NCA-GENM Question 88
Select 3A data science team is developing a multimodal AI model that integrates image and text data to generate descriptive captions for images. During testing, the model produces captions with high accuracy on the training data but fails to generalize well on unseen validation data. Which of the following steps would best assist in addressing this issue?
- A
Increase the size of the training dataset by incorporating more diverse image-text pairs.
- B
Fine-tune the model hyperparameters, such as learning rate and batch size, using the validation data.
- C
Use data augmentation techniques to synthetically expand the dataset with variations of existing samples.
- D
Replace the text-processing module with a larger language model to improve text understanding.
- E
Conduct error analysis on validation samples to identify specific failure patterns.
Show answer and explanation
Correct answers: A, C, E
Explanation
Generalization issues often stem from overfitting or insufficient diversity in training data. Adding diverse samples and applying data augmentation help improve model robustness, while error analysis provides insights into the root causes of poor performance on validation data. These approaches collectively address the failure to generalize without introducing risks like data leakage.
- A. Correct.
Increasing the size of the training dataset with diverse samples can improve the model's ability to generalize to unseen data by reducing overfitting to the training set.
- B. Incorrect.
Fine-tuning hyperparameters using validation data may lead to data leakage and overfitting, which would worsen the generalization problem.
- C. Correct.
Data augmentation introduces variations to the dataset, helping the model generalize by exposing it to a wider range of scenarios during training.
- D. Incorrect.
Replacing the text-processing module with a larger language model may enhance text understanding but does not directly address the generalization issue. It might also increase computational costs unnecessarily.
- E. Correct.
Error analysis on validation samples can help pinpoint specific weaknesses in the model, guiding targeted improvements in training or architecture.