The video of machine learning to learn [update]

I. Overview of Machine Learning

1. What is machine learning

Machine learning is a computer program that learns from experience and gets better at a specific task.
Most important for machine learning:

• Data: Experience is ultimately transformed into data that the computer can understand, so that the computer can learn from the experience.
• Model: The algorithm.Once you have the data, you can design a model and use the data as input to train the model.The trained model eventually becomes the core of machine learning, making the model the hub that can generate decisions.

2. Supervised vs Unsupervised Learning

(1) Supervised Learning

Supervised learning (Supervised learning) allows the computer to learn the rules from a large amount of known input and output paired data, so that it can make a reasonable output prediction for a new input.

• House price prediction (regression problem)

Shown below is an example of a house price forecast.It is an example of supervised learning.

As in this housing price prediction example, supervised learning provides the algorithm with a data set (which contains the correct answers), that is, we give it a housing price data set, and each sample in this data set corresponds to a correct answer(i.e. the actual selling price of the house).The purpose of the algorithm is to give more correct answers.

The prediction of house price is a regression problem (regression), because house price is a real, continuous value.

• Malignant/benign cancer (classification problem)
  
  Malignant/benign cancer (ie 0/1) is a classification problem because malignant/benign (ie 0/1) are discrete values.

(2) Unsupervised Learning

Unsupervised learning (Unsupervised learning) analyzes the inherent characteristics and structure of the data itself by learning a large amount of unlabeled data.
Unsupervised learning requires letting the algorithm discover everything from the data on its own.One of the common algorithms is clustering: using algorithms to group together news stories, market segmentation, etc.

3. Model description

Linear Regression Model

• Cost function
  
  
  Squared error cost function
  
  
  Model, parameters, cost function, objective (find a value of w, b that minimizes J(w,b))
  
  when b=0
  By simplifying the model, our goal is to find wA value that minimizes J(w)
  When w=1, J(w)=0 is calculated:
  When w=0.5, J(w)=0.58 is calculated:
  
  Take different values ​​of w, calculate J(w), and draw a graph as shown:
  
  when b!=0
  
  House price forecast
  

4. Gradient descent algorithm

where a is the learning rate.If the learning rate is too small, then large drops are feasible, but will be slow.This will take a long time because you will be taking very small steps, many steps before it gets close to the minimum.But if the learning rate is too large, the steps will be very large, and it is likely that the minimum value will be skipped due to the large step size, and the minimum value will never be reached.
As shown in the picture:

If the parameter is reachedlocal minima, then further binning reduces steps to none at all.It doesn't change the process precisely because it keeps the solution at a local minimum.

As we approach a local minimum, the rank drop will beSmaller steps are automatically taken because the derivative automatically gets smaller as we approach a local minimum, which means that the steps also get smaller automatically, even if a remains at some fixed value.


Derivation: