Adaptability: The True Mark of Intelligence

Self-driving cars were supposed to be here already. Here is why they aren’t 

Many Mission critical applications are characterized by the ever-changing requirements and dynamic environments. From cyber defense, banking, autonomous driving, to robot assisted surgery, adaptive AI, is a necessity for these applications that are essential to our well-being. They must be designed with a versatile and self-learning AI to understand and intelligently react, in real-time, to adversaries and unexpected events. Unfortunately, today’s AI systems, including the hyped “deep learning”, lack the essential features of reasoning and adaptive learning. They are only applicable in limited, supervised learning tasks, defined by rigid rules, in extremely well-defined and fixed environment such as playing Chess or Go. The real world, is a world where the rules are changed during the game. Adaptive solution will for example, detect that the Chess board had become larger, the rooks could now move like bishops or that winning is no longer to checkmate your opponent’s king but to capture all his pawns! We need an AI that could understand what changed and successfully adapt to it.

Imagine an autonomous car driving around the "Arc de Triomphe" in Paris, France. There is a saying in Paris, the real driving test does not take place in front of the driving inspector, but the first time you drive along the Arc de Triomphe. The traffic is continuously heavy, the priority rules are reversed as the drivers already engaged must give way to the cars which are entering the roundabout. Motorcycles whip between double-decker tour buses and trucks, Parisians merge assertively from the outside of the circle to the inside, while other cars, trucks and buses tries to exit towards their boulevard of destination. The sound of horns and screeching brakes is continuous. How could an autonomous vehicle with current non adaptive, zero reasoning AI, handle the sudden changes in trajectory, forced passage, sudden acceleration, braking, etc.? Meeting these requirements raises daunting challenges for the current limited AI solutions as they can’t adjust to real-life environments or make any decision in the presence of adversaries, unforeseen inputs, and unpredictable environments ......The real world of the Arc de Triomphe.