Overview of the Key Technologies of Artificial Intelligence

Key Components to Build an Artificial Intelligence Platform

Artificial intelligence is a new technical discipline that researches and develops theories, methods, technologies, and application systems for simulating the extension and expansion of human intelligence. The goal of artificial intelligence research is to let machines perform some complex tasks that require intelligent humans to complete. That is, we hope that the machine can replace us to solve some complicated tasks, not just repetitive mechanical activity but some that require human wisdom to participate in it. In this article, I will give you an overview of artificial intelligence technology and the key components that build an artificial intelligence platform.

Key Technologies of Artificial Intelligence

The key technologies of artificial intelligence can be divided into the infrastructure layer and the algorithm layer from bottom to top. In the infrastructure layer, there is basic hardware, including CPUs, GPUs, dedicated artificial intelligence chips, and high-speed networks. On top of this basic hardware, we can build algorithm frameworks, such as Tensorflow, Caffe, Mxnet, Torch, Keras, PyTorch, Theano, etc. Above the infrastructure layer is the algorithm layer. The most representative of the algorithm layer is machine learning algorithms, including a series of machine learning algorithms such as deep learning, transfer learning, general adversarial network, and reinforcement learning.

Infrastructure Layer

Basic Hardware

As we introduce the basic hardware of artificial intelligence infrastructure, which is divided into four directions: CPU, GPU, special-purpose chip, and high-speed network, let’s talk about two similar: CPU and GPU. CPU is mainly optimized for a set of serially executed tasks, while GPU is optimized to improve complex graphics and image calculation algorithms. The difference between the two is that CPU executes serially, and GPU is smaller, more efficient computing units that process calculations in parallel together. Furthermore, there is a special chip specially developed for artificial intelligence algorithms, Google’s TPU chip.

Photo by Christian Wiediger on Unsplash

To fully utilize the ability of artificial intelligence, we need a high-speed network. During the training and calculation of some complex data models, we need huge network bandwidth guarantees. Nowadays, the network has become an important part of the overall machine learning performance. Now we commonly have 10G, 20G, 40G networks. With the advent of Infiniband network technology, I believe that in the future, the network will provide a wider and faster channel for artificial intelligence learning and training.

Photo by Webaroo.com.au on Unsplash

Algorithm Framework

The second layer in the infrastructure is the algorithm framework. The algorithm framework can be simply understood as the framework for running algorithms. Just like a building framework, on which we can run our business. There is a very famous framework developed by Google, it’s called TensorFlow, which is friendly, fast and convenient. TensorFlow let us implement related AI algorithms in it, and run these methods.

Photo by :ch: Claudio Schwarz | @purzlbaum on Unsplash

Algorithm Layer

I just talked about the algorithm framework, then I will talk about the algorithm. First of all, machine learning is a core concept in artificial intelligence. All of us have to learn, and our human knowledge transfer is also carried out through such a method of learning. We learn the knowledge of our ancestors, and then create new knowledge by inference. We also hope that the machine has such ability: By learning the previous information, the machine is more like having intelligence and can react accordingly for new input in the future. This is called machine learning.

Process of Machine Learning

Do you know what is the most typical process of machine learning? First, we gather historical data, then we train the historical data. We get a model from the training result. Then as new data is input into that model, we can predict our results. This is the most typical machine learning process. We can simply say that a model is generated by training, and then the model is used to guide the prediction of new data.

We take the prediction of house prices as an example. First, we assume that house prices are only related to the area of the house. We can input one-dimensional data such as historical house prices and one-dimensional data of the area of the houses. In the house price model, we can predict the house price after entering the new house area after training. This is an example of the simplest machine learning.

Classification of Machine Learning

In machine learning, we can divide them into supervised learning algorithms, unsupervised learning algorithms, and semi-supervised learning algorithms. So what’s the difference between them?

Photo by Bruno Martins on Unsplash

First of all, let’s look at supervised learning. Supervised learning means that this learning is labeled during training. If the labels are discrete, we call it classification. If the labels are continuous, we call it regression. This is supervised learning.

Unsupervised learning is unlabeled, which is what we often call clustering.

The last kind of semi-supervised learning, that is to say, some of the samples in it are labeled, and some are not labeled, and they are put together for a more complicated mixed operation, then it becomes a semi-supervised learning algorithm.

Deep Learning Algorithms

When we talk about artificial intelligence, we often hear concepts such as machine learning and deep learning. In fact, they are an inclusive relationship, and artificial intelligence includes machine learning and deep learning, and a specific form of learning in machine learning is called deep learning. It mainly based on algorithms of neural networks. At present, deep learning has made great progress in fields of image recognition, speech recognition, natural language processing, audio recognition, social network filtering, machine translation, medical image analysis, and board game programs.

