NCA-GENM Question 15
Select 2You are building a generative AI model to create image captions by training a multimodal neural network. During the model development, you need to evaluate its performance to ensure it generalizes well to unseen data. Which of the following techniques are most appropriate to achieve this?
- A
Using cross-validation to assess model performance across multiple data splits
- B
Performing feature engineering to extract meaningful features from the input data
- C
Comparing the model's performance with baseline models like a random or heuristic-based model
- D
Overfitting the model on the training data to maximize accuracy
- E
Using a single train-test split to evaluate the model's performance
Show answer and explanation
Correct answers: A, C
Explanation
To ensure a generative AI model generalizes well, techniques like cross-validation and comparing against baseline models are essential. Cross-validation provides a more comprehensive evaluation by testing across multiple data splits, while comparing against baselines ensures that the model outperforms simpler alternatives. These methods help validate the robustness and utility of the model.
- A. Correct.
Cross-validation is a robust technique to evaluate model performance by splitting the dataset into multiple folds, ensuring the model's generalization is tested on different data splits.
- B. Incorrect.
Feature engineering is important for improving model inputs, but it does not directly evaluate model performance. It is more relevant during data preparation.
- C. Correct.
Comparing the model with baseline models is a critical step to ensure your model's performance is better than simple or heuristic-based approaches. This helps validate the utility of your generative AI model.
- D. Incorrect.
Overfitting the model on training data is harmful as it leads to poor generalization on unseen data. While training accuracy may be high, overfitting does not help in evaluating true performance.
- E. Incorrect.
Using a single train-test split can lead to biased results as the performance depends heavily on how the split is done. Cross-validation is a better option for thorough evaluation.