AI and humans: a profitable cooperation

Advances in artificial intelligence continue to progress rapidly and present our society with profound structural changes. For example, implementation at municipal level is now also foreseeable. As with any technological revolution, this is not without its concerns. These are best countered with knowledge about the background to the technology. And these, in turn, are reflected in the history of Konfuzio.

Sovereign AI - just one part of the whole

Long before ChatGPT, film history in particular shaped the public understanding of artificial intelligence. In "2001: A Space Odyssey" from 1968, for example, it appears in the form of the HAL 9000 supercomputer. Equipped with image and speech recognition as well as impressive communication skills, it forms the heart of a space mission and at the same time the autonomous control element of the spaceship. Although HAL is considered infallible, the system soon develops a destructive life of its own, culminating in the destruction of several crew members.

Some 55 years later, AI has long since made the leap from the big screen to the offices of this world and proven to be a real, extremely useful technology. The public's perception of it seems to be fundamentally different today. Or is it?

"AI revolution in Bavaria: German robot develops a life of its own"

DUP Entrepreneur Magazine, 04.08.2023

"Artificial intelligence: because we don't know what they want to do"

Guest article in the Süddeutsche Zeitung, 18.05.2023

"AI lies to humans: how ChatGPT simply tricks us"

Chip Online, 29.03.2023.

Admittedly, the image of the artificial, autonomous being is an extremely exciting one. The equation of entire systems with AI is also often promoted by software providers for marketing reasons. However, both of these things miss the point. Instead, AI usually only corresponds to individual application parts that are based on more or less complex statistical predictions and interact with other elements. The term "sovereign AI" refers to the rule-compliant and client-capable use of the models provided. This also applies to Helm & Nagel GmbH's AI technologies, which can ultimately be integrated by other software developers into their own applications. Only in this way can sovereign AI be used by companies and consumers in an adapted form.

The path to a hybrid cognitive system

Strictly speaking, AI already existed in the 1960s - albeit in a less useful form. The first chat bot "ELIZA" used a ready-made dictionary in which the program searched for possible synonyms for entered terms. ELIZA then played prefabricated phrases. So it was more or less a guessing process. This was enough to keep authors pondering for decades. The myth of completely autonomous AI was born.

AI level 1: Heuristic systems

Like "ELIZA", the simplest AI models work with the help of heuristics, i.e. certain assumptions based on a small amount of evidence and limited knowledge. This applies, for example, to evolutionary algorithms, Simulated cooling or the A* algorithm all of which are used for optimization problems. The human mind also functions heuristically. All we have at our disposal to perceive reality is our experience, which we often over-interpret. This leads to a distorted picture and makes us bad statisticians by nature.

AI level 2: Knowledge-based systems

At best, technology should compensate for the shortcomings of humans or expand their capabilities. This is why, for example, attempts are being made to transfer as much knowledge as possible to computer programs. These then learn entire sequences of actions based on the knowledge base fed in and manually programmed decisions. Although this is a useful approach that can prevent human error, it is also extremely time-consuming.

AI level 3: Machine learning

The game changer in AI development: Machine Learning enables the automated identification of correlations in large amounts of data. Once trained on a comparable, smaller basis, the models can independently form forecasts and decisions from previously unseen information. In doing so, they use statistical probabilities derived from the training data and are continuously updated - including through human corrections.

AI level 4: Hybrid cognitive systems

Hybrid intelligence is the combination of different types of intelligence. From a technical perspective, this means the joint and flexible use of various AI models in order to solve the most complex problems possible. Today, this mainly includes machine learning, in particular the newer sub-area of Deep learningwhich works with artificial neural networks. Knowledge-based systems often serve as a validation layer, which is intended to avoid errors in terms of business logic.

AI development in practice

As a manufacturer of AI solutions, Helm & Nagel GmbH has climbed each of these steps in recent years. It all began in 2016 with what is now considered a simple task:

Optical Character Recognition

The following information should be extracted from a notary's list of shareholders:

First name, last name, street, postal code, city.

