NCA-GENM Question 112
Select 3A team is building a generative AI model that integrates multimodal data, including text descriptions, satellite imagery, and audio recordings, to predict environmental changes in a specific region. During the data preparation phase, they notice that some satellite images are missing, and several audio recordings have incomplete segments. Which of the following actions should the team prioritize to address these challenges and ensure data quality?
- A
Impute missing satellite images using generative techniques or similar historical data.
- B
Discard all incomplete audio recordings to maintain consistency in the dataset.
- C
Use modality-specific preprocessing techniques to handle missing or incomplete data.
- D
Develop a unified data curation pipeline to assess and handle quality issues across all modalities.
- E
Augment the text descriptions to compensate for missing satellite and audio data.
Show answer and explanation
Correct answers: A, C, D
Explanation
Handling multimodal data requires addressing missing or incomplete information in a way that preserves the integrity and diversity of the dataset. Imputing missing data, applying modality-specific preprocessing techniques, and creating a unified data curation pipeline are critical steps to ensure the data is suitable for training generative AI models. Simply discarding incomplete data or relying solely on augmenting other modalities are suboptimal solutions that might lead to a loss of valuable information.
- A. Correct.
Imputing missing satellite images using generative techniques or historical data can help recover the missing information, ensuring that this modality remains useful for the model.
- B. Incorrect.
Discarding all incomplete audio recordings would result in significant data loss, which is not ideal unless the majority of the recordings are unusable. A better approach is to handle incomplete data with targeted techniques.
- C. Correct.
Using modality-specific preprocessing techniques is essential when dealing with missing or incomplete data, as each modality (e.g., images, audio) requires tailored strategies to address its unique challenges.
- D. Correct.
Developing a unified data curation pipeline ensures consistent handling of issues across all modalities and improves the overall quality of the dataset.
- E. Incorrect.
Augmenting text descriptions may provide some additional context, but it cannot fully compensate for missing satellite and audio data, as these modalities often contain unique and critical information.