A company wants to serve a support chatbot at high volume with lower per-request cost. It trains a compact model to reproduce the answers of a much larger model, and the compact model then handles the traffic. What is this technique called?
Show answer and explanation
Answer: B
Knowledge distillation trains a smaller student model to imitate a larger teacher model's outputs, so it keeps much of the teacher's behavior at a fraction of the inference cost. This is the approach described in the scenario. Data augmentation creates more varied training examples, dimensionality reduction compresses feature representations, and hyperparameter tuning searches training settings. None of those produces a compact model that copies a larger one's behavior.
Why the other options are wrong
- A. Data augmentation expands a training set with altered or synthetic examples. It does not create a smaller model that copies a larger one.
- C. Dimensionality reduction shrinks the number of input features, such as with PCA. It does not transfer a large model's behavior into a compact one.
- D. Hyperparameter tuning searches for better training settings. It does not produce a smaller student model from a teacher.