Usefulness of Artificial Neural Network

Neural networks as a system capable of learning, implementing the principle of induction is to say, learning by experience. By comparison with specific situations, they infer a decision system integrated with the wildcard is the number of cases encountered in learning and complexity compared to the complexity of the problem. By contrast, symbolic systems capable of learning, they also implement the coating, are based on the logic of algorithmic complexity by a set of deductive rules (eg prolog).

Thanks to their capacity classification and generalization, neural networks are generally used in problems of statistical nature, such as automatic classification of postcodes or making a purchase decision based on stock price movements. Another example, a bank can generate a set of data about customers who have made a loan made: their income, age, number of dependent children ... and if it is a good customer. If this data set is large enough, it can be used for training a neural network. The bank will then present the characteristics of a potential new client and the network will respond if good customer or not, by generalizing from the cases he knows.

If the neural network works with real numbers, the answer reflects a probability of certainty (eg 1 for 'sure he'll be a good customer, "-1" sure it will be bad client ", 0 for "no idea", 0.9 "almost sure it will be good customer").

The neural network does not always rule usable by a human. The network is often a black box that provides a response when presented a given, but the network did not give a easy to interpret.

Neural networks are actually used, for example:

  • for classification, eg for the classification of animal species by species given a DNA analysis.
  • pattern recognition, eg for optical character recognition (OCR), and in particular by banks to verify the checks by the Post Office to sort the mail according to postal code, etc.., or even to move Automated autonomous mobile robots.
  • approximation of an unknown function.
  • Accelerated modeling of known function but very complicated to calculate accurately, for example some functions of inversions used to decode the signals emitted by remote sensing satellites and data into the surface of the sea
  • Estimates Market Value:
  • Learning to value a business based on available evidence: benefits, debt to long-and short-term revenues, backlog, information technology economy. This type of application does not pose a problem in general
  • attempts to predict the frequency of stock prices. This type of prediction is highly contentious for two reasons, one being that it is not clear that the course of action is so totally convincing a recurring character (the market largely anticipates effect increases as decreases predictable, which applies to any frequency possible variation of the period towards making it difficult to reliably), and one that for the foreseeable future of a company determines at least as heavily on its stock price, if n It is more that can do his past cases of Pan Am, Manufrance or IBM can be convinced.
  • modeling learning and improving teaching techniques.



  • The artificial neural networks require real case examples used for learning (this is called the training set). These cases must be more numerous than the problem is complex and its topology is unstructured. For example, one can optimize a neural system for reading characters using the manual cutting of a large number of words written by hand by many people. Each character can then be presented as a raw image, with a topology with two spatial dimensions, or a series of almost all segments connected. The topology chosen, the complexity of the phenomenon being modeled, and the number of examples must be related. On a practical level, this is not always easy because the examples can be either absolutely limited in quantity or too expensive to collect sufficient.
  • There are problems that deal well with neural networks, especially those classification areas convex (that is to say, such that if points A and B are part of the domain, then the whole segment AB is partly). Issues such as "The number of admissions to 1 (or zero) is odd or even?" are resolved, however, very badly to say such things about N 2 power points, if it is satisfied with a naive approach but homogenous, it is precisely N-1 intermediate layers of neurons, which affects the generality of the method. Android app