Artificial intelligence – magic and practical use

Artificial intelligence: What does it take for a specific application in a company?

Artificial intelligence is in itself nothing more than an efficient form of statistical data analysis. By coupling it with a learning function, an assumption is made on the basis of already evaluated data as to the probability with which a new data set can be assigned to certain criteria. This is how the magic happens. Together with the speed of modern computers, it becomes a very powerful instrument of automation.

The reason why this technology is so important in today’s digital age is that we collect vast amounts of data. It is no longer possible to process this data using conventional means. Redundancies are accumulating in various areas. This is data with the same or almost the same content.

Nevertheless, the analysis of new data can be significantly optimized. It turns out that machines can do this much better and faster than humans in many areas.

Unorganized data – Data Lakes

Unorganized data in large quantities is referred to as data lakes. In order to organize the data in a data lake, an understanding, i.e. intelligence, is required. This understands in a semantic context how different things belong together and what connects them.

Simply organizing this data is the core of the programs that we refer to as artificial intelligence. Behind this are algorithms for machine learning and the special form of deep learning.

The more data a deep learning algorithm has at its disposal, the better it can assign another data set to certain criteria. We speak of learning algorithms.

Understanding the content

The prerequisite for organizing data is an understanding of the content of the data fragments to be organized. Normally, the data to be organized is not already available in a digital, statistically comprehensible form. Images, videos and audio are complex formats. These must first be analyzed before they can be sorted.

Let’s take the example of a lot of physical documents that come together in a company.

• Invoices
• Contracts
• Concepts
• Presentations
• Manuals
• …

When a person is asked to sort these documents, they can recognize relatively quickly which category each document belongs to based on intuitive criteria.

Various fields are relevant for an invoice, such as date, amount, sender, account, subject … are relevant. The invoices can be sorted accordingly. With contracts, it is somewhat more complicated to recognize the various fields. Nevertheless, a person can use their senses and intelligence to learn relatively quickly what to look for.

This is somewhat more difficult for a computer program. Even if invoices are relatively standardized, the individual fields are not always in the same place. It therefore requires an overview of the document and therefore a basic understanding.

In order to create this understanding of physical documents, we first need an understanding of the computer image.

The great achievement and prerequisites

Artificial intelligence programs can be used in almost all areas where we as humans are active. They adapt our way of learning. As children, we can and know almost nothing, but our brains learn quickly. We learn through the experiences we constantly have and by imitating our role models.

This basic premise was transferred from science to computers. For a long time, it could not develop because the computing power of computers was not comparable to that of our brains. This has changed in recent years. The speed of computers for calculations has doubled over the years. Today, they are able to analyze large amounts of data on specific topics in such a way that learning can be generated.

Our brain is incredibly complex. Compared to its size, it is infinitely efficient. Nevertheless, it is actually specialized in securing our lives and our evolution. This includes prioritizing and forgetting.

In contrast to our brain, programs that implement deep learning algorithms are very specialized. For example, they are exclusively dedicated to image recognition. Others take care of balance control in robots. Still others are designed to play games such as Go or chess.

Even if the programs of artificial intelligence differ greatly in their characteristics, the basis of learning is similar in all these cases and can be reused. What changes is the underlying data.

In the learning mechanism of these programs based on a specific data set, it is not possible to understand why and how a program identifies a specific data set after training the algorithm. The result is a black box in which the decisions in the assignment of the data achieve an ever-increasing probability of correct assignment. However, the many necessary calculations and weightings create a dynamic of their own. In individual cases, it is no longer possible to determine how the decision was made in detail.

On the one hand, this is what fascinates us so much about artificial intelligence, but on the other hand, it presents us with major challenges from a responsibility perspective. As humans, we have organized ourselves in our society in such a way that everyone should abide by our rules. We need our intelligence for this.

This is of fundamental importance, especially for our insurance companies.

So who should take responsibility for artificial intelligence when it makes decisions?

We probably encounter the best-known problem with autonomous driving. In a precedent case, where the computer has to decide between two foreseeable scenarios, whether it will save itself or whether it prefers to save the life of a pedestrian, for example, the discussion becomes philosophical and we will have to decide how much autonomy the computer programs should have.

In the meantime, while we are thinking about the elementary basis, the algorithms are developing very quickly.

The challenge for development

The iterative approach to learning involves a different approach to software development than with conventional programs.

The tech giants, and Google in particular, provide the learning algorithm, which is constantly being further developed by Google itself. Just like the AWS Cloud, this creates a new basic infrastructure. Google will probably give us access to quantum computers in a similar way in the near future.

In software development, which builds on these services to create application programs that use artificial intelligence, the development process is characterized by several steps.

I. A use case is needed to decide on the order based on a set of data

II.a In a requirements analysis, the first step is to define the data and structures that will be used as input. On the one hand, this raises the question of what exactly the data is that can be supplied at all, compared to what the system needs to achieve a result.

II.b In the design of a solution, the data must then be prepared for the algorithm.

III The AI algorithm must be connected to the data and to deliver a result it needs information for the output of the result.

IV. The result of the AI module is implemented in the solution to control it.
(An example would be a machine that sorts potatoes according to their size and shape. The potato is measured using a video sensor. The relevant criteria from the measurement data feed the AI algorithm, which decides which further path the potato should take on the conveyor belt).

V. This is followed by the system test and thus the training of the AI module

VI Finally, the solution is to be transferred from the test phase to the operating phase.

Compared to normal programs from software development, the operational phase of AI-controlled programs is where the major effort of allowing the system to learn begins. A conventional program is put into operation and delivers the desired result. Errors are then removed and the solution is optimized at the same time.

In a solution that uses AI, this must also happen from a software development perspective. However, this process runs parallel to the optimization of the learning algorithm.

For robots, for example, the engineering requirements are the same as for the core AI algorithm that teaches them to adjust and behave correctly.

Concrete use in a company

We can make the greatest use of AI in the publication of content. The search engines of Google and YouTube are controlled by one of the most highly developed AI instances and influence us on a daily basis. The search engine filters data according to relevance and displays it in a personalized way. The more the Google search engine can understand what an article, image or video is about, the better it can answer our search queries.

We can no longer really control which data is delivered to us when we place search queries.

As marketers, we can use Google’s AI mechanisms to effectively advertise to a target audience.

The use of in-house AI programs in industry is always about optimizing existing processes that require repetitive or parallel work. Where there is repetition and parallelism in processes, there is potential for optimization.

However, this also means that various professions are simply being eliminated due to superfluous processes.

We need to think seriously about how everyone in our population finds their place in modern trends.