Chapter 1 Overview of Machine Learning
EDITORIAL
reference book
"Machine learning practical - based Scikit-Learn and TensorFlow"
tool
python3.5.1，Jupyter Notebook, Pycharm
Excerpt part of after-school title
How would you define the machine learning?
Machine learning is a learning system that allows data from computer science.
The two most common task is supervised regression and classification.
Common tasks include unsupervised clustering, visualization, association rule learning and dimensionality reduction.
Online learning system can learn incrementally, the opposite of batch learning system. This makes it possible to quickly adapt to changing data and automation systems, and can be trained on large amounts of data.
Nuclear algorithm can handle large amounts of data outside the computer's main memory can not cope. It is divided into small quantities of data, then use the online learning techniques to learn from these in small quantities.
What is the difference between ultra-parameter model parameters learning algorithm?
Model has one or more parameters that determine the model of the new given instance will make a prediction (for example, the slope of the linear model). Learning algorithm attempts to find the optimal values of the parameters so that the model can be a good generalization to the new instance. Parameters Parameter ultra-learning algorithm itself, is not a model (for example, to apply regularization number).
What model-based learning algorithm searches are? What strategies are most commonly used? How do they make a prediction?
Model-based learning algorithm to search the optimal model generalization of model parameter values. Usually by minimizing the cost function to train such a system, the cost function is a measure of how bad the system predicts the training data, if the model has regularization, plus a penalty on the complexity of the model. The last parameter learning algorithm to find the median prediction function is finally obtained, simply examples of features available to the prediction function to predict.
Some of the major challenges faced by machine learning are: lack of data, poor data quality, data are not representative, non-characteristic information, the model is too simple lack of training data fitting, and the model is too complex to over-fitting the training data.
If your model perform well on the training data, but applied to the generalization of the results of the new instance is very bad, how is this going? We can put forward three possible solution?
This model is likely to over-fit the training data (or good luck in the training data).
Solutions are possible: for more data, the simplified model (selected simpler algorithm which reduces the number of parameters or features used, regularization model), or to reduce the noise in the training data.
What is the purpose of the validation set?
Validation set is used to compare different models. It can be used to select the best model and parameter adjustment over.
If you use the test set parameters to adjust over what will happen?
If you use the test set parameters to adjust over, there will be the risk of over-fitting set of tests, the final measure of generalization error would be too optimistic (last start of model performance is worse than expected).
What is cross-validation? Why is it better than the validation set?
By cross validation techniques, you may not require a separate validation set to achieve more models (for model selection and adjustment of parameters over). Who saves precious training data.
My CSDN: https://blog.csdn.net/qq_21579045
My blog garden: https://www.cnblogs.com/lyjun/
My Github: https://github.com/TinyHandsome
Paper come Zhongjue know this practice is essential ~
Welcome to come OB ~
by Li Yingjun children