In general, Artificial Intelligence is able to make a computer learn from experience, allow a machine to build a representation of the world, and let it take logical and common sense decisions. Today, the first two fields are already yielding nice results, but AI performance in trivial decision making is still weak and very time-consuming. The world of AI is divided in two sides. The first one aims to understand and reproduce the intelligent behaviour of human beings. This is related to philosophical questions about the nature of mind, the regularity of organic systems, and syntax and semantics problems. Mr. Maccarone's talk however was about the second aspect of AI, namely the way in which we can take advantage of the enormous memory capabilities in computers, that can manage large amounts of data and perform operations with a high degree of precision. In short, all the tasks humans cannot do.
The goal is not to replace human experts with expert systems but to use the computer in procedures, such as connecting to large databases or handling noise in an appropriate way. At present, computers are not able to deal with partial or "noisy" information but we should like them to manage problems which are non-algorithmic and also to extract significant data from incomplete information. We expect computers to study many features of the same problem simultaneously, whereas most of the operations in computers still are non-parallel. A concrete example of complex classification is pattern recognition, also applied in medical imaging, in order to detect the shape of a single abnormal cell in a picture among a host of normal ones.
Mr. Maccarone introduced the methods of neural networks, fuzzy logic, and genetic algorithms. These three systems have the common feature that they try to imitate the biology of the human brain, which allows them to deal with uncertainty, chance, and probability. The principal idea in artificial neural networking (ANN) is to teach a computer to recognise and classify patterns through intensive training. First, you train a machine with many examples of images containing the specific feature. Then you request the trained ANN to extract the feature from a previously unknown sample. The ANN consists of a large number of neurons or variables, which take on different values and communicate with each other. During the training phase the communication rules between the synapses are externally tuned, such as to reproduce the right output associated with a specified input to optimise recognition. After the training period, the ANN is able to deal with whole new samples.
Fuzzy logic forms another method to approach the human way of reasoning, as Mr. Maccarone explained. Fuzzy systems can divide the output in classes, using a non-binary multi-valued logic for qualitative classification. A water temperature of 30° C, for instance, is labelled neither hot nor cold, but hot for 75% and cold for 25%. Special algorithms are applied to convert the fuzzy output, giving no sharp answers, in a more workable, crisp computer response. In the industrial world, these systems are already widely used and they have proven to be much more efficient than other computers based on the traditional sharp and quantitative evaluation of data. This type of system can be implemented in the data mining of large databases.
The third AI classification method is the so-called genetic algorithm, relying on the idea of evolution. Initially, you can try to solve a given problem with random solutions, which is called the genome of the solution proposed to the problem. The second step consists in evaluating the fitness of the answer by letting all the solutions evolve according to their fitness through the mutual exchange of genes, which are the features of the solutions. In this way, a sort of sexual revolution of solutions is generated, as Mr. Maccarone pointed out. By mixing the genes of two solutions, you can obtain successful offspring in the end. This algorithm is very useful when managing huge amounts of data where a completely random approach to the problem is useless or too time-consuming.
One future evolution of the Artificial Intelligence idea is a combination of the three approaches mentioned by Mr. Maccarone. It is possible, for example, to train neural networks with genetic algorithms and not start from complete randomness of net connections between neurons. Instead, you might tune them by using some genetic algorithm. All of those approaches are currently being investigated by researchers and maybe, in the future, they will be used for medical imaging and industrial applications.