Error-riddled data sets are warping our sense of how good AI really is

A recent MIT study shows that the 10 most cited AI data sets are rife with data labeling errors - which confuses the public's understand of how much progress AI is really making. Correcting for these labeling errors showed different results based on the complexity of the models. In some cases, it appears that for more complex models we may actually be over-estimating their accuracy compared to more simplistic models that are available.

Read full article