Part 1: How does artificial intelligence learn?
Text: Dr. Jakob Vicari | Reading Time: approx. 15 min. MAN, THESE MACHINES! Artificial intelligence has become an integral part of the technological landscape in our factories and our homes: It communicates with us in the form of smartphone assistants, analyzes algorithms, produces medical diagnoses, and checks manufactured components. AI not only breathes life, but also value, into our data and information. Intelligent systems are constantly learning, and are capable of doing so far faster than the humans that use them. The technology is set to change our lives – and even to surpass the achievements of the human race. We can now communicate not only with the other people around us, but also with the objects and machines in our environment – machines that are capable of understanding what we say, as well as our moods and our actions.
CHAPTER 1 – How does artificial intelligence learn?
CHAPTER 2 – How does artificial intelligence work?
CHAPTER 3 – How does artificial intelligence communicate with us?
CHAPTER 2 – How does artificial intelligence work?
CHAPTER 3 – How does artificial intelligence communicate with us?
PART 1How does artificial intelligence work?What happens when a machine becomes more intelligent than the person operating it? Berlin-based startup “Twenty Billion Neurons” is aiming to find out.
Says company founder Christian Thurau. And it’s a job he takes seriously: He spent 15 years researching AI and published over 50 scientific articles before founding his own startup three years ago. In spite of his background, Thurau remains fascinated by the pace of technological change.
UNDERSTANDING VISIONAdvancements in processor performance and breakthroughs in machine learning have fueled the recent boom in artificial intelligence. At the core of many technologies are artificial neural networks: structures of artificial neurons that adjust the weight of their connections as they learn, like neurons in the human brain. When data is fed into this kind of network, the network learns by making new connections, deleting inaccurate information and adjusting the weighting attached to what it already knows. The challenge lies in organizing the huge volume of artificial neurons appropriately and training these neurons in the most logical way.
Artificial intelligence is increasingly taking on more complex tasks. The experts at TwentyBN believe that the next breakthrough will come in the form of artificial intelligence that understands videos. Thurau explains: “We’ve been able to show that artificial intelligence understands the laws of physics that apply in our world, and can recognize what is happening in visual scenarios.” Smart analysis systems of this kind – previously unfeasible due to the vast volumes of data involved – could also offer a major advantage in autonomous vehicle or production applications. But the answers aren’t there quite yet.
FROM “WHAT’S THAT?” TO “WHAT'S HAPPENING THERE?”A man reaches for a cup, drinks, holds the cup up to the camera and shows that the cup is now empty. This is an action that we’re programmed to understand from a young age as humans. But artificial intelligence doesn’t have the benefit of a childhood to learn what it needs to know, and it doesn’t have the “common sense” that we all take for granted. It does not understand that a person it has detected who then disappears behind a car will re-emerge again on the other side. Recognizing actions is one of the greatest challenges for artificial intelligence systems to overcome. The sequence of movements – as well-trained as it may be – comprises dozens of steps.
“Our cortex is trained to analyze consequences in the context of other consequences,” says Christian Thurau from TwentyBN. Video data is much more complicated than image data – and different people may complete the same actions in a slightly different order. Even being right-handed or left-handed can change the sequence.Installed in an industrial robot, the technology could determine what a human colleague is doing and respond accordingly; it could also be deployed in autonomous vehicles or in care robots.
SECOND BRAIN FOR HIRETwentyBN hires out its artificial “second brain” to companies. Organizations that take advantage of this offer gain access to a neural network: A structure of artificial neurons trained to complete tasks and learn in the same way as people. This neural network was also the inspiration behind the company’s name, which is derived from the 20 billion nerve cells in the human cerebral cortex. Thurau can train his artificial neural network to recognize images, translate text, or answer questions.
Patterns can be found all around us, whether in component production or robotic movements. Thurau comes from a research background and previously worked for the Fraunhofer Society. “A production robot could use patterns to learn to configure itself,” explains Thurau. “Or a camera could be trained to detect paintwork damage on car doors.”
THE TRAINING PROCESSAn invisible army of crowd actors is working behind the scenes to film everyday actions, which will be used to train artificial intelligence. TwentyBN is just one of many startups that pays these actors a few cents per clip of everyday people doing everyday things. Thousands of people have filmed and submitted clips of various actions to teach artificial intelligence about these concepts – be it reaching for their shoes, operating a remote control, or simply winking. “What we’re aiming to do is teach common sense,” says Christian Thurau. “Those things that are simply self-evident to us.” It would seem that training is the only obstacle standing in the way of communication between human and machine.
