Today, my family can't stand the days when I don't study at home, so I have invested a lot of money. Dad said he would accompany me to study. As long as I studied, he would accompany me as long as possible. Hey, I still feel distressed when I see my father reading a book. .
Today I watched the first part of Wu Enda ’s machine learning, and wrote about the things I still remember, one is to review, and the other is to fear that I forgot the knowledge.
I have watched Lao Wu ’s video class before, and I always think it is in English. It may be too esoteric, I ca n’t understand it. After reading the first 8 chapters of the watermelon book, I looked back at the old Wu video and suddenly found such a simple and interesting person. It ’s an esoteric, bad old man ...
Honestly, Wu Wu is very detailed and very basic (just referring to the first two chapters, because I only read these ...), even the most basic differential will be patiently explained, it ’s Daniel ’s right .
The summary part is as follows:
supervised learning: giving machine samples and conclusions, the process of making machine learning an algorithm that can adapt to these samples and conclusions is supervised learning.
Unsupervised learning: Only give the machine a large number of samples, the machine can learn anything, and see what it can learn.
Regression: A type of supervised learning, which is equivalent to I want to get some continuous results, or that there is a functional relationship between the conclusion and the input. We call it regression.
Classification: a kind of supervised learning, the feeling is to find discrete solutions. If the result I want is a few discrete results, such as judging yes or no or judging cauliflower, watermelon, tomato scrambled eggs (I have recently made delicious )Wait.
Linear regression: I personally feel that the essence is a linear equation: y = ax + b;
there are two parts that need to be solved: one part is a, b is known, x is a variable, and y is solved. This equation is solved because it is equivalent to us It is assumed that a function can be applied to the actual problem we originally wanted to solve, so this function is called a hypothetical function.
The other is j (a, b) = 1 / 2m * sum (y (a, b) -y) ^ 2, because here we are solving for the existence of a certain a, b so that the value of j is the smallest, that is The so-called cost function. In linear regression, the principle of the change or iteration of the cost function is the gradient descent principle: the
so-called gradient descent principle is that because there are two unknowns, we always need a method to change when we need to change a and b. The core of the method is of course Changing a and b can make j decrease. Then we can move j downward from the gradient direction, which is a method reflected in the mathematical expression:
variable (a, b) = variable (a, b)-α times (differentiation of j to the variable). note: Here α refers to the size of the moving steps, which is also called the gradient descent function.
The idea is also very simple, that is, the variable a = a-the moving value of the gradient is multiplied by the weight; b = b-the moving value of the gradient is multiplied by the weight.
In this way, we can make ab constantly change to find the global optimal solution (j is the minimum value in the entire real number field). When j = 0 or j1 = j2, the iteration ends. (Of course, we need to give ab an initial value before the iteration, which is generally 0)
The above is the learning experience of the linear regression part. I have to watch 20 lessons tomorrow.
