Databricks Machine Learning Associate Question 326
Single answerYou are using a hyperparameter tuning framework in Databricks to optimize a machine learning model. Initially, you set the number of trials to 10 and observe a noticeable improvement in model accuracy after each trial. You decide to increase the number of trials to 100. What is the most likely outcome of increasing the number of trials?
- A
The model accuracy will continue to improve indefinitely as the number of trials increases.
- B
The model accuracy may improve as the number of trials increases, but the rate of improvement will likely diminish.
- C
Increasing the number of trials guarantees finding the most optimal hyperparameters for the model.
- D
The model accuracy will stop improving after a certain number of trials, as the search space will have been fully explored.
Show answer and explanation
Correct answer: B
Explanation
In hyperparameter tuning, there is often a diminishing return on model accuracy as the number of trials increases. While increasing the number of trials improves the chances of finding better hyperparameter configurations, the improvement rate decreases over time as the search algorithm converges on optimal regions of the hyperparameter space. This makes Option 2 the most accurate answer.
- A. Incorrect.
Model accuracy does not improve indefinitely due to diminishing returns; the relationship is not linear, and after a certain point, the improvements become minimal.
- B. Correct.
This is correct. While increasing the number of trials provides more opportunities to explore the hyperparameter space, the improvement in accuracy typically diminishes as the most optimal regions of the hyperparameter space are approached.
- C. Incorrect.
Increasing the number of trials increases the chances of finding better hyperparameters, but it does not guarantee finding the global optimum, especially if the search strategy is not exhaustive or if the model is inherently limited.
- D. Incorrect.
While it is true that improvements might plateau, it is inaccurate to say that the search space will be fully explored, as hyperparameter spaces are often vast and rarely fully explored in practice.