What is Artificial Intelligence? How is Machine Learning Different from Deep Learning?

Artificial Intelligence (AI) is a vast branch of computer science that focuses on creating smart machines that can perform intelligent tasks. Such systems can operate because of machine learning algorithms, deep learning, or very boring things like rules.

AI should not be confused with Artificial Consciousness, when higher cognitive functions of a person are implemented in a computer, such as goal-setting, reflection, planning behavior, experiencing emotions, etc.

The concept of intellect, shared by humans and machines, can be articulated as the ability of a system, during self-learning, to create programs for solving certain problems and, as a result, to solve these problems.

One example of AI is Google’s AlphaGo system, which can beat a professional player in the board game Go with a score of 5 out of 5 games. Long before that, scientists had established that there are more than several billion variations of moves in this game. Not a single computer program could calculate all the options for the development of the party. However, the AlphaGo algorithm coped with this task. The algorithm did not calculate all the moves in advance but learned to play itself using the example of professional games and even playing with itself.

In commercial applications, AI is commonly used for:

• Natural Language Processing and speech recognition;
• automation of recognition of text, audio, images, video, faces;
• control of unmanned vehicles;
• big data analytics;
• computer vision and more.

Nowadays, AI is so prominent in our lives that many people do not think of it as AI and do not separate it from conventional computer technologies. Ask any passerby if there is AI in their smartphone and they will most likely say no. But AI algorithms are everywhere, from predicting typed text to auto-focusing the camera.

The term AI is rather generalized. In the media, AI can mean both the entire system, and its separate algorithm or narrow areas of AI, such as neural networks and deep learning.

Russia has a national standard for AI: GOST R 59277–2020 Artificial Intelligence Systems. Classification of Artificial Intelligence Systems. According to paragraph 3.18:

Artificial Intelligence is a set of technological solutions that allows to simulate human cognitive functions (including self-learning, finding solutions without a predetermined algorithm and achieving insight) and to obtain, when performing specific practically significant data processing tasks, results comparable at least to the results human intellectual activity. Note. The system of technological solutions includes information and communication infrastructure, software (including those that use machine learning methods), processes and services for data processing, analysis and synthesis of solutions.

Strong and Weak Artificial Intelligence

AI can be divided into two categories: strong and weak AI.

Weak AI, also called Narrow AI, is a type of artificial intelligence that works in a limited context and is an imitation of human intelligence. Weak AI is usually focused on a single task, such as Google search, text recognition, or voice synthesizing.

Strong AI or Artificial General Intelligence (AGI) is a type of artificial intelligence that is capable of thinking and acting like a human and performing complex tasks. For example, we can see Strong AI in science fiction films about robots: Westworld, Blade Runner, or Terminator.

How Does AI Work?

Strong AI has not yet been developed, so let’s look at how weak AI works.

To make a computer mimic human intelligence and teach it to perform a task, it needs to be trained to do so. Simply writing a program that will take into account all possible problems and contain all possible solutions is not the way to do it. Instead, an algorithm is put into the computer to independently find solutions based on statistical data. This is how Machine Learning algorithms emerged, which were then supplemented with various Neural Network and Deep Learning methods.

Machine Learning

Machine learning is an approach where a computer algorithm learns to solve a problem by itself. To start the machine learning process, a certain amount of initial data or a dataset with some properties or characteristics must be loaded into a computer. Then you need to choose an algorithm that will learn to process requests and give the most accurate answer. Algorithms can be anything from linear regression to neural networks.

For example, in AI, photos of dogs and cats are loaded with tags indicating whether they are dogs or cats. After the training process, AI itself will be able to recognize dogs and cats in new images without tags with some probability. The learning process will continue after the forecasts are issued. The more data is analyzed by the program, the more accurately it recognizes the desired images.

Basic Machine Learning Methods

Supervised learning is used when developers have a labeled dataset and know exactly which features the algorithm should look for. Such models are used in spam filters, language and handwriting recognition, fraudulent activity detection, calculation of financial indicators, product demand, or medical diagnoses.

Unsupervised learning is used when there is unlabeled or raw data and the computer needs to independently find signs and patterns. This approach is often used to cluster data, reduce its size, and find associations. For example, to define a smile in a photo, compress images or create a recommendation system.

