How AI systems develop superhuman abilities through Machine Learning.
Why should you be concerned with this issue?
Currently, we are handing over control from humans to computers in many places, with a strong upward trend. Take air traffic, for example: in the near future, pilots will only take control in a few extreme situations, as the plane will be controlled by autonomous, self-learning AI systems.
The same applies to driving, diagnosing illnesses, working on construction sites and much more. The issue around AI, machine learning and neural networks has long since ceased to be about small niche applications, but rather the fundamental way in which our society functions, because new business areas, new jobs and completely new ways of life and living realities are emerging around these new technologies. This is a development of enormous proportions that we humans should actively shape in our own interest.
For example, the question arises: What exactly is the role of us humans when our machines become increasingly intelligent and take over our tasks bit by bit?
An obvious first impulse might be to think of humans as a higher-level controller. Someone who monitors and controls the machine. But what if the competence of the machine is exorbitantly greater than that of the human and the circumstances are so complex that the human intellect is simply no longer sufficient to control the situation? Human nerve impulses can travel along neurons at 30 cm per millisecond. Computer signals along a copper wire manage the same distance in a nanosecond. That's a million times faster.
Do we really think we can assess and handle sudden dangerous situations on the road better than a driving supercomputer with laser sensors, ultrasonic sensors and 360-degree computer vision that can capture every detail in all directions? Equipped with an artificial brain that is constantly learning and whose signal transmission works a million times faster than our own? Trained with the expertise of millions of kilometers driven, the experience situations of all vehicles worldwide, which are fed into the central computer in real time and then updated in the vehicles?
It doesn't take much imagination to understand that we are going through the greatest technological development in human history.
Some people are critical of this development and may even associate it with fears. Others are curious and are looking for opportunities to shape their own career or, as entrepreneurs, to lead their own company into new, profitable hunting grounds. In any case, one should know what exactly is happening here in order to be able to make the right decisions for oneself.
So let's get into the topic of machine learning together, the learning and training method with which the outstanding performance of AI systems is made possible.
How does machine learning work?
Machine learning is modeled on human learning. It works fundamentally different than the classical programming of a computer. Don't worry, it won't get complicated and abstract here. The topic can be explained with simple examples, understandable for everyone. The classical programming of a computer is based in principle on many if-then rules. IF someone presses this button, THEN print the document. IF someone clicks on this icon, THEN open the program and so on. Of course, the whole thing can happen disconnected from human actions. For example: IF it is 12 o'clock, THEN create a backup. So we humans tell the computer via programming what to do in specific situations. So the computer follows our instructions.
Machine learning is based on a completely different concept. For example, the computer learns by observing human behavior. Without knowing it, for example, you train Google's artificial intelligence every time you run a picture-based security query and click whether there's a bike or motorcycle in the picture, or a truck or a bus. In this way, the Google AI learns how you (and the rest of the billions of users) rate the image and then draws conclusions. At a certain point, the AI has then learned enough and is smarter than humans. The same is true for driving situations in a TESLA. Is the yellow light in the sky a traffic light or the moon? Are the cracks in the asphalt a concern or not? Human behaviors are an important building block in training artificial intelligence. Of course, machine learning doesn't just involve observing people. Sets of rules, laws of physics, and more also play a supporting role in training AI systems. Simply put, you give the computer learning material on the basis of which it can train itself. This is exactly how it works in the human development from a baby to an independent living being. At some point, human intelligence, i.e. our brain, has learned enough to be able to act independently.
Of course, errors can occur in this process. A prominent example is a test in which an AI was to be taught the difference between huskies and wolves. The AI observed human trainers who classified each image as either a husky or a wolf. However, when the AI was then put to work, it incorrectly classified a husky as a wolf. The researchers then turned the AI around and had the computer show them, so to speak, what characteristics were used to determine that it was a wolf. The result surprised and shocked. The AI did not pay attention to the animals at all. On all the learning materials there were wolves in the snow. So the AI thought, IF snow, THEN wolf. A powerful example of how quickly we humans think we are doing the right thing, but misjudge situations, which can then lead to errors in execution that may have fatal consequences. Misclassifying an image is not tragic in itself, unless it is the image of a camera in a car, which is part of the handling of a dangerous situation on the road.
