Does true AI really exist?

Artificial intelligence: what it can do today

Ramin Assadollahi

A real-life situation report that shows that successful AI systems can only function with human input.

EnlargeAI has long been ubiquitous - hardly anyone knows exactly where.
© Photo: © Sergey?

Anyone who has been reading about artificial intelligence lately quickly gets a bizarre impression: Today or tomorrow at the latest we will be dealing with autonomously acting and thinking machines - be it threatening killer drones on the one hand or salvific surgical robots on the other. And the production robots will make many of our workplaces obsolete. Not to forget the Artificial Intelligences, which weigh our actions in real and digital life sometimes more, sometimes less precisely against each other and draw conclusions and predictions from them.

Toolbox AI

In practice, people who do research with artificial intelligence only have a tired smile left for these scenarios. They are about as far from reality as a weekend trip to the nearest solar system.

Today, artificial intelligence can be compared to a well-equipped tool box. Just like screwdrivers, pipe wrenches or hammers for screwing, loosening or impact applications can be found in it, there are route search engines for Google Maps, web search engines such as Bing or image classification engines such as Clarifai in the area of ​​data interpretation using AI.

And - to further spin the analogy - just as one pursues various activities when building a house (e.g. building or tearing down something) and needs the associated special tools and materials, there are also fundamentally different types of activities (e.g. classifying or marking) and materials in the field of artificial intelligence - mostly in the form of data or data-producing sensors.

Saving lives with AI?

A specific AI activity is the preparation of forecasts and possible processes based on structured data. An example from medicine: If, after an infection, a patient suffers from a fever that falls slowly and then suddenly rises rapidly, an artificial intelligence can decide, taking other factors into account, to immediately sound the alarm - even before the next regular human check - and so save valuable hours.

All data streams and signals in this context can be clearly and unambiguously differentiated from one another and are thus structured. All that is required is constant monitoring of relevant signals and a classification machine that learns to distinguish tricky from regular situations and can assess different situations better and better through continuous use.

This does not replace regular inspection rounds and thus also does not replace human responsibility in the hospital. But AI can meaningfully support staff and save one or the other life.

Understand unstructured data

The real challenge for AI tools, however, is to interpret data that is less structured or not at all. This includes, for example, image or text data. A picture can contain all sorts of elements - from a poorly photographed PowerPoint slide with a table of data points to a picture of a ball of three kittens.

A text, in turn, can simultaneously describe a specific experiment and at the same time contain a brilliant idea for decisive advances in medicine, which in turn can be further described in another text.

There is still no artificial intelligence that can look at pictures or read texts and understand what it is seeing or reading on its own. All of the breakthroughs in the use of AI in recent years are based on the fact that one (or many) people prepare training data for the AI. This means that you specify in advance what can be seen on a picture, or passages in a text that are of interest are marked. All the AI ​​then learns is to recognize such or similar structures in other images or in texts.

There are some very interesting methods that pre-analyze images and texts without human intervention and recognize structures. These predefined structures are then more accessible to AI tools, which can accelerate the process and improve the results. But it is and remains only a kind of statistical preliminary evaluation.

Human Artificial Intelligence Teachers

In other words, even the most advanced AI cannot do without the human teacher. This is the bad news.

Reading tip: Artificial intelligence is not interested in us humans

The good news is that AI tools are getting better at understanding and generalizing human input.

Interestingly, however, one area for AI tools has hardly been developed until today: language. Understanding and reliably classifying which facts are really expressed through language in a document, which contexts are highlighted and what can be concluded from them, remained an impossibility until recently.

Only recently have some approaches existed with which, in practice, certain types of facts, relationships or even entire specialist articles can be reliably and automatically extracted in a sufficiently targeted manner. As a result, it is no longer necessary for the individual researcher in the pharmaceutical industry, for example, to read thousands upon thousands of articles on the subject of Alzheimer's disease if he just wants to know which active ingredients are currently being researched in this area and where which breakthroughs have been achieved.

Easy training of AI

Since successful AI systems can only function through human input, it is also important to examine the aspect of input recording in more detail. If the training of an AI system is cumbersome, lengthy and the successful outcome is questionable, no organization will be able to carry out this process long enough. The training must therefore be simple and success must be realizable quickly. As a result, the focus in development is shifting to making the production of training data as simple as possible and an AI can even indirectly draw information from incidental use that makes it better.

Solutions from the AI ​​factory

In the meantime, instead of fully trained systems, it is even possible to develop individual AI tools that are adapted to the personal needs of the user through direct input and thus become better. This is the next generation of AI systems - a kind of "AI factory", such as that offered by ExB Labs. Such a factory is a further development of classic models in which users train their individual system using a central system - regardless of whether it is health, automotive or other data.

To stay with the initial example: In the future, internal and therefore “personal” AI systems will work in hospitals, which learn from the processes running there and thus become steadily better. At the same time, these systems allow (anonymized) knowledge to flow into higher-level systems, which in turn pick out statistically weak signals from a global perspective and can detect unexpected connections, for example between active substances and diseases.

Thus, in terms of the overall effect, at some point we actually come to a seemingly wise overall system that can draw conclusions from millions of facts that humans would not have been able to make.

But in the end, this AI will remain a tool too. And a person makes the final decisions.

Reading tip: evolution without us: will artificial intelligence kill us?