How to correctly understand the concept of deep learning (learning)

Deep learning is now a hot concept in machine learning, but the concept has become a bit of a myth as it is reproduced in various media: for example, deep learning can be thought of as a machine learning method that simulates the neural structure of the human brain, thus enabling the computer to have the same intelligence. , and such a technology will undoubtedly be promising in the future. So what kind of technology is deep learning in essence?

What is deep learning?

Deep learning is a method for modeling patterns (sounds, images, etc.) in the field of machine learning, and it is also a statistical-based probabilistic model. When models are modeled, various patterns can be identified, for example, if the mode to be modeled is sound, then this recognition can be understood as speech recognition. And analogy to understand, if the machine learning algorithm analogy to the sorting algorithm, then deep learning algorithm is one of the many sorting algorithms (such as bubble sort), this algorithm in some applications, it will have some advantages.

where does deep learning "depth" reflect

On the term "depth" in deep learning, people may think that deep learning can do more things than traditional machine learning algorithms, and it is a more "advanced" algorithm. And the fact may not be as we think, because from the point of view of the input and output of the algorithm, the deep learning algorithm and the traditional supervised machine learning algorithm input and output are similar, whether the simplest of the logistic Regression, or later SVM, boosting and other algorithms, The things they can do are similar. As with whatever sort algorithm is used, their input and expected output are similar, except that the various algorithms perform differently in different environments.

So what does deep learning "depth" essentially mean? The scientific name of deep learning, also known as deep-neural Networks, is developed from a long-ago artificial neural Network (Artificial neural Networks) model. This model generally uses the graph model in computer science to express intuitively, while deep learning "depth" refers to the number of layers of the graph model and the nodes of each layer, which has a great degree of ascension relative to the previous neural network.

      Deep learning also has many different forms of implementation, and it also has different names depending on the problem solving, application areas and even the author's idea of the title: Convolutional Neural Networks (convolutional neural Networks), Depth confidence Network (deep belief Networks), restricted Boltzmann machine (Restricted Boltzmann machines), deep Boltzmann machine (Deep boltzmann machines), Recursive automatic encoder (Recursive autoencoders), depth expression (deep representation), and so on. In essence, however, they are similar deep neural network models.

      Since deep learning such a neural network model has been in the past, why has it gone through a decline, and now again into the eyes of people? This is because in the hardware conditions more than 10 years ago, the modeling of high-level multi-node neural networks, the time complexity (possibly in years) is almost unacceptable. In many applications, the actual use of some of the more shallow network, although this model in these applications, has achieved very good results (even the state of art), but because of this time is unacceptable, limiting its application in the promotion of practical. By now, the level of computer hardware has not been the same as before, so a model of neural networks into the people's eyes.

       "   

from Google brain This project we can see that the neural network model for the computational requirements is extremely large, in order to ensure the real-time nature of the algorithm, the need to use a large number of CPUs for parallel computing.

Of course, there is another reason for deep learning now, of course, because in some scenarios, the accuracy of this algorithm pattern recognition exceeds the majority of existing algorithms. Recently, after the author of the Deep learning has modified the bug of implementing the code, the accuracy of the model recognition has been greatly improved. These factors together cause a deep neural network model, or a new craze for deep learning of such a concept.

Benefits of Deep learning

in order to identify a pattern, the usual practice is to extract the characteristics of the pattern in some way. The method of extracting this feature is sometimes manually designed or specified, and is often summed up by the computer itself, given the relatively large number of data. Deep learning puts forward a method for computer to automatically learn the feature of the pattern, and integrates the feature learning into the process of establishing the model, thus reducing the incompleteness caused by man-made design characteristics. At present, some machine learning applications with deep learning as the core have achieved the recognition or classification performance beyond the existing algorithms in the application scenario that satisfies the specific conditions.

disadvantages of deep learning

Although deep learning can automatically learn the characteristics of the pattern and can achieve a good recognition accuracy, but the premise of this algorithm is that the user can provide "quite large" magnitude of data. In other words, in a scenario where only limited data is available, the deep learning algorithm is not able to estimate the law of the data without bias, so the recognition effect may not be as good as some existing simple algorithms. In addition, due to the complexity of the graph model in deep learning, the time complexity of the algorithm is increased dramatically, in order to ensure the real-time of the algorithm, higher parallel programming skills and better hardware support are required. Therefore, there are only a few economic strength of the scientific research institutions or enterprises, in order to sufficient depth of learning algorithms, to do some of the more advanced and practical applications.

This article reprinted from:Excalibur's Column

From for notes (Wiz)

How to correctly understand the concept of deep learning (learning)