How far from strong AI are we?
Looking into the future of AI
In 1987, the US series "Star Trek: The Next Generation" introduces the Starfleet crew member Data, an AI in the form of an android. With the intellectual abilities of a human, Data worked based on parameters that allowed him to capture and analyze information. Subjective findings, intuition or social competences are not part of his system. Nevertheless, the figure strives to become more and more human-like through intensive training. For over 30 years, Data has embodied the vision of an intelligent machine – a strong AI that acts intelligently and flexibly on its own initiative.
Nowadays, that’s more than just a vision.
Research into AI was already ongoing as early as the 1950s. One of its most famous representatives is Alan Turing, a British Mathematician who is known as the founder of the theoretical information and computer technology. In his paper “Computer Machinery and Intelligence”, published in 1950, he brings up the question of whether machines can simulate the human brain and act at the same level of intelligence. The Turing Test, which he developed to test those capabilities, is still valid nowadays. But to date, no machine has passed this test.
Although there are many examples of so-called ‘weak’ or ‘narrow’ AI already being applied in everyday life – for example, in navigation systems, speech recognition, automated translations – these are primarily rule-based systems that use various methods to solve problems. The more advanced systems, such as voice assistants, self-driving cars or robots – such as those that are being developed by the American company Boston Dynamics – are still far away from the strong AI that can perform at the same level as a human brain. While we can learn facts independently using a few examples and transfer them to new problems, machines today must be trained with very large amounts of pre-structured data – this is known as ‘deep learning’.
All in all, the possibilities in the field of AI are far from exhausted. Researchers are striving to create a form such as the fictional android, Data: a machine that finds its way around the world, continuously learns new things, and can transfer existing knowledge to new tasks.
Did you know?
is the predicted growth of the AI market by 2025.
increase in the number of active AI startups since 2000.
growth in the share of jobs requiring AI skills since 2013.
Interview: Dr. Helmut Linde, Global Head of Data Science & Analytics
Where do we stand today in the field of AI research?
In recent years, AI research has made impressive progress in various fields – for example, in the recognition and synthesis of images or speech, in the motor control of machines, or in games such as chess or Go. Essentially, these advances are based on a special class of algorithms called deep learning. Its theoretical foundations are quite old and are based on ideas from brain research from the middle of the last century. However, the advances of recent years have been made possible by ever faster hardware and - thanks to the Internet - ever better access to huge amounts of training data. However, deep learning still has serious limitations: the dependency on training data, but especially the difficulty of transferring what has been learned into a new context.
What comes after Deep Learning?
I think that you have to completely rethink machine learning in order to overcome the previous limitations. On the one hand, we need input from the neurosciences - the neurosciences of the 21st century! - because so far, human or animal brains are the best proof that a physical living system can be much more intelligent than today's algorithms. On the other hand, we lack theoretical foundations: just as mathematicians have developed "probability theory”, on which statistics are based, we need a "theory of intelligence" with which we can develop new algorithms.
How far are we from an AI in the form of an android? Is it fiction or almost reality?
It is practically impossible to give a serious time estimate about it. Most experts agree that in principle it is possible to build such a strong AI. One can only speculate about when this will be achieved. But most AI researchers seem to think in decades rather than centuries, and I would agree.
How do they assess the social impact of successfully passing the Turing test?
The social impact will be greater than any other change humanity has ever experienced in its history. A strong AI could do almost any job at least as well as a human being and even much better. Furthermore, like any other software, it could be multiplied with very little effort. The potential of such a technology for economic and scientific progress is simply immeasurable. The big question is whether humanity will be able to use this power profitably for everyone. How will one distribute material abundance and secure the existence of those whose human labor is no longer needed? What are the psychological effects on society if most of its members no longer need to work? And is a strong AI even permanently controllable or do we have to fear that we will lose control over our technology?
To what extent do you at Merck KGaA, Darmstadt, Germany contribute to current research in the field of AI and how do you cooperate with other institutions (beyond your organization)?
On the one hand, of course, we do applied AI research, i.e. we try to use or adapt existing algorithms cleverly in the context of our industry so that we can derive economic benefit from them. This involves, for example, automatically analyzing medical imaging data or finding new chemicals with certain pre-defined properties for our Performance Materials business. In addition, we are currently setting up a small research group that will deal with the basics of intelligence. In essence, we want to understand how an agent - a biological brain or an AI - can get an idea of its environment and how it can abstract from irrelevant details. For example: how does a child learn so quickly what a cat is, even though each cat looks a bit different and even the same cat is constantly perceived from different perspectives and under changing light conditions? It is said that the brain forms an "invariant representation" of the cat - a kind of data set that describes a cat "in itself" and is unchangeable - i.e. invariant - regarding the many unimportant details. How exactly this works, however, is completely enigmatic. Answering this question is probably an important and necessary step towards strong AI. We believe that experts from brain research, computer science, mathematics and theoretical physics need to be brought together to make progress. Various academic institutions can be interesting cooperation partners here, such as the TU Darmstadt or the Max Planck Institute for Brain Research in Frankfurt.
What do you find fascinating about your work?
I think that the phenomenon of intelligence is one of the most exciting topics of our time. You have to realize how impressively detailed and comprehensive the scientific worldview has become in the meantime. We can measure and understand things on time and size scales that are unbelievably far beyond our direct sensory perception. We can predict tiny elementary particles such as the Higgs boson on the basis of theoretical considerations and then prove them experimentally. The same applies to unbelievably massive and distant objects such as black holes and their gravitational waves. We measure cosmic background radiation that is over 13 billion years old and pulses from femtosecond lasers that last several millionths of a billionth of a second. We now also have a pretty good understanding of how living beings have developed in the more than 500 million years since the Cambrian explosion, and how the biochemistry of their cells works in the nanometer range. And then we have this 1.5 kg organ that we call the brain, that consumes just 20 watts of energy and that we constantly carry around with us - and we actually have no idea how it works...
How are we are using AI already?
From Siri to facial recognition, AI and machine learning are becoming familiar features of everyday life, transforming the way that we prevent, diagnose and treat different illnesses – as well as improving patient care. But they could also save lives, by helping find new and better treatments for disease, faster or by screening for diseases.
The processes behind drug discovery can be challenging. Can AI and machine learning help alleviate some of the traditional bottlenecks, so we can get new and more effective drugs to patients, faster? Read our story how we are using AI for drug discovery.
Medical robots, however, can help by relieving highly-skilled healthcare professionals from routine tasks. They can also perform highly accurate surgery, improve patient rehabilitation or target medicines specifically to where they are needed inside the body. Are medical robots with human-like characteristics about to change the way we diagnose medical conditions? Learn more about PETRA, the social medical robot who can screen for diseases!
Work With Us
Our Professionals Make a Difference!
When great minds get together, they inspire each other. Our collaborative culture proves it. By working a better way—sharing ideas, voicing opinions, giving feedback and lending support—we get better results.View All Jobs