Learned systems in the domain of visual recognition and cognition impress in part because even though they are trained with datasets many orders of magnitude smaller than the full population of possible images, they exhibit sufficient generalization to be applicable to new and previously unseen data. Since training data sets typically represent such a small sampling of any domain, the possibility of bias in their composition is very real. But what are the limits of generalization given such bias, and up to what point might it be sufficient for a real problem task? There are many types of bias as will be seen, but we focus only on one, selection bias. In vision, image contents are dependent on the physics of vision and geometry of the imaging process and not only on scene contents. How do biases in these factors—that is, non-uniform sample collection across the spectrum of imaging possibilities—affect learning? We address this in two ways. The first is theoretical in the tradition of the Thought Experiment. The point is to use a simple theoretical tool to probe into the bias of data collection to highlight deficiencies that might then deserve extra attention either in data collection or system development. Those theoretical results are then used to motivate practical tests on a new dataset using several existing top classifiers. We report that, both theoretically and empirically, there are some selection biases rooted in the physics and imaging geometry of vision that challenge current methods of classification.
Rouhani et al. (Thu,) studied this question.