Researchers have found that by training large neural networks on tasks such as predicting missing words and identifying puzzle pieces, the hidden layer representation can be transferred to different tasks with great success.