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

1. You are working with a transformer-based multimodal model that processes both text and audio. You want to implement an efficient attention mechanism that reduces the computational cost associated with attending to the entire input sequence. Which of the following attention mechanisms would be MOST suitable for achieving this goal?

A) Scaled Dot-Product Attention
B) Multi-Head Attention
C) Sparse Attention
D) Local Attention
E) Global Attention

2. You're building a multimodal model that integrates text, images, and audio. The text data has many missing values. Which of the following strategies would be MOST effective for handling missing text data while leveraging the other modalities?

A) Remove all data points with missing text values to ensure data integrity.
B) Train a separate model to predict the missing text based on the available image and audio data, then impute the predicted values.
C) Use a multimodal generative model (e.g., VAE, GAN) to impute the missing text based on the learned joint representation of all modalities.
D) Ignore the missing text values during training, assuming the model can learn from the available modalities.
E) Use a simple imputation method like replacing missing text with a placeholder like 'unknown'.

3. You are building a Generative Adversarial Network (GAN) to generate high-resolution images. The generated images suffer from mode collapse, where the generator only produces a limited variety of images. Which of the following techniques would be MOST effective in mitigating mode collapse?

A) Using a simpler generator architecture.
B) Using mini-batch discrimination or feature matching in the discriminator.
C) Reducing the size of the generator's latent space.
D) Decreasing the learning rate of the discriminator.
E) Increasing the learning rate of the generator.

4. You are using NeMo to fine-tune a pre-trained language model for a specific text generation task. You want to implement a custom data augmentation technique to improve the model's robustness. Which of the following approaches is most appropriate for integrating your custom augmentation within the NeMo framework?

A) Create a custom *Dataset* class that inherits from 'nemo.core.Dataset' and implements your augmentation within the '_getitem
B) Augment the data directly within the training loop, applying transformations to each batch before feeding it to the model. method.
C) Use a separate data processing pipeline outside of NeMo and save the augmented data to disk before training.
D) Monkey-patch the existing NeMo data loading functions to inject your augmentation logic.
E) Modify the core NeMo library files to directly incorporate your augmentation logic.

5. You are deploying a text-to-speech application using NVIDIA Riv
a. The application needs to handle a large volume of concurrent requests with minimal latency. Which of the following Riva deployment configurations would be MOST appropriate?

A) Deploying Riva on a single CPU core with a small batch size.
B) Deploying Riva across multiple GPUs using Triton Inference Server with dynamic batching.
C) Deploying Riva on CPU using multiprocessing.
D) Deploying Riva on a single GPU using TensorRT for model optimization, without using Triton Inference Server.
E) Deploying Riva on a single GPU with a large batch size.

質問と回答：