Part 2: How does artificial intelligence work?
PART 2How does artificial intelligence work?The idea of creating machines that think like people has been around for some time. The concept of neural networks was first proposed by Warren McCulloch and Walter Pitts in 1943 at the University of Chicago. However, early attempts to develop an intelligent computer were plagued by frustration. Only with the advent of the Internet did scientists finally have access to the vast volumes of data needed to “teach” artificial intelligence – and it was not until the start of the 21st century that sufficient computing power was finally available to effectively evaluate this flood of data. Although computers have collected huge volumes of data over the past century, artificial intelligence holds the key to unlocking its full potential.
ALEXA’S GREAT-GRANDMOTHEREliza, developed by Joseph Weizenbaum at MIT, is the world’s most famous natural language processing computer. Her intelligence was limited to a few hard-wired rules conceived by Weizenbaum. If, for example, the user typed “I remember…”, Eliza would always respond with: “Why are you remembering this now?”, following the pattern regardless of whether the memory was related to childhood, a pet or an accident. Other sentences would simply be repeated back with the addition of a question mark. Once the initial astonishment faded, people started to notice that Eliza’s intelligence was somewhat limited after three to four questions. But the program marked a shift in the relationship between man and machine: Eliza and her successors taught users how to ask questions in a way that a computer could understand.
THE DIFFERENCE BETWEEN THE LETTER “A” AND A COFFEE STAINTraditional computer programs store their information in compartments, like in a chest of drawers. On the outside of each drawer is the question; inside lies the answer. Machine learning acts as the key to unlock the drawer by organizing information in a network. It allows conclusions to be drawn based on large volumes of data, by identifying patterns and generating its own knowledge based on experience.
As neural networks work in a similar way to the nerve cells in the human brain, they do not require extensive programming. Instead, they receive “lessons” like those taught to children, incorporating as many different examples as possible to highlight the basic principle. To teach reliable identification of the letter “A”, for example, the neural network is shown the letter “A” in various formats, including handwritten and typed in Helvetica, Garamond and Times fonts, printed messily, in bold and upside down.
In the same way as a child learns – partially through guided teaching and partially on its own initiative – how to differentiate between letters and numbers, the artificial networks start by learning to tell a coffee stain apart from a letter, and then progress to distinguishing an “A” from a “Z” – without the “teacher” having to explain all of the differences. If a letter is learned incorrectly, it is virtually impossible to relearn.
MORE THAN JUST A GAMEIn the public sphere, people became aware of the progress being made in artificial intelligence by the fact that computers were starting to beat world champions in various board games. In 2016, the AlphaGo program, developed by Google-owned company DeepMind, beat South-Korean pro gamer Lee Sedol in the board game “Go”. One year on, the program’s successor, Alpha Zero, learned chess, Go and shogi in just 24 hours and went on to emerge victorious when pitted against the then-world-champion programs.
INTELLIGENT STARTUPSA good indicator of the maturity of a technology is the number of startups working with it. Increasing numbers of startups are declaring that artificial intelligence technology is market-ready. The two-dozen pioneers who have set up AI companies in Berlin to date are convinced that the technology can be deployed to produce software that is smarter than we ever thought any computer program could be. February 2018 saw the founding of an official German Federal Association for AI [http://ki-verband.de/], which strives to promote human-centered and useful applications of AI technologies. Major technology companies such as Google, Facebook, and Apple are buying up AI specialist companies on an almost weekly basis, to ensure that their businesses are equipped for the future. Digital association Bitkom estimates that the global market for artificial intelligence software will be worth 21.2 billion euro by 2020.
MULTIPLE LEVELS FOR GREATER DEPTHDeep learning” has been a key driver of the boom in artificial intelligence.In “deep learning”, multiple neural networks are layered on top of one another. The greater the degree of connectivity between these networks, the more complex the conclusions that AI can draw. Face recognition technology, for example – a culmination of the rapid technological progress made in recent years, particularly in the field of deep learning – essentially works as follows: At the first stage, the system learns to recognize contrast between dark and light; at the second stage, it learns to identify edges and simple shapes; at the third stage, it learns more complex shapes, and finally, at the fourth stage, it learns which of these shapes are characteristic of faces.
FROM DOGS TO SCREWSDog breeds are often used as a training exercise for AI systems. Neural networks can easily distinguish between a rough-coated dachshund and an Entlebucher mountain dog. Dog genealogy is easy to train into the program, and pedigrees are surprisingly intricately structured – with a level of complexity that often poses a problem in factories handling many different types of screw, for example.
