Neural networks: life after training

摘要

There has been much work done in the use of neural networks to model an existing problem, but little has been done to address what happens after training has been completed and the model must continue to learn new information. How well does the model work on information that it has not seen before? How does it adapt to new information? In this paper we address these issues, beginning our discussion with a neural model that has been trained on parsing simple natural language phrases and how well the model can generalize. Based on these results we then investigate two techniques which attempt to allow the model to "grow" or learn information that it has never before seen.