Machine Learning is essentially a multidisciplinary field. It draws on the success of artificial intelligence, probability statistics, computational complexity theory, control theory, information theory, philosophy, physiology, neuroscience, and other disciplines. Machine Learning is a behavior in which computer programs improve the processing performance of a task through experience.
More accurate definition:
Definition: if a computer program measures the performance of a certain type of task t by P, It is developed by me based on experience. So we call this computer program learning from experience E. For a certain type of task T, its performance is measured by P.
Generally, in order to better define a learning problem, we define three features:
Types of tasks, standards for measuring task improvement, and sources of experience.
For example:
Handwriting Recognition learning problems:
Task T: Recognize and classify handwritten text in Images
Performance Standard P: classification accuracy rate
Training experience E: handwritten text database with known categories (Knowledge Base)
When the problem features are clearly analyzed, the most important thing is the design of the learning system.
Consider the learning system:
1. Exact types of knowledge to be learned
2. Representation of this target knowledge
3. A Learning Mechanism
In general, the optimization problem can be summarized as a search problem, but the corresponding search space is huge and the best search strategy is not. Select the target function. The target function is usually related to the selection policy or a function that evaluates the selection policy. The selection of the target function needs to be determined by problems, and constant attempts and comparisons are also required.
In general, learning tasks can be simplified to discovering operational descriptions of an ideal target function. It is usually very difficult to learn such an ideal target function perfectly. In fact, we usually only want to learn algorithms to obtain similar objective functions. For this reason, the process of learning objective functions is often called function appropriation)
Specify the target function
Select models (such as linear models and Artificial Neural Networks)
Learning Algorithms to train various parameters and weights
Best fit of existing data (training sample) models or functions
Continuous learning and correction
In terms of machine learning, an effective point is that machine learning often comes down to the search problem, that is, searching for a very large hypothetical space, to determine an assumption of the best fit observed data and existing knowledge of the learner. Generally, the task of the learner is to search for a certain search domain space to locate the hypothesis that best fits the training data.
The learning method is characterized by the internal structure of the search space explored by the search policy and the learner. We will find that, this viewpoint provides a formal analysis of the size of the hypothetical space to be searched, the number of available training samples, and a hypothesis consistent with the training data, which can generalize to the confidence level of no instance..
For a machine learning problem, you often need to think about the following issues:
1. What algorithms can be used to learn general objective functions from specific training data? If sufficient training data is provided, under what conditions will a specific algorithm converge to the expected function? Which algorithm has the best performance on which issues and representations.
2. How much training data is sufficient? How can we find the general relationship between the confidence level of the Learning Hypothesis and the number of training data and the hypothetical space features provided to the learner?
3. How does the learner's prior knowledge guide the process of generalization from the sample? Will a prior knowledge be helpful when it is only approximately correct?
4. What is the best strategy for choosing an effective training experience? How does the choice of this policy affect the complexity of the learning problem?
5. How can I simplify a learning task into one or more function approximation problems? In another way, what functions does the system try to learn? Can this process be automated?
6. How does the learner automatically change the representation to improve the ability to represent and learn objective functions?
Note: The above content is taken from and summarized in Mitchell, t. M Machine Learning
Machine Learning-basic problem definition, task determination, and concept understanding
