Artificial intelligence: brief history, development, prospects

Lately we have been hearing more and more about artificial intelligence. It is used almost everywhere: from the field of high technology and complex mathematical calculations to medicine, the automotive industry, and even when operating smartphones. We use the technologies that underlie the work of modern AI every day and sometimes we may not even think about it. Machine and deep learning are used to train artificial intelligence, and works created by neural networks are sold for millions of dollars. But what is artificial intelligence? How does he work? And is it dangerous?

BB will be everywhere soon!

What is artificial intelligence

First, let's define the terminology. If you imagine artificial intelligence as something capable of thinking independently, making decisions, and generally showing signs of consciousness, then we hasten to disappoint you. Almost all systems existing today are not even close to this definition of AI. And those systems that show signs of such activity actually still operate within the framework of predetermined algorithms.

Sometimes these algorithms are very, very advanced, but they remain the “framework” within which AI operates. The machines have no “liberties,” much less signs of consciousness. These are simply very productive programs. But they are “the best in the business.” In addition, AI systems continue to improve. Yes, and they are arranged in a completely unusual way. Even if we put aside the fact that modern AI is far from perfect, it has a lot in common with us.

Solutions

Artificial Intelligence Platform

When we talk about a symbiosis called AIoT, we usually mean an AI platform located at the edge of the network. Typically this solution takes the form of a small industrial PC (IPC) with an integrated industrial-grade processor. However, for real-time data analysis, such a processor requires adequate support in the form of flash memory and disk storage.

Memory and data storage

To solve the problems of implementing AI in edge applications, as mentioned above, industrial-grade storage and memory devices are needed (Fig. 3). The first step is to study and identify the risks present in each specific data collection location. This will allow components to be tailored to the precise requirements of a specific application. Let's look at several examples of implementation of the proposed solutions.

Rice. 3. To implement AIoT, industrial-grade storage and memory devices are required. Similar solutions are offered, for example, by Innodisk

How artificial intelligence works

First of all, AI can perform its tasks (of which more later) and acquire new skills thanks to deep machine learning. We also often hear and use this term. But what does it mean? Unlike “classical” methods, when all the necessary information is loaded into the system in advance, machine learning algorithms force the system to develop independently by studying the available information. Which, moreover, the machine in some cases can also search on its own.

For example, to create a fraud detection program, a machine learning algorithm works with a list of bank transactions and their end result (legal or illegal). The machine learning model looks at examples and develops a statistical relationship between legitimate and fraudulent transactions. Then, when you feed the algorithm a new banking transaction data, it classifies it based on the patterns it learned from the examples beforehand.

Typically, the more data you provide, the more accurate the machine learning algorithm becomes at performing its tasks. Machine learning is especially useful in solving problems where the rules are not predefined and cannot be interpreted in a binary system. Returning to our example with banking operations: in fact, the output we have is a binary number system: 0 is a legal operation, 1 is an illegal one. But in order to come to such a conclusion, the system needs to analyze a whole bunch of parameters, and if you enter them manually, it will take more than one year. And it’s still not possible to predict all the options. And a system working on the basis of deep machine learning will be able to recognize something, even if it has never encountered exactly such a case before.

Tips for beginners

We have prepared the TOP 3 recommendations from experts for beginning AI specialists:

• Get a specialized education. For example, in leading online schools you will be given knowledge of higher mathematics and relevant programming skills. And an experienced mentor will help you create your own neural networks.
• Get an internship. You can find a junior vacancy in one of the companies in your city or an online internship, for example, Yandex periodically recruits interns.
• Select your industry. Think about what area you would be interested in working in (finance, marketing, sales, etc.) and develop your skills in this direction.

Deep learning and neural networks

While classical machine learning algorithms solve many problems that involve a lot of information in the form of databases, they do not do well with, so to speak, “visual and auditory” data such as images, videos, audio files, and so on.

For example, creating a breast cancer prediction model using classical machine learning approaches will require the efforts of dozens of medical experts, programmers and mathematicians, says AI researcher Jeremy Howard. Scientists would have to make many smaller algorithms in order for machine learning to cope with the flow of information. A separate subsystem for studying X-rays, a separate one for MRI, another for interpreting blood tests, and so on. For each type of analysis we would need our own system. Then they would all be combined into one large system... This is a very difficult and resource-intensive process.

Deep learning algorithms solve the same problem using deep neural networks, a type of software architecture inspired by the human brain (although neural networks are different from biological neurons, they operate in much the same way). Computer neural networks are connections of “electronic neurons” that are capable of processing and classifying information. They are arranged, as it were, in “layers” and each “layer” is responsible for something different, ultimately forming the overall picture. For example, when you train a neural network on images of various objects, it finds ways to extract objects from those images. Each layer of the neural network detects certain features: the shape of objects, colors, appearance of objects, and so on.

