Popular Posts

Friday, December 12, 2008

Stuck on a Hill

In artificial neural network (ANN) research, there's a well known problem of local minima (or maxima). I've worked a bit with ANNs but much more with a (superior) learning algorithm, Support Vector Machines (SVMs). Unlike the latter, ANNs require heuristics to "converge" on an optimal solution given some very large decision surface. The heuristics, to simplify a bit, are intended to get the algorithm to converge on a global, not local, solution. Local solutions are unwanted because they can appear to be global (good) solutions, given some snapshot of the decision surface, but in fact are very bad solutions when one "zooms out" to see the larger picture. This phenomenon is perhaps best illustrated with geographical imagery. If I am walking up and down hills, en route to a very large mountain, a local maxima might be the top of a foothill. But it would hardly be a global maxima, like the top of the mountain. The point is that ANNs can converge on the foothills, telling us that it's the mountain.

ANNs notoriously suffer from this limitation, but the "blindness problem" is endemic to all statistical learning algorithms, including SVMs (though, at least in theory, you can get an optimal solution with an SVM). Using such algorithms to learn from past experience (i.e., training data), you generate an approximation function for the underlying distribution you're trying to model. You see the function as worth the time it took to gather training data, select the features, and train, if it approximates pretty well the underlying target distribution you're interested in. You can tell if it pretty well approximates the underlying distribution if it keeps getting things right, and you don't have to keep making excuses for it (say, by saying that "it's complicated").

Anyway, we can view ANNs, SVMs, and other statistical learners as essentially inductive systems, in the sense that, given a set of prior examples, they learn a rule (classifier) that allows us to predict something about new, unseen examples. They generalize, in the sense that unseen examples that match a profile (learned from the training data), even if not an exact fit, may still be classified correctly. It's not a simple one-to-one match. Hence, generalize.

The problem with the generalization performance of all such systems is two-fold. One, they're limited by the set of features that were chosen (a person selects features that are relevant, like "is followed with 'inc.'" for classifying organization mentions in free text). Two, even given optimal feature selection, the generalization performance of inductive systems is always hostage to the information in the training data. We collect examples to use to train the system, and, in the end, we hope that the training data was adequate to model the true distribution (the thing we really want to predict). Problem is, whatever hasn't occurred yet in this true distribution, can't possibly be collected from past examples, and so the entire approach is hostage to the fact that things change in real-life environments, and the larger "pattern" of the true distribution as it unfolds in time may not be captured in the training data. Whenever this happens, the approximation function does not model the true distribution. Bummer.

Now, a feature in particular of such inductive systems (we can substitute "supervised statistical learning sytems" if it sounds more fancy) is this local minima or maxima worry, which I introduced with regard to ANNs, but which is really just a handy way of introducing the general problem of generalizing from past cases to future ones writ large. And it is a problem. Consider time sequence prediction (as opposed to, say, sequence classification such as the well-known document classification task in IR research). In time sequence prediction, the goal is take a sequence of elements at time t, and predict the next element at time t+1. Applying this multiple times you can predict a large sequence of elements through some time n.

And this is where the inductive problem comes in, because if the data you're using to predict the next elements came from some set of prior elements, it's possible that the prior elements (your training data), gave you a model that will get you stuck on a foothill, or, will see the top of a small hill as a bottom valley, and so on. You can't be sure, because the future behavior of the true distribution you don't have. And this is why, in the end, induction--however fancy it gets dressed up in mathematical clothing--not only can be wrong, in theory, but often is, in practice.

We can't see into the future, unfortunately. If we could, we could fix the inductive problem by simply adding in the additional information about the true distribution that we're missing in our training data. But in that case, of course, we'd hardly need the approximation.

2 comments:

mijopo said...

"And this is why, in the end, induction--however fancy it gets dressed up in mathematical clothing--not only can be wrong, in theory, but often is, in practice"

I see two issues coming together here, one is why and how ML algorithms tend to fail. It's an interesting and important CS question, but it's not clear to me how we get from there to impugning induction. Some problems with ML result from the fact that our training data isn't great, but does it follow that induction itself is often wrong in theory in practice, I'm not quite sure why we can't/shouldn't just stop at "often ML tools are wrong". After all, aren't our ML algorithms just tools to help us perform inductive reasoning rather than the inductive reasoning itself?

Erik said...

They are indeed just tools. The problem comes in when various disciplines, like economics, or the social sciences, or various other sciences, try to imbue these tools with powers that they (or we) don't have, like trying to predict future outcomes with messy, complex data. Very, very hard.

And yes, I agree that induction itself is strictly speaking independent of the modern computational techniques used to help perform inductive inferences. But, then, this raises the question of the general limits of induction for us in general; that is, the power of humans to perform induction irrespective of the use of computational models.

More specifically, the species of induction I'm getting at is time sequence prediction (as opposed to, say, sequential learning, like named entity recognition, or sequence classification, like document classification). How good are we at seeing the future with the use of computational models? How good are we without them?

So, if the idea is to consider our inductive practices without the use of computational methods, a look at past predictions reveals that we're dismal. If the idea is to include modern computational methods, it may be surprisingly to some that, in fact, we're still dismal. My point is simply that we can't see into the future, with or without the computational methods.