The Next Evolution of AI: Transparent and Thoughtful

Think about the following situation: you are driving in a town and see a traffic sign with a top speed of 120 km/h. What would you do? Hit the gas and max out the speed limit? Probably not—because you know that you are only allowed to drive a maximum of 50 km/h in built-up areas.

But what would today’s artificial intelligence (AI) do? It would recognize and correctly interpret the traffic sign without any difficulty. Without additional knowledge, it would conclude that the speed limit here is 120 km/h. It would never occur to the AI that the sign could have been placed on the side of the road by accident or with malicious intent.

Because the AI lacks a typical human ability: reflection. In a situation like this, any human being would reflect and form hypotheses. Or they would ask their passenger for a second opinion. This is exactly what we want to achieve with the “third wave of AI”: The aim is for AI systems to become more humanlike in the future—by taking knowledge of the world into account when making decisions, and asking people for their opinions when necessary.

This is demonstrated by a current example. Today we can already teach AI systems ethical behavior by presenting them with many human texts from which they extract our prejudices—because ethical rules are, in a certain sense, really just prejudices. A contemporary AI would therefore say that the public wearing of a mask is not accepted in Germany. However, it would not take into account that in the coronavirus pandemic, masks are a part of our everyday life. In the future, an AI could search current news and take the knowledge gained into account in its decisions.

In order to better understand the difference to existing AI systems, it is helpful to take a look at the past: during the first wave of AI from 1956 to the 1980s, the intelligent behavior of humans was pre-programmed. The computer was able to draw its conclusions from a large number of if-then relationships with the help of a programmed logic. This is how the first “expert systems” were created, which gave recommendations for action in a specific area of activity.

The second wave of AI started in the 1980s and continues to this day. Today, computers are able to learn from examples—also called “data”—and thus develop “intelligent” behavior. It is no longer necessary for the human programmers to imagine every eventuality and manually pre-program countless rules. Only the learning algorithm is still programmed. Among the best-known successes of such AI systems are the victory of the IBM chess computer Deep Blue over world champion Garry Kasparov (1996) and the victory of the AI software AlphaGo over the world-class South Korean go player Lee Sedol (2016). Today, we all benefit from the second wave of AI—for example through voice recognition in smartphones or cars that can drive to some extent without our intervention.

The third wave is just starting to take off. And it builds on what we have developed in the past two waves of AI. The new approach could be summarized roughly as follows: combining low-level perception (120 km/h) and high-level reasoning (traffic rules) and contextualizing decisions (in town) and communicating them in a human-like way (asking the passenger).

Technically, we use a combination of neural networks, probability models and logic to achieve this—as Nobel Prize winner Daniel Kahneman describes for us humans in his book Thinking, Fast and Slow: the fast, instinctive, and emotional system (neural networks and probability models) works together with the slower, more logical system (program logic) that thinks things through.

One of the important developments for the third wave of AI is “neural symbolic AI”: while the well-known Convolutional Neural Networks (CNN) can recognize traffic signs from the pixels of a video camera, for example, they lack a special ingredient that makes us human: common sense—that intuitive sense of context that seems so natural to us. It enables people to deduce the meaning of a new word, the properties of unknown substances, or even social norms based on a few experiences. Such conclusions go far beyond the available data.

The technology has not yet reached this stage. Today’s neuronal networks are inclined to confuse a school bus tipped over on its side with a snowplow. Humans are different: once we have learned what a school bus is, we have no major difficulties recognizing it even in unfamiliar situations. This is because we humans can abstract, generalize solution strategies, and apply them to similar, albeit different, situations.

Thanks to neural symbolic AI, machines will soon be able to do this as well. In the future, we will use neural networks which, alone or in combination with other methods of AI, can grasp the true meaning of traffic signs, understand them as “symbols” and then apply program logic to “put two and two together”—for instance that you are not allowed to drive 120 km/h in a town.

Thanks to the third wave of AI, machines will soon be partners to humans who think and understand us just like we do. In climate research, for example, they will combine data and models for the atmosphere, the oceans and the cloud system with economic models and biosphere models to create a “big picture.”

Future AI systems will also be much more transparent to us. This is because they will be able to put their decisions into words, abstract from unimportant details, and make it easier for us to understand how they arrive at their decisions: why does the navigation system take this particular route? Why does the financial software recommend exactly this investment? Why does an AI reject a certain applicant and give preference to another? With the third wave of AI we will solve the problem of the “black box” and make the processes transparent.

For the third wave of AI, the players are in the process of planting their flags. Companies in the US, particularly, have already recognized the potential of this development and are currently buying many startups in the field. We are also well positioned in Europe because there are many capable scientists and developers here—but unfortunately less venture capital sloshing about. However, the train has not yet left the station and it is up to us to help shape the future of AI. My message to politicians, industry, and science is therefore that we should create an AI that is more friendly and more useful. An AI for the benefit of all. Let’s ride the third wave!

Guest author Prof. Dr. Kristian Kersting is Professor for Artificial Intelligence and Machine Learning at TU Darmstadt and winner of the first German AI Prize (2019). Since 2020, he has also been the co-speaker together with Prof. Mira Mezini for the new Hessian Center for Artificial Intelligence, in which 13 universities from the state participate. As part of the effort, the state of Hesse is establishing 20 additional professorships and providing 38 million euros in the five-year start-up phase. The focus of the work will be the third wave of AI.

Text first published in the Porsche Engineering Magazine, issue 1/2021.

Text: Prof. Dr. Kristian Kersting

Copyright: All images, videos, and audio files published in this article are subject to copyright. Reproduction in whole or in part is not permitted without the written consent of Porsche Engineering. Please contact us for further information.

You have questions or want to learn more? Get in touch with us: info@porsche-engineering.de