|
|
|
AI in historical perspective
THE idea of machine intelligence was popularized recently by Steven Spielberg’s movie Artificial Intelligence. Spielberg’s fantasy world is not merely wishful thinking but has the potential to become a reality one day. Broadly speaking the field of Artificial Intelligence aims to create intelligence in man made systems. This intelligence can range from low-level insect intelligence to human level intelligence.
Contrary to the common misconception, AI is not limited to the field of Computer Science but also encompasses certain fields of Philosophy, Physiology, Linguistics and Psychology. The main idea behind AI is that human beings and animals acquire knowledge from their environment and manipulate it in an efficient manner. It is the manipulation of information and interaction with the environment that entails intelligence in an organism. AI is gradually becoming part of our daily lives and its applications range from video game to military applications.
Speculations on intelligence exhibiting machines is centuries old and dates back to the time of the ancient Greeks. It was not until the invention of the digital computer that AI progressed from mere speculation to reality.
Work on Artificial Intelligence began in the early 1950s but the term Artificial Intelligence was coined in 1956 at a conference at Dartmouth University. The Dartmouth conference had a far-reaching consequence on the AI community and the paradigm advocated at the Dartmouth Conference dominated this field for more than three decades. Currently, there are two main approaches towards creating intelligence in machines: first approach is the top-down approach, which strives to create human level intelligence via computer programs. The second approach, called the bottom-up approach, seeks to create higher-level intelligent systems from lower level systems gradually.
In the early days AI was concerned mainly with logic and finding out ways to teach machines to learn. In the next decade or so great leaps were made in making programs that could solve complex mathematical and logical problems. The AI community at the time thought that these were the really hard problems of intelligence and that human level intelligence was close to hand.
In the end it turned out that nothing could be more wrong. Several robots were built in 60s and 70s that could navigate in toy worlds that were simplified versions of the real world but as their world became a little complex the robots were clueless. While those computers could solve the hardest of mathematical problems but it proved exponentially difficult to get them to recognize faces, navigate in a room full of obstacles. By the end of the 70s it appeared as if AI had come to a dead end.
Pressured from the inside and the outside the AI community began to look for real world applications of its work. After scaling back their earlier grandiose claims AI researchers started building expert systems that had a wide range of applicability. Expert systems are systems that have human like specialization abilities within a limited domain of knowledge. For instance, expert systems have been built that can diagnose diseases, prescribe medicine, find defects in machines, etc. By the closing years of the twentieth century expert systems were a multi-billion dollar industry but the enthusiasm for them was dying down, as expert systems developed for a particular domain of knowledge could not be used for another domain of knowledge.
It would not be an exaggeration to say that these early approaches failed miserably in solving real world problems but valuable lessons were learned from them. From the time of its inception to mid 70s, serial processing theorists who considered the human brain to be a digital computer dominated AI, although sufficient evidence existed that the human brain is parallel in nature.
By the mid 80s enthusiasm for AI was slowly waning and Japan’s disastrous foray into AI was starting to become apparent. It was during this era that Rodney Brooks of MIT presented his revolutionary subsumption architecture for AI according to which perception is not necessary for low-level intelligence. A number of robots have been built according to the subsumption architecture that have performed considerably well as compared to the traditional approach in AI. Rodney Brooks and other who followed his architecture were able to solve problem like getting a robot to navigate a room, the kind of problems that had earlier plagued serial AI.
An important question about AI is that how will we recognize intelligence if we are presented with an intelligent machine. In 1950 Dr Alan Turing published his seminal paper “Computing Machinery and Intelligence” in which he argued that if a machine could fool a person into thinking that it is intelligent then the machine should be considered intelligent. This test, called the Turing Test has lost some of its significance partly because of the fact that it has been possible for a machine to fool a person without having semantics. There is a plethora of literature on this controversy and no one has been able to pin down an exact definition of AI. The one that I think as the best is, “Intelligence is what intelligence does. “ Of course it should be borne into mind that even this definition leaves something to be desired.
Another approach in AI that has gone in and out of fashion several times is that of Neural Networks, a network based on simple models of the mammalian brain. In a Neural Network information is not stored at a specific place in the network but is distributed. Neural Network and even AI can trace its origin to a 1943 paper by Warren McCulloch and Pitts, in which they described how the brain could produce highly complex patterns from simple neurons. Progress in Neural Network was gradually made over the next few decades but since AI was dominated by serial processing paradigm. Neural Network was sidelined for a long time.
In the early 80s interest in Neural Network began to grow and commercial applications for the Neural Network were found for the first time. Neural Networks are especially well suited for control systems and can handle distorted input easily, a task which is sometimes immensely complex for serial processing machines. Neural Networks also have a wide range of applicability and are being employed by several countries for military applications.
After almost fifty years of research and several forays into blind alleys, AI finally seems to be going somewhere as it becomes part routine life but the big goal of self-conscious machines seems to be still far off. This has not, however, deterred the US government that recently announced an ambitious five-year project to create a self-conscious machine.
In less than two decades the computational power of an average desktop computer will equal the computational power of a human being and by the middle of this century an average computer’s computational power will be greater than those of all human beings combined. However Deep Blue and Kasparov have taught us that a super fast machine is not necessarily intelligent. With more and more people boarding the AI bandwagon, may be one day in the not so distant future, a machine, a mere artefact will transcend its humble origins and will join humanity in the rank of being a person. As Kasparov notes, “Maybe the highest triumph for the Creator is to see his creations re-create themselves.”
The writer
|
|
|
|
