Reading figures and tables

Figures and tables hold the evidence. Learn to read them before the text, and you will understand papers much faster.

For every figure, ask

• What is on each axis, and in what units? Is the y-axis starting at zero? A cut axis can make a small difference look huge.
• What is being compared? The new method against which baselines?
• Are there error bars or shaded areas? They show how much results vary. Check the caption: they may mean standard deviation, standard error or a confidence interval, which are different things.
• How many runs or samples? One run can be luck.
• Is the scale linear or logarithmic? On a log scale, equal steps mean multiplying, not adding.

For every results table, ask

• Which number is the main metric, and is higher or lower better?
• Is the best number in bold, and how far ahead is it? A 0.2-point lead with no variance reported may be noise.
• Are all methods tested on the same data, with the same budget?
• Are the baseline numbers re-run by the authors, or copied from older papers that used different setups?

ExampleA table shows the new model at 91.3% and the best baseline at 91.1%. No variance is reported. Before believing the new model is better, you would want several runs with different random seeds, and the spread of results.

Read the captions

A good caption tells you what the figure shows and how to read it. Many papers put important details there, such as the number of runs.

RememberIf the figures and the text tell different stories, trust neither until you understand why.