LOCATING THE ERROROnce a certain level of complexity is reached, it is not always possible to predict the response of an artificial intelligence system to a particular problem – not even for the people who developed it – which can make troubleshooting a challenge. We are approaching a new age in software. An age in which software learns from us how to solve problems, and develops based on what we need it to do – and will sooner or later become a more effective problem-solver than the foremost human experts.
“The main benefit of artificial intelligence in production is that we can use it to optimize processes to a degree never seen before,” says Dr. Walter Heibey, Head of Technology Innovation at Sanofi. Complex planning tasks no longer need to be left to a human struggling with doubt and second-guessing decisions. The machines themselves can arrange production. OPTIMAL USEThe way in which humans communicate – or in some cases fail to communicate – is what can often make life difficult in business. “Machines communicate without reservations – unlike people,” says Heibey’s colleague Jan Tessnow, an expert in smart manufacturing at MHP. For Heibey, it is important to emphasize that AI is not about replacing human workers. The technology becomes truly useful when multiple intelligent systems communicate with one another, as they are starting to do in production and logistics settings, where an intelligent robot-controlled vehicle may collect and re-order parts from the warehouse – bringing together the traditionally separate areas of production and logistics.
FORGIVE AND FORGET“If I no longer need to check whether the special component I need has been delivered – because it’s there, right in front of me – that’s the kind of magic that will enable artificial intelligence to gain widespread acceptance,” says Heibey. On the other hand, we must be prepared to forgive the computer’s mistakes. Artificial intelligence is not infallible – and these are the kinds of challenges faced where intelligent systems are deployed. Tessnow talks about these challenges in terms of a “maturity level” that companies and people reach. “Artificial intelligence will only work in companies where people talk to one another rather than entrenching themselves in silos.”
Part 3: How does artificial intelligence communicate with us?
PART 3How does artificial intelligence communicate with us?Whether on the Audi production line or in the form of a health assistant at the patient’s bedside: Artificial intelligence systems have become an integral part of our environment. We don’t always recognize them at first, if they’re working away in the background to optimize processes. When we are aware of them, we use them naturally and without inhibition – but we also have high expectations.
The major challenge for artificial intelligence systems tasked with interacting with people does not necessarily lie in delivering the right result. We already have faith in the computer’s ability, and we even accept that, in some situations, its performance is superior to ours. However, it is important that computers are open with us about how they make their decisions, and that they interact with empathy. Anyone who has ever argued with their GPS about the best route or screamed at their voice assistant for playing the wrong song will tell you that.
THE EMPATHETIC MACHINE“A doctor can glance at a patient and make an assessment of their health. Ada does this based on four to five questions – sometimes a few more...,” says Martin Hirsch, doctor and creator of Ada. Ada is a “smart health companion”: an app that provides guidance and helps its user feel better. It follows a process not unlike a visit to a family physician: The user describes their symptoms and answers a series of questions based on their own experiences and feelings. Ada then uses artificial intelligence to come up with an initial diagnosis. “It is really helpful when patients take our analysis results along to their doctor,” says Hirsch.
Dr. Martin Hirsch PhD, Founder and Chief Scientific OfficerMartin grew up in a family of scientists, and is a grandson of the celebrated Nobel Laureate Werner Heisenberg. He studied theoretical medicine before working as an independent researcher. His special interests include cognitive neuroscience, semantic knowledge representation and the use of technology to support human thinking and decision-making.
“Trust is based on interaction,” says Hirsch. “It is established in the first few seconds of a meeting.” With this in mind, the team dedicated a great deal of time to Ada’s approach. “She balances keeping her distance with an air of familiarity; she’s not excessively friendly. She’s more of a quiet introvert,” says Hirsch of Ada’s character. Hirsch and his team worked on Ada for six years. They learned, among other things, that artificial intelligence doesn’t ask the question that would provide the most information at any given stage, but instead opts for a question that links back to the previous question. “Users want a solution that acts and feels like there’s another person on the other end of the line.”But what does that mean for us and how we define ourselves as human – if a voice assistant can recognize and respond to our moods? “We often receive feedback from users saying: Finally, someone is listening to me. Generally, doctors these days don’t have the time to spend five minutes taking a history,” says Hirsch. Ada is not designed to replace physicians, but to support and complement their work. Our brains organize their storage capacity highly efficiently. “The human brain’s capacity to store data is limited. But data volume doesn’t pose a problem for computers,” says Hirsch. In the case of rarer disorders or illnesses, Ada’s performance surpasses that of doctors. Hirsch believes that people distrust the creators of the program rather than the AI technology itself.