This is a typical use case for so-called OCR technology (Optical Character Recognition). This enables the conversion of optical text into machine-readable formats. Since in this case each copy is the same data type with the same location, the task is relatively easy to solve, even with heuristic approaches.

However, it did not stop there. Since 2018, artificial intelligence has made the greatest progress in its entire history. At the same time, Helm & Nagel GmbH has continued to develop and modify the latest inventions to adapt them to its customers' processes. A decisive turning point came from a specific area of application for the technology.

natural language processing

Natural language processing is an AI technology that aims to process natural language. In 2018, by far the most powerful AI systems to date saw the light of day. NLP Models GPT-1 and BERT were the first representatives of the so-called transformers, which use a special system architecture developed by Google. This enables modern language models to generate coherent texts and process complex sentence structures. 

For Helm & Nagel GmbH, this progress meant that text captured by OCR could now also be processed in terms of content - an important milestone and the basis for data-based knowledge acquisition.

Computer Vision

However, most documents contain more than just text and linguistic content. Optical elements can be combined with Computer Vision a technology based on complex neural networks. Their quality has increased significantly in recent years. This has made high-precision image recognition possible in a wide range of applications. This also applies to the layout information of a document.

Equipped with these three basic technologies, Helm & Nagel GmbH was now able to process all the key dimensions of a document using AI. However, linking various individual models can be complex and time-consuming. Added to this is the training and fine-tuning for different formats. But research has a solution for these problems too.

Multimodal language models

The latest generation of NLP models is able to analyze visual elements such as graphics, images or videos in addition to purely linguistic content. This makes it possible to process diversified content without necessarily requiring multiple model integration. There is also a new method called "instruction tuning", which multimodal language models to date unseen task types prepared. In many cases, this eliminates the need for fine-tuning for special data formats.

This also represents the latest technical tool in Helm & Nagel GmbH's repertoire. Even if the technology is not yet fully developed, it can already offer partial benefits. It is clear that the potential here - especially for document processing - is far from exhausted. However, the use of these elements alone is no guarantee for successful AI applications. Equally important are reliability, security, fairness, resilience, transparency, explainability and data protection. And let's not forget the most important factor of all:

Human-in-the-loop

Even if AI can carry out actions and make decisions that were previously reserved for humans, the technology often comes up against professional limits. To prevent errors, it is up to humans to monitor the processes and correct them if necessary. This not only leads to greater (data) security, but also improves the accuracy of the models. Each human correction can be seen as a new data point that is given a high weighting in future forecasts.

From Helm & Nagel GmbH's point of view, the result of all these techniques and concepts was the product launch of Konfuzio. The history of its development, which took place close to the general development process, is based on a deep technical understanding and a continuously optimized infrastructure. This is reflected in data-based findings and the automation of highly differentiated processes.

Conclusion

Every technology is developed with the aim of generating a practical benefit. In doing so, it adapts more or less to human needs. This perspective is particularly complex in the case of artificial intelligence, as the changes brought about directly affect people's scope of action and even replace parts of it. In addition to great potential for growth, there is also a certain risk - e.g. in terms of data security - that cannot be denied. However, a scenario like the one in "A Space Odyssey" is out of the question. When we talk about complex systems having a "life of their own", the degree of explainability is simply too low and needs to be increased through intensive testing. Data-based models are simply not capable of developing their own agenda. In addition, access to the real world is severely limited.

Both factors can be controlled one hundred percent by humans: They define the database. He decides against autonomous driving without a steering wheel. They develop hybrid cognitive systems that can be adapted almost at will and seemingly create resources out of nothing. Various techniques such as NLP or computer vision are used, which are modeled on central dimensions of human perception. These developments are progressing at an incredible pace - faster than any transfer of knowledge, change of attitude or structural adjustment. It is therefore also the task of media communication to support the most appropriate and profitable use of artificial intelligence.