Neural Network

When it comes to deep learning algorithms, we have to talk about convolutional neural networks and recurrent neural networks. The neural network is similar to the neural transmission of the human brain, from one input unit to the next input unit to get a result. This is the principle of a simple neural network, which is to simulate the transmission of information from nerves in the human brain. It transfers information from one neuron to another and then passes down. With the rise of neural network algorithms, deep learning algorithms can solve many practical problems.

BP Neural Network

After the invention of the neural network algorithm, many problems have been solved to a certain extent. At the same time, people are constantly optimizing this algorithm. First, a very widely used and very classic one is the BP neural network. BP neural network has one more hidden layer than the original neural network. There are additional hidden layers in the input layer and the output. It can greatly reduce the amount of calculation and the difficulty of calculation by way of gradient descent.

Convolutional Neural Network

But after we have the BP neural network, we find that the computational load of the BP neural network is still very large. It sometimes fails to give the optimal solution within our acceptable time range, or it takes too long to give the optimal solution, which does not meet the needs of some of our applications. Then came the convolutional neural network (CNN), which is also a kind of neural network algorithm in essence, but it optimizes the content in the BP neural network, it makes the calculation faster, and it can get the most on many problems. Excellent solution. It improves the efficiency of its calculation by processing related information highly concurrently. At the same time, it greatly reduces the computational complexity between BP neural networks. Therefore, the convolutional neural network can currently reach the optimal solution in a fast time on many problems.

Recurrent Neural Network

Recurrent neural network models currently trained through this learning algorithm have reached a result that can write poems. It is not a single-term output, it is to use the previous output as a re-input, re-enter the model line calculation. This model is amazing because it has a certain memory ability.

Transfer Learning

In traditional machine learning, we have trained a few different models for specific problems, and each model can solve this type of problem. In transfer learning, we try to store the results of training in a certain field (we call it the original domain) as knowledge. We train the algorithm in the original domain. After training, we hope it can solve new problems, that are called targeted tasks. After we put in our target task, it can infer new problems based on the knowledge deposited in solving problems in the original domain and can get results without the need for training. We call it Transfer Learning, which transfers the results of previous learning to new problems.

Generative Adversarial Network (GAN)

The word “Adversarial” in the generative adversarial network (GAN) means that there is a competitive relationship between the two networks generated. Among these two networks, one is responsible for generating samples, and the other is for determining the correctness of the samples. The group that generates the samples hopes to fool the group that discriminates the samples. The group that discriminates the samples hopes not to be fooled by the results of the group that generates the samples, they have a certain relationship between competition and confrontation. It is more accurate to generate a learning result in such a relationship.

Reinforcement Learning

Reinforcement learning is also an important branch of machine learning. Its essence is to solve the problem of decision-making, that is, to make decisions automatically, and to make decisions continuously. It mainly includes two parts, the first is the agent, and then the environmental action reward. The goal of reinforcement learning is to get the most cumulative rewards. Let’s take an example, a baby learning to walk is a process of reinforcement learning. The baby is an agent, and he hopes to change the state of the environment through an action such as walking. Every step he takes is a change of state. If he takes the right step, we give him a reward. Otherwise, if he doesn’t take the right step, then we don’t give him a reward. Through this process, A little baby slowly learned to walk.

Artificial Intelligence Platform

The goal to build an artificial intelligence platform is to provide users with the infrastructure, algorithms, computing power, and related data needed for artificial intelligence research and development. Many large firms have planned and established unified AI R&D cloud platforms, which are deployed in the mixed cloud resource pool and directly connected to the IT big data platform. The unified AI research and development cloud platform is mainly composed of a data platform, a deep learning platform, and an AI capability platform. Among them, the data platform can provide open public data sets, and provide the massive data required for data mining, which can well explore the advantages of big data stored in most of the large firms, and provide data annotation and data sharing services. The deep learning platform is an AI cluster based on AI hardware and integrates a mainstream deep learning algorithm framework to provide infrastructure support for AI development, such as algorithms and computing power. The AI ​​capability platform provides core AI capabilities such as speech semantics, graphic images, and intelligent data analysis, as well as data model management and sharing mechanisms.

Summary

Let’s summarize what we have learned in this part. In this part, we first got an overview of the key technologies of artificial intelligence. It includes the infrastructure layer and algorithm layer. Then we explained some of the major infrastructures in the infrastructure layer, including hardware CPU, GPU, network, and dedicated chips and the framework of the underlying technology algorithms we build on the infrastructure layer, very representative of Google’s Tensorflow. We also introduced some classic artificial intelligence algorithms such as adversarial learning, reinforcement learning.

To add on, the development of artificial intelligence must also rely on big data technology. It requires tons of data for support. Technology innovation has just begun, and there are more new technologies that we need to keep learning. Technology innovation has just begun, and there are more new technologies that we need to keep learning.