Given John’s recent posts on CMU’s new machine learning department and “Deep Learning,” I asked for an opportunity to give a computational learning theory perspective on these issues.<\/p>\n
To my mind, the answer to the question “Are the core problems from machine learning different from the core problems of statistics?” is a clear Yes. The point of this post is to describe a core problem in machine learning that is computational in nature and will appeal to statistical learning folk (as an extreme example note that if P=NP– which, for all we know, is true– then we would suddenly find almost all of our favorite machine learning problems considerably more tractable).<\/p>\n
If the central question of statistical learning theory were crudely summarized as “given a hypothesis with a certain loss bound over a test set, how well will it generalize?” then the central question of computational learning theory might be “how can we find such a hypothesis efficently (e.g., in polynomial-time)?”<\/p>\n
With this in mind, let’s spend the rest of the post thinking about agnostic learning. The following definition is due to Kearns, Schapire, and Sellie:<\/p>\n
Fix a function class C. Given a training set drawn from an arbitrary distribution D and *arbitrarily labeled* (that is, we are given samples from an arbitrary distribution D over some set X x Y<\/i>), efficiently output a hypothesis that is competitive with the best function in C with respect to D. If there exists a provably efficient solution for the above problem we say C is agnostically learnable with respect to D. Notice the difference between agnostic and PAC learning: we *do not assume* that the data is labeled according to a fixed function (I find this to be a common criticism of the PAC model).<\/p>\n
To make this a little more concrete, let C be the class of halfspaces, let X = Rn<\/sup>, and let Y = <\/i>{-1,1}<\/i>. Let opt = min{h in C}<\/sub> <\/i> (classification error of h with respect to D). Then the goal is, given an *arbitrarily* labeled data set of points in Rn<\/sup> <\/i>, find a hypothesis whose true error with respect to D is at most opt + eps<\/i>. Furthermore the solution should provably run in time polynomial in n<\/i> and 1\/eps<\/i>. Note that the hypothesis we output *need not be a halfspace*– we only require that its error rate is close to the error rate of the optimal halfspace.<\/p>\n Why choose C to be halfspaces? At the heart of many of our most popular learning tools, such as Support Vector Machines, is an algorithm for learning a halfspace over a set of features in many dimensions. Perhaps surprisingly, *it is still not known*, given samples from an arbitrary distribution on X x Y<\/i>, how to efficiently ‘find the best fitting halfspace’ over a set of features. In fact, if we require our algorithm to output a halfspace as its hypothesis then the agnostic halfspace learning problem is NP-hard. Recent<\/a> results<\/a> have shown that even in the case where there is a halfspace with error rate .0001 with respect to D it is NP-hard to output a halfspace with error rate .49999 (achieving .5 is of course trivial).<\/p>\n The astute reader here may comment that I am ignoring the whole point of the ‘kernel trick,’ which is to provide an implicit representation of a halfspace over some feature set in many dimensions, therefore bypassing the above hardness result. While this is true, I am still not aware of any results in the statistical learning literature that give provably efficient algorithms for agnostically learning halfspaces (regardless of the choice of kernel).<\/p>\n My larger point is that a fundamental approach from statistical learning theory (running an SVM) is deeply connected to strong negative results from computational complexity theory. Still, as I explain in the next paragraph, finding the ‘best fitting halfspace’ remains an open and important computational problem.<\/p>\n Negative results not withstanding, it now seems natural to ask ‘well, can we find any efficient algorithms for agnostically learning halfspaces if we allow ourselves to output hypotheses that are not halfspaces?’ Briefly, the answer is ‘to some extent Yes<\/a>,’ if we restrict the marginal distribution on X to be one of our favorite nice distributions (such as Gaussian over R^n). The solution runs in polynomial-time for any constant eps<\/i> > 0 (for various reasons involving the ‘noisy XOR’ topic three posts ago, obtaining efficient algorithms for subconstant eps<\/i> seems intractable).<\/p>\n The complexity of agnostically learning halfspaces with respect to arbitrary distributions remains open. A solution (if one exists) will require a powerful new algorithmic idea and will have important consequences regarding our most commonly used machine learning tools.<\/p>\n To respond to Andrej Bauer’s comment three posts ago:<\/p>\n ‘I guess I have a more general question. I understand that for the purposes of proving lower bounds in computational complexity it’s reasonable to show theorems like “can’t learn XOR with NOT, AND, OR.” But I always thought that machine learning was more “positive”, i.e., you really want to learn whenever possible. So, instead of dispairing one should try to use a different bag of tricks.’<\/p>\n I think Andrej is right. Solving future machine learning problems will require a more sophisticated ‘bag of tricks’ than the one we have now, as most of our algorithms are based on learning halfspaces (in a noisy or agnostic setting). Computational learning theory gives us a methodology for how to proceed: find provably efficient solutions to unresolved computational problems– novel learning algorithms will no doubt follow.<\/p>\n","protected":false},"excerpt":{"rendered":" Given John’s recent posts on CMU’s new machine learning department and “Deep Learning,” I asked for an opportunity to give a computational learning theory perspective on these issues. To my mind, the answer to the question “Are the core problems from machine learning different from the core problems of statistics?” is a clear Yes. The … <\/p>\n