Does a computer really think? 
For decade now the proponents of ‘Strong AI’ have tried to persuade us that it is only a matter of century or two (some have lowered the time to fifty years!) until electronic computers will be doing everything a human mind can do. Stimulated by science fiction read in their youth, and convinced that our minds are simply ‘computers made of meat’ (as Marvin Minsky once put it), they take for granted that pleasure and pain, the appreciation of beauty and humour, consciousness, and free will are capacities that will emerge naturally when electronic robots become sufficiently complex in their algorithmic behavior .
Some philosophers of science notably John Searle strongly disagree. To them a computer is not essentially different from mechanical calculators that operate with wheels, levers, or anything that transmit signals. Because electricity travels through wires faster than other forms of energy (except light) it can twiddle symbols more rapidly than mechanical calculators, and therefore handle tasks of enormous complexity. But does an electrical computer understand what is it doing? .
There is no dispute over computers exhibiting intelligent behavior in many situation. The bedrock of manifestation of intelligence by computers is its knowledge bases and inference mechanism, heuristic and optimizer based search approaches in lessening computational complexity. However most AI programs use pattern recognition and search techniques which lead one to believe that these are in essence non-intelligent in nature. One need to provide a lot of information to the computer to initiate a modicum of learning in it. Only thereafter the computer may manifest some degree of intelligence in task solving.
However the human mind is just too complex to duplicate. Computers certainly cannot think in the same way humans do but they can be very useful for increasing our productivity. This is done by several commercial AI technologies .
1. Roger Penrose, The Emperor’s New Mind, Penguin Books, Chapter 1
2. Turban, Rainer, Potter, Introduction to Information Technology second edition, Pg 394, Wiley India ISBN 978-81-265-0968-3
What are the challenges for building intelligent systems?
It is an acknowledged fact that to build an effective intelligent system one would require a lot of data. Learning won’t be effective with a small data size. Greater the volume of data, the better is the learning. For example, the successful intelligent based application emerged in those domains where data is quite readily available e.g. Image Recognition system. Computer requires several orders of data in excess than humans to perform the same learning task. The owners of data are the big companies like Google, Amazon, Facebook, Microsoft and a few more. These organizations do have an edge over others to build intelligent applications.
The basis of any intelligent decision making is knowledge. How to capture knowledge, represent and codify it for later use in inference making/reasoning is a non trivial task. We are still far, far away from building Automated Real Time Reasoning system. Current Machine Learning techniques like Decision Tree and Neural Networks have obtained limited success in a narrow domain. Even Neural learning is based on synaptic weight changes based on Hebb’s work which dates back to 1959 …
Which tasks are considered manifestation of intelligence by a computer?
In order to be perceived as intelligent, the computer must be able to do some or all of the following:
1. Learn – Learning enables an entity to be able to perform correctly and efficiently an already encountered problem in the past that have been solved. In a game of chess between a human and a computer, if the computer encounters a similar board position and a move by the player which led to its defeat, the computer would have learnt from its mistake and try to circumvent the situation by playing a different game so as not to repeat his failure.
Building an image classifier through feature selection and supervised learning would help us in image classification. If the image is that of a fish, it would result in a fish classification (categories of fishes).
2. Recognize – You may recognize your friend even after thirty years when he / she has changed considerably. This ability in humans need to be incorporated in intelligent machines.
3. Decision making under uncertainty – Human beings make excellent decisions based on imprecise variables, as in driving when the man at the wheel has to navigate traffic on the roads, and decide on the driving speed and turns to take based on the velocity of the approaching cars on several lanes.
4. Natural Language Processing – When a computer system is able to carry on a normal conversation with a person fooling him/her into thinking that it is actually a human who is communicating…
5. Acumen and Intelligence – When a system employs an intelligent guess, as in the case of Turing Test, where an interrogator is trying is trying to determine whether the response is being generated by a computer or given by a human, to fool the interrogator the computer system must hold back the answer to a complicated multiplication problem for a human (which it can do easily) and output the answer in approximately the same time that a human would take to compute such results.
6. Real Time Decision Making - Sensing traffic density in different car lanes at a junction through sensors, and allowing traffic flow for greater duration for lanes with heavy traffic than lanes with lesser traffic.
7. Intelligent Inference /Knowledge Based Reasoning – Deriving knowledge / inference from a given set of facts. Feeding parametric values of a patient into intelligent software may result in proper diagnosis of the ailment and enable right prescription of drugs.
8. Fuzzy Logic / Intelligence – Fuzzy Logic/Intelligence has been employed in auto focus cameras, washing machines, robot and cars to provide functionalities that normally would have required human intervention and intelligence to execute them.
9. Searching using heuristics – Capability to perform intelligent search like the informed search algorithms – Best First Search, A*, AO*
10. Swarm intelligence – For combinatorial problems i.e., problems having a large number of possible alternatives, collective intelligence in the bird and animal kingdom such as ants (ant colony optimization) and birds (particle swarm optimization) for arriving at optimized values could be incorporated for problem solving by AI systems.
11. Planning – An agent interacts with the world via perception and actions. Perception involves sensing the world and assessing the situation. Actions are what the agent does in the domain. Planning involves reasoning about actions that the agent intends to carry out. “Planning is the reasoning side of acting” (Ghallab et al., 2004). This reasoning involves the representation of the world that the agent has, as also the representation of its own actions .
12. Case Based Reasoning – CBR systems make use of knowledge/experience gained in the past for current problem solving.
13. Prediction - Neural learning can enable good prediction and has resulted in successful applications for stock markets …
14. Soft Computing – The ability to process/perform computation with imprecise data and variables so necessary in many real life situations …
1. Deepak Khemani, A First course in AI, Chapter 7, Planning, Pg 192, McGraw Hill Higher Education, ISBN-13: 978-1-25-902998-1
Q. How is AI and Robotics relevant in a developing nation like India & the Third World?
Being cost intensive, research in AI and Robotics have taken place mostly in developed countries like USA and Japan. How does developing countries like India benefit from this technology? Without our realization we are actually surrounded by AI applications in our daily lives. Automated applications like Road Traffic Controller or Washing machines in our homes have their roots in Fuzzy logic – a branch of AI. AI applications are too numerous having a great impact and no country can afford to stay away from it. Whether it is Sentiment Analysis or use of AI concepts like Machine Learning for Big Data, business application of AI will make the technology commercially viable and support its growing popularity. With greater cost-effectiveness, expensive fields like Robotics can be useful for providing elderly care and working in disaster zones.
Weak AI and Strong AI
Classification of AI broadly falls into two categories: Strong AI and Weak AI. Most of the application around us would classify as Weak AI. AI mimics intelligence/intelligent behavior but need to operate in conjunction with manual input. It does not possess the ability to read the environment and act on its own, the way the human brain is capable of decision making based on sensory inputs. Remote controller would be an example of Weak AI as it requires manual intervention.
Strong AI still remains a distant goal for practitioners. We haven’t succeeded yet with Strong AI and no application exist that exhibit characteristic of strong AI. Strong AI is the ultimate dream of Computer Scientists to build a computer system equipped with an ‘inner brain’ that would automate intelligent behavior. Work on Neural Network based learning and application is in this direction.
(12th July, 2017)