Reinforcement learning is training an AI agent to survive in the environment in which it exists. The environment can be anything from a video game to the real world. Reinforcement learning rewards the agent for doing the right thing and punishing mistakes. The algorithm does not have to remember all its previous experiences and calculate all possible scenarios of events. It must learn to act according to the situation. For example, reinforcement learning is used by engineers for drones, robot vacuum cleaners, and company resource management.

Ensembles are groups of algorithms that use multiple machine learning methods at once and correct each other’s errors. Ensembles are used in search engines, computer vision and face recognition.

Deep learning is a subset of machine learning in which a computer uses neural networks with many hidden layers of neurons as an algorithm to work.

Neural networks

Neural networks are an attempt to simulate how the human brain works with a computer. A precise description of how our brains work has not yet been created, but it is usually modeled using the concept of neurons and neural networks. The brain is believed to contain approximately 100 billion neurons.

Put simply, an artificial neuron can be represented as an adder having several inputs and outputs. Each entry has its own weight. The information from the entry with the highest weight will prevail. The neuron receives information from all inputs (the sum of all data multiplied by the corresponding weight coefficients), processes it in accordance with some function (the activation function can be linear, like a hyperbolic tangent or sigmoid). The neuron then sends the result to the output. The output can be connected to the inputs of other neurons.

Neurons are grouped into layers and connected in various ways. Information in a neural network can be transmitted not only in one direction but also through feedback (recurrent networks).

Neural network training comes down to choosing input weights across the entire neural network. The dataset is run through the neural network in several iterations until the required accuracy of the neural network operation is obtained. Neural network training is a very time-consuming and computational resource-consuming process. However, after training the neural network, the answer will be received much faster.

To speed up the operation of neural networks on computers, special processors are used, for example, one of the well-known machine vision processors from Intel — Movidius Myriad X with a performance of 1 Teraflop and power consumption of no more than 1 W. These accelerators allow AI to be embedded in virtually any compact PC or camcorder.

Types of Neural Networks and Their Application

The most popular networks today are Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). CNN is often used for face recognition, finding objects in photos and videos, improving image quality, and more. Recurrent networks have found application in machine translation of text and speech synthesis. For example, since 2016, Google Translate has been operating based on the RNN architecture.

Generative adversarial networks (GANs) have also gained popularity. They are based on two neural networks, one of which generates data, for example, an image, and the other tries to distinguish correct samples from incorrect ones. Since the two networks compete with each other, an antagonistic game arises between them. GAN is often used to create photorealistic photographs. For example, the “This Person Does Not Exist” image repository consists of portrait photos of “people” generated by a generative neural network.

AI in Industry

Of course, AI is now actively used in industry. Here are some examples.

Datana has introduced AI at the Ashinskiy metallurgical plant in the electric steel-making shop. The most expensive materials for steel smelting are ferroalloys, the price of which can vary from 60 thousand rubles to 3 million rubles per ton. Optimizing the consumption of ferroalloys significantly saves costs. Datana has trained AI to predict the chemical composition of steel and recommend the correct amount of ferroalloys. Calculation for only one steel grade showed that the use of AI during pilot production brought savings in the consumption of ferroalloys of the order of 8 %.

Cognitive Pilot has robotized forage harvesters. Neural network-based AI makes it possible to introduce an autonomous control system into the combine with the ability to detect obstacles and people.

Gazprom Neft has implemented a Cognitive geologist project, which helps select the optimal scenarios for creating geological models, automate routine operations at the geological exploration stage and increase the efficiency of investment decisions.

One of Uralchem’s branches has introduced a system for intelligent control of the drying and granulation process — a granulator-dryer drum operator advisor. The system increased the stability of work and reduced the influence of the human factor in process control, increasing output by 2–6 % while maintaining quality.

Conclusion

Artificial intelligence is the science that allows creating smart machines and the ability of a computer to learn and make decisions on its own.

Machine learning is one area of artificial intelligence that allows a computer to learn new things without being explicitly programmed to do so.

Deep learning is a machine learning method in which a computer uses neural networks with many hidden layers of neurons as an algorithm to operate.

Of course, AI is not yet as advanced as the human brain, but it can already solve applied problems. In the next few years, we will see significant AI development, as well as new ways of using it.