Transfer learning is essentially using a deep neural network (many hidden layers) trained to perform a task on a specific data set, and using that trained network to make predictions on a similar, but different, data set by editing the final output layer.
In facial recognition, one can take advantage of this by using an already trained model (let’s say on celebrity faces) which extracts important (and somewhat unknown) features of the human face that only a computer can distinguish as unique identifiable features. The model can than be edited to extract those same features from faces from a different data source (say CA State legislators) with incredible accuracy.
Read more about transfer learning and facial recognition in the following paper. Although the model was only able to achieve 91% accuracy, this is due to the fact that our data source had several mislabeled faces.
Link to Kaggle Competition