The explanatory power of Artificial Neural Networks
(1998) Methodos: The explanatory power of models in the social sciences — Location: Louvain-la-Neuve (14.November.1998)
Files
43-Theexplanatorypowerofartificialneuralnetworks.pdf
Restricted Access - Adobe PDF
- 111.62 KB
Details
- Authors
- Abstract
- Many engineering problems include some kind of recognition: from automatic character recognition to the control of steel quality in a steelworks, through the fault detection in nuclear plants or the prediction of financial rates, it is impossible to enumerate all domains where the key challenge is to identify an input-output relationship between variables or concepts. When the physical relationship is difficult to tackle, models are developed to approximate it. There are many ways to develop such models. Linear ones are used in many cases, even if it known that the linearity limitation will make the model inadequate. Non-linear models are the solution, but they suffer from many limitations, related to the concept of recognition itself: what is the relation to be recognised if it is only known through examples? Artificial neural networks (ANN), i.e. models based on the remote analogy with the information processing in a human brain, try to answer to this question. ANN models are built (trained) on examples, the purpose being to keep the equilibrium between a correct training and a useful (in some cases meaningful) representation. Despite the fact that ANN are known to be "blind", or non-explanatory, we intend to show that it is possible to feed to or to extract knowledge from these models; the step towards an explanatory power is then straightforward. But the real question is to know to what extend it is possible to interpret the results of such a "non-explanatory" model: what is the real difference between extracting representable knowledge from a computational model, and using a "blind" model to predict, classify or recognise some relationship?
- Affiliations
Citations
Verleysen, M. (1998). The explanatory power of Artificial Neural Networks. Proceedings of Methodos: The explanatory power of models in the social sciences, p. 13 pages. https://hdl.handle.net/2078.5/254019