Machine intelligence/Artificial intelligence

We are entering an age of intelligent machines. Computers are becoming increasingly adept at mental tasks beyond just calculating results and storing and retrieving data. They are becoming better at the sort of pattern-recognition and hypothesis-generation that is the essence of human creative intelligence. Computers can now devise “creative” proofs of mathematical conjectures, devise scientific experiments ^[1] and compose haiku ^[2]. Computers can be programmed to recognize patterns, to generate and test novel hypotheses and to display creativity, adaptation and ingenuity, albeit still in narrowly-defined areas.

With such ingenuity now being displayed by artificial intelligence, it is possible to replace human labour not just at physical jobs like factory assembly, but also mental jobs like astronomy, drug-discovery and strategic planning.

Humanity now has an incredible opportunity to use artificial intelligence systems to free us from boring jobs and to advance art, science, mathematics, to create highly-efficient supply-chains for industry, to detect disease faster and more accurately and to design more efficient machines. It is the responsibility of an advanced civilization to learn how to use AI to best serve the needs and wants of Man.

Types of AI

But how is it possible for a computer to be creative, adaptive and insightful? I mean, they're just mindless machines that follow rules, right? Well, right, but there are mathematical models of how complex rule-governed systems (like the brain) form creativity and insight. Some of the models used to create artificial intelligence —

• Expert systems were the earliest type of artificial intelligence and are still useful today. They are made by quizzing human experts about their thought processes and decision-making strategies so that they can be emulated by a computer.

• Bayesian nets make generalizations from data and use them to form beliefs. By looking at these beliefs, a Bayesian net can calculate the probability of a particular explanation or outcome. For example, in a medical AI, a Bayesian network might notice that a fever often goes hand-in-hand with a positive result for a certain virus. The network forms a belief that these two are connected. When faced with a complex set of data, it finds the beliefs that best fit the data to form explanations. It then eliminates hypotheses that contain contradictory beliefs. In this way, it uses both inductive and deductive logic. It then spits out a judgement of the comparative likelihood of various explanations. Bayesian networks work probabilistically, so no beliefs are set in stone, but the more beliefs an explanation conforms to, the more likely it is deemed to be.

• Markov models model the probability of subsequent states following on from the current state. For example, a weather-simulation may know that low pressure in the Gulf of Mexico is often followed by high winds in Brazil.

• Neural networks are a sophisticated from of AI based on the processes observed in the human nervous system. Input to a neural network system is first fed into one layer of neurons. The input data points are randomly connected to different neurons. If a neuron gets enough input from important enough sources, the neuron fires. This firing becomes the input for a second layer of neurons, again randomly connected to the first, and the process repeats. At the end, the result emerges from the last layer of neurons. There are several variables that affect how exactly a message spreads through the neural network: the importance given to different connections, the placement of connections between different neurons and different layers, the amount of input a neuron needs to fire etc. These variables are random at first, but if the result that comes out is correct, they are reinforced, and if it was wrong, they are negatively indicated. In this way, the system learns over time what works and what doesn't. It constantly corrects itself. Like a biological nervous system, an artificial neural network starts out knowing nothing, but acquires competence over time and with experience.

Computer simulations of actual neurological processes taking place in the human brain are becoming more and more detailed and powerful. We can now scan the brain in real-time with ever-increasing resolution and this has given us great insight into chaotic, self-organizing, fractal methods of computing information. The rate at which neurological scans be translated into computer simulations is impressive.

• Genetic algorithms are a fascinating emulation of the process of genetic evolution. Species evolved because the ones that got the job done passed on their genetic information, which got mixed by sexual reproduction with the genetic information of other thriving organisms, and over generations this led to the fittest genetic information coming to the forefront. In exactly the same way, a genetic algorithm is programmed to evolve over generations. Say for example you are designing a car — you enter a range of weights, angles and shapes and specify that you want high aerodynamic efficiency. The computer generates a thousand possible designs by tweaking the variables, throws out the unaerodynamic ones and splices together the ones that are up to par. This creates a new generation with the best qualities of each of the best members of the first generation. This is repeated for many generations until the very best qualities have been distilled into a small number of optimal designs. Like biological evolution, this method sometimes gives surprising and exotic solutions to the problem posed, giving the sense of creativity. Genetic algorithms can be applied to all kinds of things, from robots that can teach themselves to walk, to animals that evolve within virtual worlds, to bots that learn to play the stock market.

The best AI is when a combination of these methods are all used together. A sophisticated AI program might run all these methods in combination and have another AI monitoring them all to find the strengths and weaknesses of each, so as to amplify the intelligence of the whole.

One key thing to note is that AI learns from experience; it is not a dumb, rule-bound machine, it is intelligent and adaptive. This allows it to replace humans in many tasks. With the exception of expert systems, all the artificial intelligence methods above are only smart if exposed to a large amount of high-quality learning environments. This is another reason why open collaboration is vital; global collaboration results in the greatest possible amount of learning data for a system, which leads to the most competent possible AI.

Uses for AI

• Generating educational material (using genetic algorithms)
• Design of vehicles, engineering. Genetic algorithms are frequently used to evolve optimal engineering designs by running simulations.
• Optimize logistics. Logistics has recently become far more a task for computers than for humans.
• Medical diagnostics
• Autonomous astronomers: The Moving Object and Transient Event Search System (MOTESS) is an AI observatory that has discovered hundreds of asteroids without human intervention.
• Drug discovery
• 'Autonomic' computers that repair their own bugs. This is one of the early examples of self-maintenance and repair technology.