CNN 学习步骤总结--CNN 0

个人笔记，可以忽略这篇文章

0. Convolutional Neural Networks

这一节要对对神经网络有一个大致的把握，最重要的是了解每一层的意义，为什么一层比一层更抽象？是怎么提取更大的特征的。

图像位置大小等这些是怎么解决的？

带着问题探究怎么去优化每个神经元的 weight 以及 bias.

接着探究梯度下降等数学知识，所以学到了积分。最后发现学的积分只能针对单一变量，那就要学习多变量积分。

Convolutional Neural Networks–CNN 1

Tensors in Neural Networks–CNN 2

But what is a neural network?

Numbers and Linear Functions–CNN 3

1. Differential, Derivative and Calculus

Differential, Derivative and Calculus–CNN 4

The Chain rule

The essence of calculus

2. Linear Regression

Linear Regression

3. Gradient Descent，Cost Function

Gradient Descent(cost function)–CNN 5

Gradient Descent

4. Gradient descent, how neural networks learn

Multivariable functions

Gradient

Gradient and graphs 这个必须看

Gradient descent, how neural networks learn

---

Function Notation

---

Recall the notation that R stands for the real numbers. Similarly, $R^2$ is a two-dimensional vector, and $R^3$ is a three-dimensional vector.

Scalar-valued functions

In one-variable calculus, you worked a lot with one-variable functions, i.e., functions from R onto R. If $f(x)$ is such a one-variable functions, we can write $f:R\to R$ as a shorthand way of expressing that $f$ is a function from R onto R.

A function like $f(x,y)=x+y$ is a function of two variables. It takes an element of $R^2$, like (2,1), and gives a value that is a real number (i.e., an element of $R$), like $f(2,1)=3$. Since f maps $R^2$ to $R$, we write $f:R^2\to R$.

Vector-valued functions

A vector-valued function in three dimensions can be written $f:R\to R3$. For example, if $f(x)=(1-x,x^3,cosx)$, then $f(0)=(1,0,1)$. We sometimes write vector-valued functions using the standard unit vector i, j, and k, as in $f(x)=(1-x)i+x^{3}j+(cosx)k$.

---

Find More Information:

https://mathinsight.org/function_notation