Abstract:

Visual Inference Protocols have been shown to be useful in several Statistical Inference tasks. The plot of a dataset is viewed as another statistic of the data, and as such, it can be "recalculated" a few times on a permuted dataset, producing several plots. These additional plots can be put side to side with the original plot as in a typical lineup and if the original dataset indeed contains a signal over noise, e.g. a correlation in a scatterplot  the original plot must emerge to the naked human eye as "different" or "strongest" of all plots. But a plot is an image. Recent advances in Deep Neural Networks have proved computers can extract all sorts of interesting information from images. Will a computer be able to "see" the correlation the same way a human can? Will it choose correctly the original plot in a lineup of permuted plots? In this talk we will show how well a computer can do Visual Inference in several tasks, after training a deep neural network on tens of thousands of sampled datasets. As often is the case it is interesting to focus on the cases where the computer got it wrong and wonder: was it really wrong?
