Monthly Archives: November 2017

This post was prompted by a line—more properly, part of a line, from John Ashbery’s “Unctuous Platitudes” (Houseboat Days), which you can hear Ashbery read here (mp3 link). Here is the line:

The weather has grown gray with age.

It’s an effective metaphor, immediately giving the reader a feeling not just for the state of the weather, but for the feeling associated with it. This feeling is complex and difficult to summarize, but involves the sense that the weather is well-worn, that it is not something new. It oppresses as familiar things oppress.

But my interest is less in what this particular metaphor does in “Unctuous Platitudes” than in what it suggests about metaphor more generally. How, exactly, does this metaphor work? I suggest that it works by a form of overfitting.

Overfitting is generally discussed in scientific contexts, where it is an example of bad practice. In a well-controlled experiment, data is generated whose patterns predominantly reflect the operation of the cause of interest, while the effects of others causes are (a) minimized and (b) appropriately distributed. Appropriately distributed in the sense that, while each individual data point reflects the operation of many causes and so deviates from the expectation were only the cause of interest operating, these deviations are not systematic, and wash out as many data points accumulate.

In fitting a curve to such data, the goal is thus not to account for each individual data point exactly, but to capture the general trend they reveal, which, if all goes well, is the product of the cause of interest. It is, however, always mathematically possible to find a complex curve that fits the data exactly. This is known as overfitting. The image below provides an example of proper curve-fitting (black line) and overfitting (blue line).


The temptation of overfitting is that it allows one to capture the data at hand arbitrarily well, as the image shows. The cost is that predictive power is lost. Overfit curves tend to be completely wrong about where the next data points will be. What proper curve-fitting loses with respect to the particular data set, it thus gains by being much better suited to predicting future data. As prediction is one of the essential functions of scientific hypotheses, overfitting is indeed bad science.

Poetry, by contrast, does not aim at prediction (usually; never say never and all that). It generally aims much more at capturing precisely a particular scene (whether real or invented) in all of its specificity. In poetry, the data at hand are all there is to capture. In scientific contexts, the temptation to overfit is countered by the need to predict. There is no such need in poetry; therefore, there is no penalty for overfitting.

All of this suggests that metaphor may be a form of overfitting, at least in many cases. The Ashbery metaphor with which I began is just such a case. In comparing the gray of the weather with the gray of age, Ashbery certainly does not capture any causal regularity. Weather does not proceed from a starting point to an end point, but is broadly cyclical. Extend the metaphor, then, and it is woefully wrong about what the weather will be like the day after this poem takes place. But that day never occurs, has never been written, and there is no need to account for it. By ignoring it, Ashbery is able to condense a number of features of the day’s weather, the products of innumerable causes, into a single description that captures it, if not with perfect precision, at least with more than a scientifically respectable curve would enjoy.