NVIDIA Generative AI Multimodal 認定 NCA-GENM 試験問題:

1. You are building a multimodal generative A1 application that uses CLIP to align text and image embeddings. You observe that the generated images lack detail and fidelity to the text prompt. Which of the following strategies would be MOST effective in improving image quality, and how could prompt engineering and Triton Inference Server play a role?

A) Using a larger batch size during CLIP training and increasing the learning rate. Triton is not directly involved in model training.
B) Increasing the CLIP model's text encoder's hidden layer size and using more aggressive data augmentation during CLIP training. Triton can be used to serve the augmented CLIP model at scale.
C) Training a separate image super-resolution model to enhance the generated images after they are produced by the CLIP-guided generator. Triton can manage the concurrent execution of the generator and super-resolution models.
D) All of the above
E) Refining the text prompts to be more descriptive and specific, incorporating stylistic details and relevant keywords. Triton can optimize the prompt embedding process.

2. A research team has developed a novel multimodal model that fuses text, image, and audio dat a. They want to quantitatively evaluate the model's performance in comparison to several existing state-of-the-art models. Which of the following evaluation metrics would be MOST appropriate to assess the model's ability to generate coherent and relevant text descriptions based on the combined multimodal input?

A) Frechet Inception Distance (FID).
B) Structural Similarity Index Measure (SSIM).
C) BLEU (Bilingual Evaluation Understudy) and ROIJGE (Recall-Oriented Understudy for Gisting Evaluation).
D) Inception Score.
E) Perplexity.

3. You're building a multimodal model that predicts customer satisfaction based on their written reviews and associated call center audio recordings. You've pre-trained separate text and audio encoders. What's the MOST effective strategy to fuse these modalities for the final prediction task?

A) Train a separate attention mechanism to weigh the contributions of each modality before concatenation.
B) Concatenate the final hidden states from both encoders and feed them into a fully connected layer.
C) Fine-tune only the text encoder, keeping the audio encoder frozen.
D) Average the output embeddings of both encoders element-wise.
E) Add the hidden states of both encoders element-wise.

4. Consider a system that generates captions for images, and a key metric is BLEU score. You observe that while the BLEU score is high, the generated captions often lack detailed descriptions of the objects and relationships within the image. Which of the following strategies would you employ to improve the descriptive richness of the generated captions?

A) Train the model to minimize cross-entropy loss between predicted and ground truth captions.
B) Increase the beam size during decoding to explore a wider range of possible captions.
C) Fine-tune the model using Reinforcement Learning with a reward function that encourages detailed descriptions, such as CIDEr or SPICE.
D) Implement early stopping based solely on BLEU score during training.
E) Reduce the size of the vocabulary to focus on the most common words.

5. You're developing a multimodal model that combines text and audio for sentiment analysis. The text component is performing well, but the audio component contributes very little to the overall accuracy. What's the MOST likely reason and how could you address it?

A) The audio data is too large. Downsample the audio data to reduce computational cost.
B) The audio data is irrelevant. Remove the audio component entirely.
C) The audio features are not properly aligned with the text features. Use a cross-modal attention mechanism to improve alignment.
D) The audio data is not preprocessed correctly. Apply aggressive noise reduction techniques.
E) The text component is simply too dominant. Reduce the weight given to the text component in the final prediction.

質問と回答：