Intelligence is something we humans thrive on. We think of ourselves as the most intelligent beings, certainly on this planet but possibly in the entire universe.
However, it is a notoriously difficult task to define what “intelligence” really is.
Among various definitions, perspectives, and outlooks, the standard consensus is that a being is intelligent if “it” can respond to events and stimuli around it and be able to manipulate either the surroundings or itself to make things better for itself.
This definition suits artificial intelligence nicely since it can be adapted to non-living beings almost readily.
Artificial intelligence, more commonly known by its abbreviation “AI,” is the field of study that analyses this process of understanding or gaining intelligence; it is also concerned with building systems or agents that display such intelligent behaviour.
Given today’s pervasion of AI in almost every field of innovation and development, starting from driverless cars to the recommendation of products online to personalised healthcare to natural language conversations, it is important to understand what artificial intelligence really is, and its capabilities and inabilities.
Comparison of AI systems with humans is natural. Throughout history, most such systems have been modelled on humans. However, humans may not always show what is called “rational” behaviour, in the sense that a human may choose an action that does not necessarily produce the best outcome for themselves. There is, thus, a dichotomy of human behaviour versus rational behaviour.
Perspectives Of AI
Using this dichotomy, the field of AI can be analysed from four broad perspectives. These perspectives test the ability of an AI system from four different angles.
The first is the ability to “act humanly,” that is, whether the system can mimic a human in its actions.
The most famous thought experiment in this field is called the Turing Test, named after British mathematician Alan Turing.
In this experiment, a set of questions is asked to a human being as well as an AI system and the responses are collected. The human interrogator does not know who is who, and the AI system passes the Turing Test if the interrogator cannot distinguish between the two.
This does not require the AI system to be correct or perfect. In fact, since its role is to mimic a human, and humans are error-prone, a perfect set of responses may actually give the game away.
The Turing Test requires the AI system to have the capabilities of natural language processing (to understand the questions written in a human language), knowledge representation (to store and process what it knows), automated reasoning (to answer a question by processing the stored knowledge), and machine learning (to adapt to new questions and draw conclusions from previous experience).
Researchers have proposed extending the Turing Test to the Total Turing Test, which requires the AI system to interact with humans and objects in the real world. This requires the additional capabilities of computer vision and speech recognition (to perceive the real world) and robotics (to manipulate objects in the real world).
The second important perspective of AI is the ability to ‘think humanly’.
Testing this ability requires the development of a model of the human mind and thoughts. Cognitive science and psychology are two important subjects that deal with this aspect. Testable hypotheses of the human mind are designed and experiments performed to test the validity of the hypotheses.
The third and fourth perspectives deal with rationality, a subject that has been discussed and debated in philosophical treatises over centuries across the world.
One of the important ways to understand rationality is through the use of ‘logic’. Can a conclusion be arrived at logically?
The stock example, due probably to Aristotle, is if the predicates ‘Socrates is a man’ and ‘all men are mortal’ are true, is the conclusion ‘Socrates is mortal’ valid?
The conclusion can be arrived at by applying deductive reasoning. This logical argument structure is called syllogism. The third perspective of an AI system, to ‘think rationally’, tests this aspect.
While statements such as ‘all men are mortal’ are certain, most real-life statements, such as ‘India will win the next cricket world cup’, cannot be determined to be either completely true or completely false. The field of probability and statistics here comes to the rescue. Uncertain information about predicates is handled by associating them with probabilities.
The fourth perspective is to go beyond thinking and test the ability to ‘act rationally’.
A rational AI system not only thinks rationally but also takes action such that it achieves the best outcome. It is easy to understand this for board games such as chess and ‘go’, where an AI player is pitted against a human opponent. The objective is to win the game and the move that is most likely to achieve it is the best move. IBM’s Deep Blue and Google’s AlphaGo systems caused quite a flutter when they beat the best human players.
The last perspective, however, opens a Pandora’s box. Acting ‘rationally’ may not always be acting the ‘best’ in terms of human interests or interactions.
Consider, for example, a chess-playing machine. If the goal is to win the game, the machine is free to do whatever it deems advantageous as long as the rules of the game are not violated. It can, for example, shine a light on the eyes of the human opponent or increase the temperature of the room to uncomfortable levels to disturb the thinking process.
While some such conducts can be disallowed explicitly, it is not always easy or possible to list all the possibilities that a machine may take to achieve its goal. Asimov’s Three Laws of Robotics, for example, only lists broad rules where human beings cannot be hurt. Hence, the paradigm of ‘acting rationally’ can be modified to ‘acting the best for a human’.
This leads to important discussions on where AI is headed. We will return to it in part two of this article.
Technical Paradigms Of AI
We now discuss the field of artificial intelligence from a technical standpoint.
The first broad paradigm of AI is problem solving. A large part of problem solving involves searching. Given a set of rules and an objective, an AI system searches its next move among a maze of possibilities such that, eventually, the goal is reached.
