We've been acquainted with this type of learning in Chapter 2, How Neural Networks Learn, and now, we are going to explore the features of this learning paradigm in a detailed fashion. Unsupervised learning algorithms in essence aim at finding patterns within datasets by using only the information presented in the datasets themselves. Here, the unsupervised learning algorithm will adjust the parameters (weights in the case of neural networks) without any error measure, and this is the crucial feature that distinguishes unsupervised from supervised learning. The learning itself is triggered only on the basis of the fact that in neurology, similar stimuli produce similar responses. So, applying this fundamental knowledge to artificial neural networks, we can say that similar data produce similar outputs, and these outputs can be grouped in clusters.

Although this learning may be used in other mathematical fields such as statistics, its core functionality is intended and designed for machine learning problems such as data mining and pattern recognition. Neural networks are a subfield in the machine learning discipline, and provided that their structure allows iterative learning, they serve as a good framework to apply this concept on.

One wishes to apply unsupervised learning algorithms when there is no defined target on the data, as well as there is a need to find hidden patterns amongst the data. Most of the unsupervised learning applications are aimed at clustering tasks, which means that similar data points are clustered together, while different data points from different clusters. Further, one application that unsupervised learning is suitable for is dimensionality reduction, wherein one wants some multidimensional data to be classified or reorganized in a less-dimensional domain. In the references [Duda et. al, 2001; Hinton et. al, 1999; Rummelhart & Zipser, 1985; Kohonen, 1982] the reader may find a useful list of articles that show more examples of applications of unsupervised learning.