AI Artificial Intelligence - Introduction to Deep Learning

First of all, I declare that many articles from different photo sources are used in the following. If there is any infringement, please inform me and delete it immediately. Preface This document is for friends who are ready to get started or want to know the hot AI technology at present. First, under popular science, AI is a big subject, and the way to realize it is traditional machine learning. Algorithms, but at present the best and most powerful algorithm is deep learning algorithm, so now commonly referred to as artificial intelligence technology generally refers to deep learning. The concept of machine learning was put forward as early as the 1980s. Because of the limitations of computer performance and traditional algorithms, there was no large-scale application until the emergence of deep learning algorithms in 2006. Compared with traditional machine learning algorithms, breakthroughs have been made in many areas, such as speech recognition translation. Progress in the domain. At present, large companies attach great importance to the layout of AI. At present, the two most influential technologies in the future are AI and gene editing. A lot of people know that AI is probably just hearing Alpha Go beat the strongest human players, face recognition, automatic driving, these areas, will feel that these things are far away from themselves, that this technology or concept is not like other development languages can actually be used for projects, or that this technology is very difficult. Doctors and masters can play well. Once no one taught me. I looked for a course on the Internet and talked about a lot of mathematical formulas. It seemed that they were really big and almost frightened. But I'm telling you it's not that hard to learn. Next I'll take you through some basic concepts and explain the principles in the most straightforward terms. 2. Overview of Artificial Intelligence

At present, the main branches of deep learning algorithms are probably these. Each algorithm absorbs the advantages of the predecessors and then improves continuously. Thanks to the spirit of open source abroad, this kind of open source promotes the effect of group accelerating innovation. In recent years, deep learning has developed very fast. In order to promote deep learning, Google has made great progress. The development of Degree Learning in the field of vision, ImageNet Visual Competition started in 2004. ImageNet contains more than 15 million artificially annotated images, or labeled images, that illustrate the content of the images, more than 22,000 categories. Among them, at least 1 million of the networks listed in the SLIM category above the bounding box are those that have won the championship in recent years. Until 2017, the competition was declared off, because the best network recognition rate has reached 97% accuracy, to some extent, has surpassed the human eye. So there's no need to hold it again. Let me briefly describe the application areas of these algorithms. CNN: It is mainly used in the field of computer vision, including classification, recognition and target detection.

RNN: The main application areas are time series feature data, such as speech recognition, translation, NLP natural language processing, weather forecasting, stock quantitative investment and so on. GAN: Self-coding feature engineering, image style migration, automatic picture coloring, image super-resolution restoration, etc. Enhanced learning: mainly used in the game field, open industry. OpenAI, the source framework, provides a lot of basic operations. It allows AI simulators to play games, learn game strategies and constantly improve themselves. Like Alpha Go, 3. The current mainstream deep learning framework

At present, TensorFlow ranks first in the overall ranking. It is the most popular framework for in-depth learning. It has abundant information and many documents. It is easy to use Python language. Keras is a top-level encapsulation framework, which encapsulates Tensorflow, Theano and CNTK once again, and can switch different backend networks. However, I suggest that novices still switch from T. The ensorflow framework starts, familiarizes itself with it, and then uses Keras. 4. What is linear separability? Linear inseparability. Let's start with the simplest classification problem. First, let's talk about a basic mathematical expression of neural network neurons: y = f (w x + b) y is the output, f is a non-linear activation function, W is the weight, x is the input, B is the bias. Don't worry, these concepts have slowed down through the following examples. solution

We need to separate X from O in the graph. This is a simple case. A straight line can do it. y = Wx + B. For example, if you enter a number into the equation now, you get Y above the right of the line, which is the classification of X, and below the left of the line, which is the classification of O. This is the simplest case, if the data becomes this way.

Can it be separated by a straight line? The answer is no, so what? This is not possible in two-dimensional space, which maps to multi-dimensional space.

The mapping is also very simple. Traditional machine learning is called kernel function and deep learning force is called non-linear activation function such as Sigmod.

For example, Wx + B mainly makes some linear changes, such as rotation, translation, zooming and zooming, to distinguish the two-dimensional data. Adding the non-linear activation function, the data will be distorted. For example, in Figure 1 below, the blue and red regions can not be cut by a hyperplane. After twisting and deforming into Figure 2, it can be separated.

Figure 1

Figure 2

You can open this website http://playground.tensorflow.org to experience data flow, change, and finally cut them apart. 4. What is a neural network?

Biological neuron cells

The above image is a biological neuron. The synapses around the cells are signals received, processed by the neurons, and then sent to the next neuron for computation. Mathematicians have invented this algorithm by imitating the principle of mathematics.

Please read the Chinese version for details.