We discussed in the past many general, "philosophical" problems that are very relevant to supervised learning. Namely, we are dealing with finite, often small, and potentially even biased sample from unknown underlying distribution; we can run into overfitting, i.e. find a "dependence" that's not really there; we have to understand that the information a dataset can possibly provide, given its size and noise level, is limited - for instance we might be unable to confidently determine the slope of the second order term aX^2 in regression model even if we know from the "physics" of the problem that this term should be present, so with the data in hand we might need to settle for a simpler, "physically incorrect" model with only a linear term (as we cannot discern anything beyond that simple trend in the data at hand). When we apply unsupervised learning methods to the data, however, we are trying to find any structure present while "keeping our mind open" and not assuming any particular model. Hence all these general problems so common in supervised learning are much less relevant