Unveiling Unexpected Training Data in Internet Video.

The article examines the data mining of internet video in the training of machine learning systems. Alleged shortcomings of unsupervised learning based on internet videos are identified and the author calls for the adoption of a "self-supervised" approach. The example of computing depth maps based on real estate videos found on the YouTube platform using a parallax scene format called MPI is proffered along with an example of speech isolation and speaker identification.