Called the Inclusive Images Competition, the data challenge focuses on how new algorithmic techniques can be used to successfully train deep learning computer vision models on highly geographically skewed datasets, yet not inherit their biases. Using the example of a wedding ceremony, Google demonstrates how important such geographic cultural generalization is and how biased many of the most popular training datasets are. One of the greatest challenges confronting the deep learning era is the availability of training data. Google could easily have simply funded a global competition to acquire training imagery from every corner of the earth. And indeed, the company has actually embarked upon just such a sweeping effort to globalize its own training datasets.

Source: Forbes September 11, 2018 15:11 UTC