The surface layers of neural networks exhibit common features. The deeper layers already reveal the actual objects. The figure shows a diagram of a simple neural network. Green indicates input neurons (incoming information), blue indicates hidden neurons (data analysis), yellow indicates output neuron (decision)

Smartphone applications

Imagining our life without mobile phones is not an easy task. Various applications on our mobile phones have become an integral part of our daily lives. Many of these applications are based on artificial intelligence.

The built-in smart head-based assistants in our phones, such as Alice, Siri, Alexa and Google Assistant, are the most obvious examples of AI that most of us know and use.

More and more mobile technology platforms are developing solutions that use AI to manage various aspects of the device, such as battery management, event prompts, etc. One of the standard features, such as using portrait mode on the phone, also uses AI.

AI-enabled services now make up a significant portion of updates on iOS and Android platforms.

Are neural networks an artificial human brain?

Despite the similar structure of the machine and human neural networks, they do not possess the signs of our central nervous system. Computer neural networks are essentially the same auxiliary programs. It just happened that the most highly organized system for carrying out calculations turned out to be our brain. You've probably heard the expression “our brain is a computer”? Scientists simply “replicated” some aspects of its structure in “digital form.” This only made it possible to speed up calculations, but not to endow machines with consciousness.

This is interesting: When will artificial intelligence learn to reason?

Neural networks have been around since the 1950s (at least in concept form). But until recently, they did not receive much development, because their creation required huge amounts of data and computing power. In the last few years, all this has become available, which is why neural networks have come to the fore and received their development. It is important to understand that there was not enough technology for their full appearance. There are still not enough of them now in order to take the technology to a new level.

Stages of definition.

Closer to reality - AI in business processes of a modern office

Large modern organizations, for the most part, have gone through the stage of digitalization of basic business processes - office work, meeting management, working with contracts, requests from citizens and organizations, etc. Further development of projects occurs in two directions: the first is to cover new areas, such as personnel records management, business trip management, and the second is to reduce the labor intensity of performing daily tasks.

The use of intelligent services such as Directum Ario helps to reduce labor intensity and get rid of routine.

1 minute average processing time for one letter

Office work

. Manual transfer of information from received correspondence to the electronic document management system (EDMS) is eliminated:

1. documents are picked up from a scanner or email and then sorted into sets
2. Ario services also recognize text and extract the necessary information
3. all copies are classified and entered into the EDMS with automatic filling of cards.

Smart search Directum Smart Search will help you quickly find information, even if exact criteria are not specified, and the request is entered in free form and reflects only an approximate meaning.

more than 95% accuracy in determining invoices and cost items

Accounting

. The processing of incoming kits is simplified, and the correctness of their completion is checked automatically:

1. incoming primary accounting documents are recognized using Ario and distributed so that the accountant receives them in the form of ready-made sets for verification;
2. the artificial intelligence program checks the completeness of the kit, the correctness of the indicated amounts, and compares them with the order, specifications and nomenclature. The invoice and cost item are determined as accurately as possible.

In payment documents, the AI ​​system determines the accounting account and cost item by payment purpose and counterparty.

The preparation of advance reports is also simplified - documents are created automatically based on electronic tickets and photographs of receipts. Ario services fill out the fields of the report and send it for approval.

5 minutes to process the contract and identify changes

Contract activities

. The efficiency of already automated processes increases: Ario services check how different the content of the digital version of the contract is from the signed paper copy. All discrepancies are highlighted. In addition, the artificial intelligence system monitors the presence of mandatory details and checks the terms of the contract from the point of view of possible risks: fines, penalties, payment terms.

1-2 minutes average request processing time

Working with requests

. Ario services identify incoming requests by content and classify requests by type, for example, support requests and requests from legal entities and individuals.

Registration cards are filled out automatically, incl. the person responsible is determined. The registrar needs to check the correctness of filling.

more than 92% correctness of data extraction from documents

HR processes

. Ario services help to recognize personal documents (passport, certificates, diplomas, etc.) received from a scanner or mail. The HR service specialist receives documents for the candidate with already completed cards in the system.

More details on the Ario page

Vendors, including Directum, offer not just services, but a whole set of solutions to cover a whole area of ​​problems. Moreover, the company's flagship product is the Directum RX intelligent process and document management system.

What are deep learning and neural networks used for?

There are several areas where these two technologies have helped make significant progress. Moreover, we use some of them every day in our lives and don’t even think about what is behind them.

