel matrices. More specifically, the classification or regression performance at the output of these random neural networks are functionals of random matrices that fall into the wide class of kernel random matrices, yet of a slightly different form than those studied in Section 4. Perhaps more surprisingly, this connection still exists for deep neural networks which are (i) randomly initialized and (ii) then trained with gradient descent, via the so-called neural tangent kernel [Jacot et al., 2018] by considering the “infinitely many neurons” limit, that is, the limit where the network widths of all layers go to infinity simultaneously. This close connection between neural networks and kernels has triggered a renewed interest for the theoretical investigation of deep neural networks from various perspectives including optimization [Du et al., 2019, Chizat et al., 2019], generalization [Allen-Zhu et al., 2019, Arora et al., 2019a, Bietti and Mairal, 2019], and learning dynamics [Lee et al., 2020, Advani et al., 2020, Liao and Couillet, 2018a]. These works shed new light on our theoretical understanding of deep neural network models and specifically demonstrate the significance of studying simple networks with random weights and their associated kernels to assess the intrinsic mechanisms of more elaborate and practical deep networks.

## 计算机代写|机器学习代写machine learning代考|Regression with Random Neural Networks

Throughout this section, we consider a feedforward single-hidden-layer neural network, as illustrated in Figure $5.1$ (displayed, for notational convenience, from right to left). A similar class of single-hidden-layer neural network models, however with a recurrent structure, will be discussed later in Section 5.3.

Given input data $\mathbf{X}=\left[\mathbf{x}_1, \ldots, \mathbf{x}_n\right] \in \mathbb{R}^{p \times n}$, we denote $\Sigma \equiv \sigma(\mathbf{W} \mathbf{X}) \in \mathbb{R}^{N \times n}$ the output of the first layer comprising $N$ neurons. This output arises from the premultiplication of $\mathbf{X}$ by some random weight matrix $\mathbf{W} \in \mathbb{R}^{N \times p}$ with i.i.d. (say standard Gaussian) entries and the entry-wise application of the nonlinear activation function $\sigma: \mathbb{R} \rightarrow \mathbb{R}$. As such, the columns $\sigma\left(\mathbf{W x}_i\right)$ of $\Sigma$ can be seen as random nonlinear features of $\mathbf{x}_i$. The second layer weight $\boldsymbol{\beta} \in \mathbb{R}^{N \times d}$ is then learned to adapt the feature matrix $\Sigma$ to some associated target $\mathbf{Y}=\left[\mathbf{y}_1, \ldots, \mathbf{y}_n\right] \in \mathbb{R}^{d \times n}$, for instance, by minimizing the Frobenius norm $\left|\mathbf{Y}-\boldsymbol{\beta}^{\top} \Sigma\right|_F^2$.

