Machine Learning Introduction

This is briefing from wiki machine learning for easily understanding

Machine Learning

Machine Learning(ML) is the study algorithm and statistic model that computer system uses to improve the performance on a specific task. The ML algorithm uses data model of training data to make prediction or decision without explicitly programming on the specific task.

Overview

ML is classified as 3 major type:
1.Supervised Learning
The algorithm builds of data model from both input data and out put data.
e.g. for a image identification data training set, it includes the input, which is the picture with target image, and the output, which is label to indicate the picture is belong to classification or not.
2.Unsupervised Learning
The algorithm builds of data model from input data only without desired output data. Therefore it is used for analysis of data structure like grouping or clustering.
3.Active Learning
The algorithm accesses the desired output with limited input
4.Reinforcement Learning
The algorithms are given feedback of form of positive or negative reinforcement in a dynamic environment like unanimous driving or gaming against human opponents.

Theory

The object of leaner is to generalize from the experience so as to perform the new task or data prediction more accurately. Because the training data set is infinite but the future is uncertain, the algorithm cannot guarantee the performance. Therefore bias-variance-decomposition is a way to qualify the generalization error.

Model

1.Artificial Neural Networks

The Neural Network itself is not a algorithm, but a framework to implement the different algorithm. The neuron that receives signals from previous layer will process the it by its algorithm (non-linear) then pass the result to the connected neuron in the next layer. Normally the signals are real numbers for computation. The connection between neuron is called edge, the edge and neuron have wight that will be adjusted during the data training procedure. Deep learning normally consists of multiple hidden layers.

Support Vector Machines
SVM are sets of related supervised learning methods used for classification and regression. Although it is a linear model, it can perform a non-linear classification by using kernel trick, which is implicitly mapping their inputs into high-dimensional feature spaces.

Bayesian networks

The theory is an acyclic graphic model is probabilistic graphic model that represents a set of random value. For example, Bayesian Network can represents the symptoms and disease.

Limitation

Effective machine learning is difficult due to finding pattern wrongly with limited data sets. Like lacking data, data bias, or even wrong algorithm.

Bias

Machine learning approach can easily suffer from different data biases, for example, health care measurement error or a group of ethnic customer with bias already.