The subject is complex and today's systems are still immature in some places. But AI systems have the potential to far exceed our human abilities. Let's take a look at some impressive examples from the current state of the art.
How an AI uses machine learning to convert monochrome night vision images into real images.
Our human eyes see only a certain part of the color spectrum. Infrared light, for example, is not perceptible to us. If we were in a room flooded with infrared light, the room would appear pitch black to us. However, there are special cameras that can capture light that is invisible to us. The result are monochrome (mostly green) images, which we know from night vision devices.
Researchers at UC Irvine have developed an AI (artificial intelligence) that is able to colorize night vision images deceptively using machine learning, according to Interesting Engineering. To do this, the AI was fed footage in true color and night vision footage and trained to develop a mechanism to colorize even when only the night vision footage is available. The trained algorithm can therefore calculate what the image looks like in true color on the basis of monochrome information and display this immediately. The result is night vision images in complete darkness that can be displayed as if it were a normal camera image taken in broad daylight.
Thanks for reading Frontscale by Christian Sauer! Subscribe for free to receive new posts and support my work.
What the diagnosis of skin cancer has in common with the detection of construction defects.
Today, there are already very good AI systems that support dermatologists in their diagnosis. These systems work on a similar principle to the night vision example above. The AI is fed with thousands of images. Based on the images and a diagnosis associated with the image, the AI learns the principle by which an image represents skin cancer. If the AI has been fed enough material, the computer's capabilities quickly exceed those of most human specialists. Surely you would also feel very comfortable if your local doctor was assisted by a computer system trained by the best specialists in the world.
Of course, this approach can be applied to other industries. In the construction industry, an AI that is properly trained can detect any construction defects, mold growth or storm damage. The principle of machine learning is based on humans using image and data recognition to train a computer to take on specific tasks and thus significantly increase the effectiveness and/or efficiency of certain processes, depending on the topic.
How an AI learns to fly in flight.
Back in 2016, an article was published in the Economist describing the development of sophisticated autopilots for commercial aviation. These systems could already do almost everything that human pilots could do. Taking off and landing in the most difficult conditions, route-finding, and approaching distant destinations. To mitigate the risk of mechanical failure, three separate systems were installed. Only when these three systems agree does the autopilot perform an action; otherwise, the helm is handed back to a human.
A software company recently considered developing a new generation of autopilots. Similar to the AlphaGo supercomputer, the system could "learn the art of flying" over dangerous situations while faced with just that. As a quick reminder, AlphaGo was fed the rules of the game "Go." "Go" is the most complex game of mankind, significantly more multifaceted than chess. Within one day, the computer was able to beat the best human players and, as part of the learning process, developed moves and tactics not found in two thousand years of human history.
So, using this technology, autonomous supercomputers could soon take control in airplanes, capable of making faster and better decisions in dangerous situations.
The evolution that can be seen across the different topics is as follows: AI-based systems have evolved through various stages of evolution in the past, and this evolution is currently continuing. It started as support tools for human work, moved on to semi-autonomous systems that do the work while being supervised by humans, and is now currently evolving into fully autonomous systems that will, at some likely not-too-distant point in the future, do without humans altogether. A brief excursion into the world of science fiction shows that the solution probably lies in defining certain rules and laws that robots and AI systems (obligingly) have to abide by. I refer here to Asimov's laws:
A robot must not (knowingly) harm a human being or, by inaction, (knowingly) allow harm to be done to a human being.
A robot must obey orders given to it by a human being - unless such an order would conflict with rule one.
A robot must protect its existence as long as that protection does not conflict with rule one or two.
Please ask yourself how likely it is that intelligent beings always stick to rules and laws. And whether machines that can write their own code will leave the man-made laws in the next update is also a moot point. On the other hand, AI systems could perhaps finally solve complex problems for which we humans are not intelligent enough. One or two topics immediately come to mind.
So will the future be good or bad?
In my opinion it has a lot to do with how we WANT to look into the future.
I am an optimist and believe the future will be really good.