Navigating around obstacles for robots to conclude a task is a prime example. Sometimes, objectives are modelled as games with utility functions for each move. Should country X build up a nuclear arsenal? The decision is not unilateral because it depends on how enemy countries are behaving. The field of game theory developed by economists is used to solve and analyse such games.
The third important type of problem solving involves constraint satisfaction problems. Given a set of variables with their domains, can each variable be assigned a value without violating a given set of constraints?
Important application areas include job scheduling, such as in a car assembly system. Constraints, such as a wheel axle needing to be fixed before putting on the wheels, must be respected, and the objective is to find a parallel assignment of tasks to limit the total assembly time to less than the target time.
Propositional logic and first-order logic form the basis of the second important sub-field, which is reasoning and logic. The example of Socrates earlier highlights the use of logic. Knowledge representation and reasoning builds upon such logic systems.
An important concept is that of an ontology that describes the categories and relationships the objects in the system can have. It is often organised in a hierarchy with inherited properties. Thus, if the task is to find a human who knows Sanskrit, the system can return a woman since it can reason that a woman is a human being.
Since in real life most situations are uncertain and facts and relationships are mostly ‘likely’ rather than ‘certain’, this leads to the next paradigm, that of uncertain knowledge and reasoning. Probabilistic reasoning using tools such as Bayesian networks and hidden Markov models derive probabilities of events or inferences.
An interesting example is trying to guess the weather outside by sitting in a room and only observing if visitors are carrying umbrellas. Decision-making systems use such probabilistic reasoning to meet a goal in collaborative as well as adversarial environments.
The fourth paradigm of AI, machine learning (or ML), is undoubtedly the most popular paradigm in both research as well as common parlance. It is so popular and pervasive that even students of AI often mistake ML to be AI.
Machine learning is the ‘art’ of making a machine or system learn how to achieve an objective without providing an explicit way of doing it.
Driving a car is a great example. When a human is taught driving, only some general rules are mentioned, such as that pressing the brake stops the car and turning the wheel changes the direction of the car. No human is or can be taught to rotate the wheel by x degrees at y speed so as to negotiate a turn of z degrees on the road for all combinations of x, y, and z. That comes from the experience of driving a car.
The same is true for machine learning systems. A system is given a lot of examples to learn from. In the supervised learning setting, each such example is additionally endowed with a class tag, while in an unsupervised setting, the tag is missing. The system then undergoes ‘training’ using these examples; often, it uses a ‘validation’ set to assess how well it has learned, and repeat the training if needed.
Students use this kind of validation when they try to solve previous years’ examination papers; if they do not do well, they go back to training.
After the machine is trained, given a ‘test’ object, the machine tries to reason about it correctly. The reasoning is typically classification, where the task is to predict what class the object falls under, or regression, where an exact value is predicted.
Examples of classification include identifying a handwritten digit, deciding whether an email is spam or normal, and diagnosing whether a medical image indicates disease. Unsupervised learning problems include clustering and anomaly detection. Detecting anomalies automatically is especially important in network intrusion detection systems.
In recent years, the semi-supervised learning setting has also come up, where a few examples with the class are given while a lot more are without a class tag.
Machine learning also includes reinforcement learning, where a machine is ‘rewarded’ when it produces a good outcome and ‘punished’ when it does not. The immediate parallel that can be drawn is training animals to perform in circuses. Reinforcement learning is used in various real-life applications, including driverless cars (to control acceleration and braking), stock price predictions, and recommendation systems.
Important machine learning models include decision trees, support vector machines, and artificial neural networks (or ANN).
ANNs are particularly important since they try to mimic the working of a human nervous system where information is processed and then passed on from one neuron to the next, layer by layer (neurons are called nodes in ANNs).
An extremely successful family of ML models is a type of ANNs, called deep neural networks (or DNN). The process of inferencing using DNNs is called deep learning.
In essence, DNNs are simply variants of ANNs that have multiple layers of hidden nodes (this multiplicity of layers lends the name ‘deep’). They are astonishingly accurate in solving a wide range of real-life problems and, in many areas, have outperformed human experts. Their stunning successes in even ‘humanesque’ tasks, such as language processing and conversation, is stupefying.
This success is in part due to the architecture of such machines. It has been shown that given enough training data, DNNs can model any mathematical function to any arbitrary precision. This, however, requires the use of an enormous number of hidden nodes and layers.
The advancement of computing paradigms, tagged data, and available hardware, such as GPUs (or graphical processing units), have contributed massively to this success. Consequently, it is not uncommon nowadays to encounter DNNs with hundreds of crores of parameters.
This was the first of two parts covering the basics of artificial intelligence. The next part will cover the applications, issues, and future of AI.
This article has been published as part of Swasti 22, the Swarajya Science and Technology Initiative 2022. We are inviting submissions towards the initiative.
Other Swasti 22 reads:
The Basics Of A Quantum Computer, Explained
National Science Day: The Raman Effect And One Of Its Key Applications, Explained