Introduction to Machine Learning in Visual Field

There is a problem.

How can machines become more intelligent by self-learning?

The answer is that machines, like humans, need to improve their intelligence through continuous learning and experience gained. They learn constantly through neural networks.

The human brain is the smartest machine. There are billions of neurons in the brain. These neurons receive information synchronously, classify data elements according to their types, and then search for and store and archive the data. This is a simplified description of how our brains work.

Human Brain. png

People see cars. png

Neural networks, like the brain, receive input information, process information, and transmit output information. In order to be able to process information, it must be adequately trained.

Neural network. png

The architecture of the neural network is multi-layered. Each layer receives input and produces output. Each middle layer uses the output of the previous layer as input. All these different layers are called Deep Neural Network (DNN). The image application of DNN is called Convolutional Neural Network (CNN), which is suitable for both 2D and 3D images.

Multilayer Neural Network. png

So how do developers train deep neural networks? What is the goal of training?

Training Neural Network. png

Neural network needs a lot of empirical data, such as a large number of classified images, so that it can eventually classify new images. Image databases provide data, and software databases provide a framework for data processing. Google has developed its own machine learning framework, TensorFlow. TensorFlow is used in many of Google's products, such as oil pipes, such as Google Photo and other Google products. Many big companies, such as IBM, Microsoft and Facebook, have invested billions of dollars in the technology.

In recent years, researchers have achieved success in the field of object recognition. In the field of robots and biometrics, in the early warning system of automatic driving, handwritten recognition and automatic translation tools have been successfully applied.

Object recognition. png

Applying 1-Robot Grab.png

Application of 2-Autopilot.png

Application of 3-Handwriting Recognition.png

Application 4-Automatic Translation of App.png

Today, tool technology has matured, not only for large companies such as Google or Microsoft, but also for small companies.

So what is the technical principle behind this? What is the real value of neural networks?

Why do we need machines to look at images? Because the number of image sensors is larger than the population, it is impossible for people to physically check every picture or video they get.

Man and Sensor.png

Searching for specific images on the Internet, we will use search engine image search, such as Google's image search engine. The working principle and algorithm of the image search engine is based on 22-layer deep convolution neural network, which we call Inception. The classification result of this neural network is much better than that of manual classification.

Inception.png

The next generation of neural networks will be the generation of adversarial networks. It is a combination of two neural networks. One is used to generate training sets and the other is used to generate other data sets.

GN.png

Finally, machines can really teach themselves:)

This article is translated from the following links.

Https://dwz.cn/oI1rSs8b

Author: Hauge's World

Link: https://www.jianshu.com/p/c8bd21eef401

Source: Brief Book

The copyright of the brief book belongs to the author. For any form of reprinting, please contact the author for authorization and indicate the source.