Basic statistics and basic AI: Neural networks

Abstract

There is a little book with the title What do you believe, but cannot prove? (Brockman 2006). In it, the editors compile the answer to that question given by 50 of the greatest thinkers alive. The editors did not solicit my answer, but if they had it might have been I believe but cannot prove that Artificial Intelligence (AI) and statistics are mostly the same; and when they are not, the differences are within the natural variations occurring within each field. Fortunately, AI and statistics have already been compared and contrasted thoroughly.1 As a result, there is a large body of knowledge that can be employed to create a proof for my claim. However, I also believe that the body of knowledge accumulated is actually more interesting and fruitful than the original question itself, i.e., whether AI and statistics are the same. In fact, I would claim that it does not matter at all whether AI and statistics are, or are not, the same. One characterization that is probably reasonable is that the difference between AI techniques and traditional statistics is not in kind, but in degree. In other words, one may argue that most techniques belong to a continuum, but with different techniques having different degrees of AI-ness or statistic-ness. By the way, in this chapter, AI is synonymous with machine learning. Certainly, however, any problem dealing with data is inherently statistical. As such, tools and concepts that have been traditionally developed in statistical circles should not be ignored. The concepts of a random variable and a probability distribution (or histogram), or the difference between a population and a sample, are all extremely important in the analysis and modeling of data. For instance, the results of a study that ignores the difference between a sample and a population are apt to be generally unreliable (or ungeneralizable), because the results are based on a single sample taken from a larger population, and therefore do not pertain to the latter. The question of whether one can generalize the results from the sample to the population is again one that statistics is designed to address. Alas, some AI-related studies generally ignore such issues. © 2009 Springer Netherlands.