• Computer vision is the ability of software to understand the content of images and videos. This is one area where deep learning has made great progress. For example, deep learning image processing algorithms can detect various types of cancer, lung diseases, heart diseases, and so on. And do it faster and more efficiently than doctors. But deep learning is also ingrained in many of the applications you use every day. Apple Face ID and Google Photos use deep learning to recognize faces and improve the quality of photos. Facebook uses deep learning to automatically tag people in uploaded photos and so on. Computer vision also helps companies automatically identify and block problematic content such as violence and nudity. Finally, deep learning plays a very important role in enabling self-driving cars to understand their surroundings.
• Voice and speech recognition. When you speak a command to your Google Assistant, deep learning algorithms convert your voice into text commands. Several online applications use deep learning to transcribe audio and video files. Even when you “shazam” a song, neural network and deep machine learning algorithms come into play.
• Internet search: Even if you are looking for something in a search engine, in order for your request to be processed more clearly and the results to be as correct as possible, companies have begun to connect neural network algorithms to their search engines. Thus, the performance of the Google search engine increased several times after the system switched to deep machine learning and neural networks.

Case: increasing the ROI of contextual advertising by 2.2 times using conversion forecast

The project in question was implemented by the Dentsu Russia agency for the Eldorado retailer. His goal was to optimize advertising costs so that performance-related metrics would increase; and the frequency of contact with those who do not intend to make a purchase in the near future has decreased. After all, when purchasing traffic to a website in systems such as Google Ads, Yandex.Direct, Facebook, the main costs come from clicks from users who do not place orders. And if they do, they don’t always buy it back.

Dentsu used a model from OWOX BI, which, using machine learning technology, calculates the probability of making a purchase for each site user from the moment of his first visit, and determines whether it is worth spending the advertising budget on him further.

Such a model is trained on historical data on the behavior of site users, CRM data on purchased orders, and aggregated and anonymized data from tens of thousands of OWOX client projects.

As a result, the Eldorado network received:

1. Calculation of the probability of making a purchase for each site user, taking into account the order redemption rate.
2. This calculation is updated with every user action or inaction. That is, if a user visited the site and performed a series of actions, he is assigned a probability of X%. But if it doesn't return within the next few days, then the likelihood will decrease.
3. We identified 10 user segments, divided by probability in 10-point increments (10%, 20%, 30%, and so on). Then these segments are transferred to advertising services, and their effectiveness is analyzed in the context of advertising campaigns. This allows you to adjust your bids: reduce them for audiences with low probability and increase them for audiences with high probability.

To calculate the probability, the model takes into account more than 60 parameters, for example:

• number of sessions and hits within the conversion window;
• actions on the site during the session;
• temporary pauses between sessions;
• total number of actions;
• session device, operating system;
• what traffic sources the user had within the conversion window;
• number of actions on each page within a session;
• time of a specific session, total time of sessions within the conversion window.

The human mind is not capable of conducting such an analysis, but machine learning is. As a result, the ROI of advertising campaigns (taking into account the redemption of orders) increased by 2.2 times, and the difference in income from purchased orders increased by 2.7 times in favor of campaigns with OWOX BI audiences compared to the control group.

The Limits of Deep Learning and Neural Networks

Despite all their advantages, deep learning and neural networks also have some disadvantages.

• Data dependency: In general, deep learning algorithms require huge amounts of training data to perform their tasks accurately. Unfortunately, for many problems, there is not enough high-quality training data to create working models.
• Unpredictability: Neural networks develop in some strange way. Sometimes everything goes as planned. And sometimes (even if the neural network does its job well), even the creators struggle to understand how the algorithms work. The lack of predictability makes it extremely difficult to eliminate and correct errors in neural network algorithms.
• Algorithmic bias: Deep learning algorithms are only as good as the data they are trained on. The problem is that training data often contains hidden or obvious errors or flaws, and algorithms. For example, a facial recognition algorithm trained primarily on photographs of white people will perform less accurately on people of other skin color.
• Lack of generalization: Deep learning algorithms are good at performing focused tasks, but are poor at generalizing their knowledge. Unlike humans, a deep learning model trained to play StarCraft will not be able to play another similar game: say, WarCraft. Additionally, deep learning does not handle data that deviates from its training examples well.

How many times have you wondered while looking at an online advertisement, “I was thinking about buying this product”?

This is the work of AI.

There is no magic here when the AI ​​reads your thoughts. It simply tracks what you do on the Internet. The products you view on various shopping sites or search engines are tracked and advertisements related to those products are targeted to you.

