Google

The performance gains from output ensembling are not limitless. As the number of models in an ensemble grows, a point of diminishing returns is often reached where adding more models provides only marginal improvements in output quality, or none at all.

Diminishing Returns in Output Ensembling

A machine learning team has an ensemble of 5 models for a critical task. They observe that each new model added to the ensemble provides a smaller accuracy boost than the last. They are now considering two options for their next development cycle, both requiring similar engineering effort:

*   **Option 1:** Add 5 more models to the ensemble, which is projected to increase overall accuracy by 0.2% and significantly increase computational costs.
*   **Option 2:** Focus on improving the quality of the training data for the existing 5 models, which is projected to increase overall accuracy by 1.5% without increasing computational costs.

Which option should the team choose? Justify your recommendation by explaining the underlying principle that makes one option more efficient than the other in this scenario.

Ensemble Scaling Strategy

Evaluate the junior engineer's proposal from the case study below. Is this a sound strategy for improving the system? Justify your reasoning.

Evaluating an Ensemble Scaling Strategy

A development team is creating an ensemble to improve the accuracy of a text summarization model. They measure the quality score of the summaries as they increase the number of models in the ensemble, with the following results:

| Number of Models | Quality Score |
|------------------|---------------|
| 1                | 78.0          |
| 3                | 84.0          |
| 5                | 86.5          |
| 7                | 87.5          |
| 9                | 87.7          |

Based on t

Learn Before

Related