If the user grants permission, Ada can use Apple Health data to see how the patient is sleeping, how much exercise they are getting and which restaurants they visit – meaning that the app has much more data at its disposal than a doctor. The startup is now working on a way to integrate genetic analysis data into the program. The patient bringing an initial diagnosis along to an appointment opens up a new level of freedom for physicians.MUTUAL UNDERSTANDINGThis freedom is also an important factor for Walter Heibey from Sanofi: “People need to learn to put their trust in the algorithm. If you can succeed in doing that, you allow yourself the scope to concentrate on more important things.” Intelligent computer systems won’t take our jobs – but they are likely to change them. In automotive production, for example, intelligent algorithms can change the sequence of components on the conveyor belt if the truck containing the required parts is late making its delivery. “What’s important is making the system’s decision transparent to the user rather than it coming across as a dark, mysterious force working in the background.”
MORE THAN SCIENCE FICTIONHaving conquered our day-to-day gadgets and devices, artificial intelligence is now gaining a foothold in industry. AI will be critical as we analyze ever-more-complex data and make predictions using networked systems – working towards the ultimate vision of a self-managed factory. In the vast majority of cases, AI systems will be invisible workers. But if machines can think like health assistant Ada and recognize actions like the technology of TwentyBN, they could become valued colleagues.
In the future, it will be less about what computers can do and more about whether people can work alongside, trust, and share information with their new coworkers.
Scroll down to continue
Swipe to continue
Swipe to continue
MAN, THESE MACHINES!
How does artificial intelligence work?
Introduction Christian Thurau
Artificial intelligence is increasingly taking on more complex tasks.
Installed in an industrial robot, the technology could determine what a human colleague is doing and respond accordingly; it could also be deployed in autonomous vehicles or in care robots.
How does artificial intelligence work?
Eliza, 1964
In the same way as a child learns – partially through guided teaching and partially on its own initiative – how to differentiate between letters and numbers, the artificial networks start by learning to tell a coffee stain apart from a letter, and then progress to distinguishing an “A” from a “Z” – without the “teacher” having to explain all of the differences. If a letter is learned incorrectly, it is virtually impossible to relearn.
Deep learning” has been a key driver of the boom in artificial intelligence.
Once a certain level of complexity is reached, it is not always possible to predict the response of an artificial intelligence system to a particular problem – not even for the people who developed it – which can make troubleshooting a challenge. We are approaching a new age in software. An age in which software learns from us how to solve problems, and develops based on what we need it to do – and will sooner or later become a more effective problem-solver than the foremost human experts.
OPTIMAL USE
How does artificial intelligence communicate with us?
Dr. Martin Hirsch PhD, Founder and Chief Scientific Officer
MUTUAL UNDERSTANDING
In the future, it will be less about what computers can do and more about whether people can work alongside, trust, and share information with their new coworkers.
MAN, THESE MACHINES!
How does artificial intelligence work?
Introduction Christian Thurau
Artificial intelligence is increasingly taking on more complex tasks.
Installed in an industrial robot, the technology could determine what a human colleague is doing and respond accordingly; it could also be deployed in autonomous vehicles or in care robots.
How does artificial intelligence work?
Eliza, 1964
In the same way as a child learns – partially through guided teaching and partially on its own initiative – how to differentiate between letters and numbers, the artificial networks start by learning to tell a coffee stain apart from a letter, and then progress to distinguishing an “A” from a “Z” – without the “teacher” having to explain all of the differences. If a letter is learned incorrectly, it is virtually impossible to relearn.
Deep learning” has been a key driver of the boom in artificial intelligence.
Once a certain level of complexity is reached, it is not always possible to predict the response of an artificial intelligence system to a particular problem – not even for the people who developed it – which can make troubleshooting a challenge. We are approaching a new age in software. An age in which software learns from us how to solve problems, and develops based on what we need it to do – and will sooner or later become a more effective problem-solver than the foremost human experts.
OPTIMAL USE
How does artificial intelligence communicate with us?
Dr. Martin Hirsch PhD, Founder and Chief Scientific Officer
MUTUAL UNDERSTANDING
In the future, it will be less about what computers can do and more about whether people can work alongside, trust, and share information with their new coworkers.