AI also uses information related to demographics such as your age, gender, profession, etc. to determine advertisements for products that are most likely to be your preferences.

The future of deep learning, neural networks and AI

It is clear that work on deep learning and neural networks is far from complete. Various efforts are being made to improve deep learning algorithms. Deep learning is an advanced method in creating artificial intelligence. It has become increasingly popular in the last few years, thanks to the abundance of data and increasing computing power. It is the core technology behind many of the applications we use every day.

Schemes and ways to solve problems will soon replace a lot of things.

But will consciousness ever be born on the basis of this technology? Real artificial life? Some scientists believe that at the moment when the number of connections between the components of artificial neural networks approaches the same indicator that exists in the human brain between our neurons, something similar can happen. However, this statement is very doubtful. For true AI to emerge, we need to rethink the way we build AI-powered systems. Everything that exists now is just application programs for a strictly limited range of tasks. No matter how much we would like to believe that the future has already arrived...

What do you think? Will humans create AI? Share your opinion in our Telegram chat.

Who can't AI replace?

In 2016, Chinese engineer Andrew Ng announced on Twitter that computer speech recognition accuracy would improve to 99%, a figure almost reached today. For example, the Google voice assistant learned to recognize the medical names of drugs with an accuracy of 91.8%. In 2022, about 70% of US residents used Amazon smart speakers.

Does this mean that voice robot assistants will now appear in all companies? Yes and no. The solution to standard problems, of course, can be delegated to a machine, but from time to time situations arise when we really want to talk to a live employee, although the robot persistently asks us to press the next key.

In addition, the tasks of workers in many specialties do not fit into simple algorithms, while some involve a combination of monotonous (that is, completely programmable) operations and soft skills and attentive communication. Therefore, in many professions there will be a symbiosis of machine and man.

One of the leading entrepreneurs in the field of AI, Li Kaifu, places special emphasis on human feelings in his book “Superpowers of Artificial Intelligence”. After undergoing treatment for cancer, he realized the value of sharing his love with others: “There is no algorithm that could do for my recovery what my family did.”

Therefore, for the graph predicting the disappearance of professions, Lee chose the “compassion” and “creativity-optimization” . In his opinion, AI will replace humans only in areas that require neither compassion nor creativity. In all other areas, various options for the symbiosis of machines and humans await us.

The future division of labor between humans and AI, according to Li Kaifu, is based on the need for compassion (Y-axis) and creativity/strategy/optimization (X-axis). Screenshot of Li Kaifu's TED Talk, August 2018

Professions for which emotional contact and creativity are important cannot be replaced by computer algorithms. The sector with a predominance of the human role (top right) includes, for example, the following business specialties:

• Director of Advertising and Marketing,
• merger expert
• CEO.

That is, those who simultaneously know how to think structuredly and interact with people.

In the part where AI will be included in a person’s permanent work (upper left), Lee attributed:

• teachers,
• tour guides,
• event organizers,
• psychologists.

The computer will perform the analytical part, and the person will deal with the emotional part of communication.

The last sector where there was a place for a person (lower right) equally divides the work process. These include:

• scientists who constantly work with databases and have already fully automated some manual processes;
• people of creative professions who can be helped by AI with a selection of current trends;
• economists.

Professions that are fully automated (lower left sector):

• a hematologist whose job is to evaluate blood tests;
• a dishwasher (to which one could probably add a food delivery person, a dark kitchen picker, and even a fast food chain chef);
• customer service workers.

Examples of professions in the model of the future division of labor between humans and AI proposed by Li Kaifu.
Screenshot of TED Talk by Li Kaifu, August 2018 While it may be tempting to delegate all repetitive tasks to algorithms, complex and emotionally charged tasks require human input.

However, among other things, a person will also need to serve the AI: handle exceptions issued by the machine, control the process.

Smart input

Are you wondering where AI is used in input?

You don't need to look any further than your own mobile device. Typing has become more convenient with the integration of AI, where it predicts words, phrases and emojis based on your regular usage and writing style. Phrases can also be written by simply swiping across the screen.

Few email apps provide subject suggestions, and in many cases text, based on the email content you enter.

Financial services

Banking is one of the services that has embraced technological inventions earlier than most other areas. Banking has now moved beyond the need to move to a physical space to conduct transactions from your mobile phone.

Banks are using AI in many areas, including detecting fraudulent activities, analyzing customer investment trends, providing customer services, etc.

Have you ever received a notification from your bank when you made a transaction from a new device? This is a case of using AI to detect any potential fraud. Notifications received from banks and financial institutions regarding their services and products are examples of AI in understanding your preferences, requirements and financial strength to offer relevant products.