NCA-GENM Question 103
Select 3You are developing a multimodal generative AI model that processes both text and image inputs. During evaluation, you notice that the model performs well on text-only inputs but struggles with combined text-image inputs. Which approach would most effectively test the data quality and consistency in this multimodal setting?
- A
Check if the text and image data are properly aligned and correspond to the same context.
- B
Ensure that the text data is free from grammatical errors and spelling mistakes.
- C
Verify that the images in the dataset have consistent resolution and format.
- D
Randomly shuffle the text and image pairs in the dataset to test generalization.
- E
Test the dataset for balanced representation across all input modalities.
Show answer and explanation
Correct answers: A, C, E
Explanation
In a multimodal setting, maintaining alignment between modalities (e.g., text and image) and ensuring data quality for each modality is essential for testing and improving model performance. Additionally, balanced representation across modalities helps the model generalize better, avoiding bias toward one modality over the other.
- A. Correct.
Checking alignment between text and image data is critical in multimodal settings to ensure the inputs are contextually meaningful.
- B. Incorrect.
While eliminating grammatical errors can improve text data quality, it does not address multimodal consistency, which is the specific issue here.
- C. Correct.
Ensuring consistent resolution and format for images helps maintain data quality and prevents errors in the model's image processing pipeline.
- D. Incorrect.
Randomly shuffling text and image pairs disrupts the alignment between modalities, which would further degrade performance instead of testing consistency.
- E. Correct.
Testing for balanced representation across modalities ensures that the model is exposed to diverse and unbiased multimodal inputs.