Article Directory
Machine Learning
Machine Learning (Machine Learning) is a specialized computer simulation to study how to implement human learning or behavior, in order to acquire new knowledge or skills, re-organize existing knowledge structures so as to continuously improve their performance discipline. Can machines like humans can have the ability to learn it? In 1959 the United States Samuel (Samuel) designed a chess program that has the ability to learn, it can improve your chess constantly in chess. Four years later, the program defeated the designer himself. It took another three years, this program defeated the United States to maintain a victorious 8-year-old undefeated champion.
Although machine learning development for decades, but there is still a lot of no good to solve problems:
such as image recognition, speech recognition, natural language understanding, weather forecasts, gene expression, content recommendation, and so on. Currently we through machine learning to solve these problems is the idea of such a (visual perception as an example):
Obtaining data from the beginning by a sensor (e.g., CMOS). Then after preprocessing, feature extraction, feature selection, then the inference, prediction or recognition. The last section, which is part of machine learning, most of the work is done in this area, there are also a lot of paper and research.
And the three middle part, summed up expression characteristics. Good features of expression, the accuracy of the final algorithm has played a very key role, and the system main computing and test work is spent in the most. However, this practice is generally manually performed. By manual extraction feature .
However, manually select the feature is a very laborious, heuristic method (requires specialized knowledge), you can not select a large extent on experience and good luck, and it takes a lot of time to adjust. Since the hand-select the feature is not very good, you can not automatically learn some features of it? The answer is energy! Deep Learning is used to do this thing to see one of its aliases UnsupervisedFeature Learning, will be the name suggests, Unsupervised means that people do not choose to participate in the process characteristics.
feature
Machine learning system is characterized by raw materials, the impact on the final model is beyond doubt. If the data is well expressed became a feature, usually linear model can achieve satisfactory accuracy. For that feature, what we need to consider it?
Grain size characteristics represented by the
learning algorithm in a feature on what size says can only play a role? On a picture, the pixel-level features of no value. For example following a motorcycle, from pixel level, do not receive any information, it can not distinguish between a motorcycle and a non motorcycles. And if time is characterized by having a structural (or meaningful), such as whether the car handle (handle), whether the wheel (wheel), it is easy to distinguish the motorcycle and non-motorized vehicles, learning algorithms can play a role .
Primary (superficial) wherein represents a
complex pattern, often a few basic structures. The image not only the existence of this law, there are also sound. They never found a voice marked 20 basic structure of the sound, the sound can be synthesized from the rest of the 20 kinds of basic structure.
Structural features represent
small graphics can be composed of basic edge, more structured, more complex, with a conceptual graphical representation of how it? This feature requires a higher level of representation, such as V2, V4. So look at the pixel level V1 is the pixel level. Look V1 V2 is the pixel level, this level is progressive, high-level expression expressed by the combination of the underlying. Professional point that is the base basis.
Intuitively, it is to find a small patch make sense of it and then carried combine, get on the floor of the feature, recursively learning feature upwards.
From the text, one doc What does it mean? We describe one thing, what is there to represent more appropriate? With a word Well, I do not think the word is the pixel level, at least, should be a term, in other words each doc by the term constitution, but it represents the concept of competence is enough Well, probably not enough, the need to step, reaching topic level, with the topic, then it is reasonable doc. But a large number of gaps at each level, such as doc concept represented -> topic (one thousand - Man of the order) -> term (10 Wan middleweight) -> word (a mega).
A person looking at a doc, the eyes see is the word, word formation term of these words automatically cut in the brain, in the manner of the concept of organization, prior learning, get topic, and then to high-level learning .
Depth study
The basic idea of the Deep Learning
Suppose we have a system S, which has an n-layer (S1, ... Sn), which is the input I, output is O, vividly expressed as: I => S1 => S2 => ...... = > Sn => O, O if the output equal to the input I, that is, without any loss of information input after I go through this system change (Oh, Daniel said it was not possible. there is information theory "information layer by layer missing" statement (information processing inequality), set up a process to get information b, then c to b to get treatment, it can be shown that:. a and c of mutual information does not exceed the mutual information indicating that a and b of the information processing information does not increase, most of the processing information is lost. of course, if you lose the useless information that is more than good), maintained unchanged, which means that the input I go through every layer of Si without any loss of information, that is, at any layer of Si, which is original information (i.e., the input I) is represented by another. Returning now to our theme Deep Learning, we need to automatically learning feature, assume we have a bunch of input I (such as a stack of images or text), suppose we designed a system S (n-layer), we adjust system parameters, so that its output is still the input I, then we can get a series of feature-level input I get automatically, that is, S1, ..., Sn.
For deep learning, its idea is for stacking a plurality of layers, that is to say the output of this layer as a lower layer input. In this way, we can achieve the input information classification expressed.
In addition, the front is assumed that the output exactly equal to the input, this restriction is too strong, we can slightly relax the restrictions, for example, we make a difference as long as the input and output can be as small as possible, this relaxation will lead to another type of different Deep Learning methods. Above is Deep Learning the basic idea.
