The secret life of artificial intelligence
In March 2016, something unexpected happened: Lee Sedol, one of the world’s best Go-players, became the first world-class professional to be beaten by Alpha Go, Google’s artificial intelligence. Go is considered the most complex and sophisticated board game of all and many AI experts
were not expecting this event to happen for at least another 10 years. For the first time a technology emerged that would have the potential to replace not only the physical labour force of man, but the highest and most exclusive human good of all: the ability to think.
One consequence of this development has been that Google has now placed artificial intelligence at the heart of their existence. Instead of “Mobile First” the new mantra is now “AI First” (Artificial Intelligence First). But what does this actually mean? Software has always been present in Google’s DNA. However, until the beginning of the 21st century, GOFAI (good old fashioned AI) was largely based on symbolic logic: here reality is being mapped top-down via a set of rules (i.e. symbolism) and traditional programming logic.
Put simply, this corresponds to “if-then” rules as you may know them from Excel: positive numbers are being marked green and negative numbers in red. This logic works well as long as the complexity of the world in which AI operates is manageable. But it fails miserably as soon as complexity increases exponentially.
Symbolic AI had led to initial successes, but previous projects based on symbolic logic such as image recognition, automated translation and speech recognition were of limited use because human reality is too complex to be captured by rigid rules. But how was Alpha Go suddenly able to beat people who had been masters for decades? How could it be that Alpha Go (or the updated version Alpha Zero) was able to conjure strategies that even the Go masters could no longer comprehend?
Journey into our past
If we want to get to the bottom of this mystery, we have to go far into our past, very far! Our journey goes back into Earth’s history, long before the first man walked on our planet. At some point there were the first unicellular organisms, then multicellular organisms and later (about 580 million years before our time) the first jellyfish emerged. They developed a network of nerves and were therefore able to “feel” their environment. They could absorb information and this helped them to survive better in an uncomfortable setting. Evolution took its course. And here we are at the central point: this principle of evolution, i.e. to filter out certain qualities through trial and error that are beneficial to reproduction and survival, is the core idea behind the success of Alpha Go.
Certain behaviours or bodily functions contribute to the survival of a species and then reappear in the next generation. Specific characteristics have thus been enhanced because they have led to the survival of the organism. Reinforcement Learning, a method in AI that has contributed significantly to the success of Alpha Go, offers a similar idea: an agent moves in a its environment and learns by trial and error which behaviours are appropriate and contribute to a predefined goal. These are then reinforced or not.
So instead of programming a set of rules top-down (which would have been a questionable undertaking anyway in a complex board game like Go), the AI learns how to win bottom-up by completing millions of games at high speed. The first step is to learn from people playing Go and eventually the AI starts playing against itself. So it trains its “brain” by playing up countless game variations and thus becoming a master of its genre at an exponential speed. This experiential way of learning is closey resembling the way small children learn to walk, ride a bike or articulate themselves. What the AI researchers have started simulating was nothing less than the functioning of the human brain. We feed the system with inputs via neural networks, which you can imagine as an artificial brain. These inputs are then propagated through multiple layers and eventually produce an output, which helps the AI to recognise and learn patterns. This new paradigm is also known as machine learning. Machine learning algorithms perform certain tasks without having been explicitly programmed for them. In a sense, they learn from experience.
Evolution on steroids
Accordingly, there has been a paradigm shift in AI research: we are moving from symbolic AI to significantly more effective machine learning algorithms.
There are two fundamental factors, which contribute to this process: computers operate at a much faster speed than the human brain and we started simulating the functioning of the human brain. Taking these two ideas together, it should become obvious that we are on our way of creating something profound: we are designing evolution on steroids. This is the real reason why AI is mutating into the most important technology of the 21st century and we all should pay attention.
This means that just as a child learns to walk and slowly discovers the world through trial and error, AI learns to distinguish cats and dogs, to navigate in a machine production environment or to translate poetry from German into English. Symbolic programming logic (Top- down) is now more and more replaced by “learning through experience” (Bottom-up).
AI First at Google thus means that products and processes are progressively generated by machine learning instead of more traditional programming logic. This is all well and nice for young digital companies such as Google and Facebook. However what could this mean for more established corporate players? Haven’t many bricks and mortar businesses just started digitising recently? And they are now being forced to embrace the next trend?
Leapfrogging into our future
The English language presents a nice term for this dilemma: “leapfrogging”. It is mainly used in economics and business studies and describes the idea to leap over certain technologies in order to be radically innovative and to overtake existing market leaders. In developing countries, many consumers have jumped from using no technology at all to mobile devices as they simply could not afford anything else. Perhaps it is time for the business world to “leapfrog” and put experiential AI at the center of their efforts instead of designing their businesses around a set of rules.