| Pattern recognition is both an art and a science. We are able to see structure and recognize patterns in our daily lives and would like to find out how we do this. We can perceive similarities between objects, between people, between cultures and between events. We are able to observe the world around us, to analyze existing phenomena and to discover new principles behind them by generalizing from a collection of bare facts. We are able to learn new patterns, either by ourselves or with the hclp of a teacher. If we will ever be able to build a machine that does the same, then we will have made a step towards an understanding of how we do it ourselves.
The two tasks, the recognition of known patterns and the learning of new ones appear to be very similar, but are actually very different. The first one builds on existing knowledge, while the second one relies on observations and the discovery of underlying principles. These two opposites need to be combined, but will remain isolated if they are studied separately. Knowledge is formulated in rules and facts. Usually, knowledge is incomplete and uncertain, and modeling this uncertainty is a challenging task: who knows how certain his knowledge is, and how can we ever relate the uncertainty of two different experts?
If we really want to learn something new from observations, then at least we should use our existing knowledge for their analysis and interpretation. However, if this leads to destruction of all inherent organization of and relations within objects themselves, as happens when they are represented by isolated features, then all what is lost by (not incorporated in) the representation has to be learned again from the observations. |