What does "Exposure Bias" mean?
Table of Contents
Exposure bias happens when a model trained to generate text makes predictions based on correct sequences during training. However, when the model is actually used, it has to rely on its own previous words to predict the next ones. This difference between training and real use can lead to mistakes and less accurate results.
Why is it a Problem?
When a model learns only from the correct input, it doesn't get enough practice in making decisions based on its own outputs. This can make it less reliable when it needs to create text on its own. As a result, the quality of the generated text can suffer, and the model may struggle with tasks like translation or story writing.
Solutions to Exposure Bias
To address exposure bias, researchers use methods like scheduled sampling. This approach slowly introduces the model to its own predictions during training. By doing so, the model becomes better at handling the real situation where it has to work with its own generated text. Newer methods enhance this even further by focusing on making the model’s behavior more like a teacher model, which helps improve its overall performance.