Remark 5.1 (Random neural networks, random feature maps and random kernels). The columns of $\Sigma$ may be seen as the output of the $\mathbb{R}^p \rightarrow \mathbb{R}^N$ random feature map $\phi: \mathbf{x}i \mapsto \sigma\left(\mathbf{W} \mathbf{x}_i\right)$ for some given $\mathbf{W} \in \mathbb{R}^{N \times p}$. In Rahimi and Recht [2008], it is shown that, for every nonnegative definite “shift-invariant” kernel of the form $(\mathbf{x}, \mathbf{y}) \mapsto f\left(|\mathbf{x}-\mathbf{y}|^2\right)$, there exist appropriate choices for $\sigma$ and the law of the entries of $\mathbf{W}$ so that as the number of neurons or random features $N \rightarrow \infty$, $$\sigma\left(\mathbf{W} \mathbf{x}_i\right)^{\top} \sigma\left(\mathbf{W} \mathbf{x}_j\right) \stackrel{\text { a.s. }}{\longrightarrow} f\left(\left|\mathbf{x}_i-\mathbf{x}_j\right|^2\right) .$$ As such, for large enough $N$ (that in general must scale with $n, p$ ), the bivariate function $(\mathbf{x}, \mathbf{y}) \mapsto \sigma(\mathbf{W} \mathbf{x})^{\top} \sigma(\mathbf{W y})$ approximates a kernel function of the type $f\left(|\mathbf{x}-\mathbf{y}|^2\right)$ studied in Chapter 4. This result is then generalized, in subsequent works, to a larger family of kernels including inner-product kernels [Kar and Karnick, 2012], additive homogeneous kernels [Vedaldi and Zisserman, 2012], etc. Another, possibly more marginal, connection with the previous sections is that $\sigma\left(\mathbf{w}^{\top} \mathbf{x}\right)$ can be interpreted as a “properly scaling” inner-product kernel function applied to the “data” pair $\mathbf{w}, \mathbf{x} \in \mathbb{R}^p$. This technically induces another strong relation between the study of kernels and that of neural networks. Again, similar to the concentration of (Euclidean) distance extensively explored in this chapter, the entry-wise convergence in (5.1) does not imply convergence in the operator norm sense, which, as we shall see, leads directly to the so-called “double descent” test curve in random feature/neural network models. If the network output weight matrix $\boldsymbol{\beta}$ is designed to minimize the regularized MSE $L(\boldsymbol{\beta})=\frac{1}{n} \sum{i=1}^n\left|\mathbf{y}_i-\boldsymbol{\beta}^{\top} \sigma\left(\mathbf{W x}_i\right)\right|^2+\gamma|\boldsymbol{\beta}|_F^2$, for some regularization parameter $\gamma>0$, then $\beta$ takes the explicit form of a ridge-regressor ${ }^1$
$$\beta \equiv \frac{1}{n} \Sigma\left(\frac{1}{n} \Sigma^{\top} \Sigma+\gamma \mathbf{I}_n\right)^{-1} \mathbf{Y}^{\top},$$
which follows from differentiating $L(\boldsymbol{\beta})$ with respect to $\boldsymbol{\beta}$ to obtain $0=\gamma \boldsymbol{\beta}+$ $\frac{1}{n} \Sigma\left(\Sigma^{\top} \boldsymbol{\beta}-\mathbf{Y}^{\top}\right)$ so that $\left(\frac{1}{n} \Sigma \Sigma^{\top}+\gamma \mathbf{I}_N\right) \boldsymbol{\beta}=\frac{1}{n} \Sigma \mathbf{Y}^{\top}$ which, along with $\left(\frac{1}{n} \Sigma \Sigma^{\top}+\right.$ $\left.\gamma \mathbf{I}_N\right)^{-1} \Sigma=\Sigma\left(\frac{1}{n} \Sigma^{\top} \Sigma+\gamma \mathbf{I}_n\right)^{-1}$ for $\gamma>0$, gives the result.

# 机器学习代考

## 计算机代写|机器学习代写machine learning代考|Regression with Random Neural Networks

$$\beta \equiv \frac{1}{n} \Sigma\left(\frac{1}{n} \Sigma^{\top} \Sigma+\gamma \mathbf{I}_n\right)^{-1} \mathbf{Y}^{\top},$$

## 有限元方法代写

tatistics-lab作为专业的留学生服务机构，多年来已为美国、英国、加拿大、澳洲等留学热门地的学生提供专业的学术服务，包括但不限于Essay代写，Assignment代写，Dissertation代写，Report代写，小组作业代写，Proposal代写，Paper代写，Presentation代写，计算机作业代写，论文修改和润色，网课代做，exam代考等等。写作范围涵盖高中，本科，研究生等海外留学全阶段，辐射金融，经济学，会计学，审计学，管理学等全球99%专业科目。写作团队既有专业英语母语作者，也有海外名校硕博留学生，每位写作老师都拥有过硬的语言能力，专业的学科背景和学术写作经验。我们承诺100%原创，100%专业，100%准时，100%满意。

## MATLAB代写

MATLAB 是一种用于技术计算的高性能语言。它将计算、可视化和编程集成在一个易于使用的环境中，其中问题和解决方案以熟悉的数学符号表示。典型用途包括：数学和计算算法开发建模、仿真和原型制作数据分析、探索和可视化科学和工程图形应用程序开发，包括图形用户界面构建MATLAB 是一个交互式系统，其基本数据元素是一个不需要维度的数组。这使您可以解决许多技术计算问题，尤其是那些具有矩阵和向量公式的问题，而只需用 C 或 Fortran 等标量非交互式语言编写程序所需的时间的一小部分。MATLAB 名称代表矩阵实验室。MATLAB 最初的编写目的是提供对由 LINPACK 和 EISPACK 项目开发的矩阵软件的轻松访问，这两个项目共同代表了矩阵计算软件的最新技术。MATLAB 经过多年的发展，得到了许多用户的投入。在大学环境中，它是数学、工程和科学入门和高级课程的标准教学工具。在工业领域，MATLAB 是高效研究、开发和分析的首选工具。MATLAB 具有一系列称为工具箱的特定于应用程序的解决方案。对于大多数 MATLAB 用户来说非常重要，工具箱允许您学习应用专业技术。工具箱是 MATLAB 函数（M 文件）的综合集合，可扩展 MATLAB 环境以解决特定类别的问题。可用工具箱的领域包括信号处理、控制系统、神经网络、模糊逻辑、小波、仿真等。