A PICTURE is said to be worth a thousand words. That metaphor might be expected to pertain a fortiori in the case of scientific papers, where a figure can brilliantly illuminate an idea that might otherwise be baffling. Papers with figures in them should thus be easier to grasp than those without. They should therefore reach larger audiences and, in turn, be more influential simply by virtue of being more widely read. But are they? Bill Howe and his colleagues at the University of Washington, in Seattle, decided to find out.
First, they trained a computer algorithm to distinguish between various sorts of figures—which they defined as diagrams, equations, photographs, plots (such as bar charts and scatter graphs) and tables. They exposed their algorithm to between 400 and 600 images of each of these types of figure until it could distinguish them with an accuracy greater than 90%. Then they set it loose on the more-than-650,000 papers (containing more than 10m figures) stored on PubMed Central, an online archive of biomedical-research articles.