5 Main Approaches to AI Learning - dummies

By John Paul Mueller, Luca Massaron

An algorithm is a kind of container. It provides a box for storing a method to solve a particular kind of a problem. Algorithms process data through a series of well-defined states. The states need not be deterministic, but the states are defined nonetheless. The goal is to create an output that solves a problem. In some cases, the algorithm receives inputs that help define the output, but the focus is always on the output.

Algorithms must express the transitions between states using a well-defined and formal language that the computer can understand. In processing the data and solving the problem, the algorithm defines, refines, and executes a function. The function is always specific to the kind of problem being addressed by the algorithm.

Each of the five tribes has a different technique and strategy for solving problems that result in unique algorithms. Combining these algorithms should lead eventually to the master algorithm that will be able to solve any given problem. The following discussion provides an overview of the five main algorithmic techniques.

Symbolic reasoning

One of the earliest tribes, the symbolists, believed that knowledge could be obtained by operating on symbols (signs that stand for a certain meaning or event) and deriving rules from them. By putting together complex systems of rules, you could attain a logic deduction of the result you wanted to know, thus the symbolists shaped their algorithms to produce rules from data. In symbolic reasoning, deduction expands the realm of human knowledge, while induction raises the level of human knowledge. Induction commonly opens new fields of exploration, while deduction explores those fields.

Connections modelled on the brain’s neurons

The connectionists are perhaps the most famous of the five tribes. This tribe strives to reproduce the brain’s functions by using silicon instead of neurons. Essentially, each of the neurons (created as an algorithm that models the real-world counterpart) solves a small piece of the problem, and using many neurons in parallel solves the problem as a whole.

The use of backpropagation, or backward propagation of errors, seeks to determine the conditions under which errors are removed from networks built to resemble the human neurons by changing the weights (how much a particular input figures into the result) and biases (which features are selected) of the network. The goal is to continue changing the weights and biases until such time as the actual output matches the target output. At this point, the artificial neuron fires and passes its solution along to the next neuron in line. The solution created by just one neuron is only part of the whole solution. Each neuron passes information to the next neuron in line until the group of neurons creates a final output. Such a method proved the most effective in human-like tasks such as recognizing objects, understanding written and spoken language, and chatting with humans.

Evolutionary algorithms that test variation

The evolutionaries rely on the principles of evolution to solve problems. In other words, this strategy is based on the survival of the fittest (removing any solutions that don’t match the desired output). A fitness function determines the viability of each function in solving a problem. Using a tree structure, the solution method looks for the best solution based on function output. The winner of each level of evolution gets to build the next-level functions. The idea is that the next level will get closer to solving the problem but may not solve it completely, which means that another level is needed. This particular tribe relies heavily on recursion and languages that strongly support recursion to solve problems. An interesting output of this strategy has been algorithms that evolve: One generation of algorithms actually builds the next generation.

Bayesian inference

A group of scientists, called Bayesians, perceived that uncertainty was the key aspect to keep an eye on and that learning wasn’t assured but rather took place as a continuous updating of previous beliefs that grew more and more accurate. This perception led the Bayesians to adopt statistical methods and, in particular, derivations from Bayes’ theorem, which helps you calculate probabilities under specific conditions (for instance, seeing a card of a certain seed, the starting value for a pseudo-random sequence, drawn from a deck after three other cards of same seed).

Systems that learn by analogy

The analogyzers use kernel machines to recognize patterns in data. By recognizing the pattern of one set of inputs and comparing it to the pattern of a known output, you can create a problem solution. The goal is to use similarity to determine the best solution to a problem. It’s the kind of reasoning that determines that using a particular solution worked in a given circumstance at some previous time; therefore, using that solution for a similar set of circumstances should also work. One of the most recognizable outputs from this tribe is recommender systems. For example, when you buy a product on Amazon, the recommender system comes up with other, related products that you might also want to buy.

The ultimate goal of machine learning is to combine the technologies and strategies embraced by the five tribes to create a single algorithm (the master algorithm) that can learn anything. Of course, achieving that goal is a long way off. Even so, scientists such as Pedro Domingos are currently working toward that goal.