Application of Machine Learning
- Image Identification
- Speech Recognition
- Medical diagnosis
- Statistical arbitrage
- Association learning
- classification
- prediction
- extract
- return
- Probability
Image Identification
- An object to be sorted is a digitized image, which measured value can describe the output of each pixel.
1. black and white image: a pixel saturation as a measured value. If a black and white image has N × N pixels, the number of measured values is N² .
2. Color Image: providing . 3 measurements, corresponding to the saturation of the RGB. For the N × N pixel image, only the measurement 3N² months. - Face detection (recognition)
classification may be present or absent, or a face pattern, each person in the database or as a separate classification. - Character recognition
to some notes into smaller images, each containing a single character. - Industrial component recognition
is determined whether the part is defective, whether certain criteria.
Speech recognition
measurement may be a value set represents a speech signal, the signal is typically divided into "segments" comprise words or phonemes is different in each segment, the speech signal may represent a variety of ways.
Signal may be made at different times - represents a frequency band intensity or energy. For example, some countries use speech recognition technology in the aircraft, and successfully applied to each program. The setting radio frequencies, providing commands to the autopilot, the aircraft designated turning point and the determined control parameters and weapon launch monitor.
Medical diagnosis
- After accurate identification to determine whether a disease is present.
- Each disease under consideration is a separate classification, does not suffer from any disease is also a separate classification.
- Machine learning can improve the accuracy of medical diagnosis by analyzing patient data.
- This application is the result of measurement is typically of medical testing (blood pressure, temperature, and various blood tests), and / significant degree and the basic information of the patient (age, sex, weight, etc.) of the emergence of symptoms / does not appear. In the basis of these measurements, the physician to determine the impact of the patient's disease.
Statistical arbitrage
- 一般是指短期、涉及大量证券的自动化交易策略。
- 根据多种证券的历史相关性、最近一段时间的价格变动和常规经济金融变量为一组证券设计交易算法，将这些作为观测值，预测可以转化为分类或者估算问题，使用预期回报的估算做出交易决策（买入、卖出等）。
关联学习
- 关联学习是扩展对产品之间各种联系的认识的过程，由机器进行的产品间的关系学习。
- 例如，经过对客户购买行为的分析，看上去毫不相干的产品可能展现出某种相互关联，就像沃尔玛超市的啤酒和尿布。
- 这体现了大数据的一个转变，考虑的不再是因果关系，而是相关关系。
- 我们在研究大量销售数据找出某种关联时可以开发一个规则，衍生出一个概率检验，从而学习某种条件概率，使用公式p（d/h），p是概率，d是根据h调整的产品，h是客户已经购买的一个或者一组产品。
分类
- 分类是根据个体特征（标识为自变量），将所研究群体中的每个个体纳入一组指定类别中的过程。
- 使用对象的观测值确定所属类别。例如银行的借贷和放贷问题。
- 将元素分类为不同类别的过程称为分类问题。将与元素有关的数据作为输入，而分类程序的任务是为输入指定一个类，然后生成输出。
预测
银行借贷中，预测还款风险成为一项重要的应用。
提取
信息提取（IE）是自动地从非结构化数据中提取结构化信息的过程，它产生的输出通常在关系数据库中维护。提取过程以一组文档为输入生成结构化数据，输出采用摘要形式，如Excel工作表和关系数据库中的表。
回归
使用一个模型表达各个参数之间的关系y＝g（x）。
在函数需要优化的时候，我们可以在这些输入的一系列不同设置下获得观测值，从而找到所需的设置，拟合一个回归模型，然后继续修改输入以得到更好的模型，这一过程被称作响应面设计。
Probability
In most applications, classified, not by the value of the feature vector uniquely determined. Depending on the application and the specific observations, the characteristic value may be noisy, i.e. the observed feature values may have some inherent uncertainty or randomness.
