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

1. Consider a scenario where you're building a multimodal model to generate image captions. You've pre-trained a large language model (LLM) on a massive text corpus and a convolutional neural network (CNN) on ImageNet. How would you effectively combine these pre- trained components for your image captioning task, considering the need to maintain high caption quality and training efficiency?

A) Freeze the CNN, extract image features, and train the LLM to generate captions from these features.
B) Freeze the LLM, train the CNN to predict text embeddings, and then decode these embeddings into captions.
C) Fine-tune both the CNN and the LLM jointly on the image captioning dataset.
D) Train the CNN and LLM separately on unrelated datasets and then combine them at inference time using a simple averaging of their outputs.
E) Use a transformer-based encoder to process both image features and text embeddings before feeding them to the LLM decoder.

2. Consider the following code snippet, where you are trying to load image and text data for a multimodal model. What is the most likely cause of error if the code fails during the image loading step?

A) The system doesn't have CUDA drivers installed.
B) The text data is not in IJTF-8 encoding.
C) The learning rate is set too high.
D) The batch size is too large.
E) The image files are corrupted or in an unsupported format.

3. You are fine-tuning a large pre-trained language model for a specific downstream task. During training, you observe that the model performs well on the training data but generalizes poorly to the validation dat a. Which of the following strategies could help improve the model's generalization performance?

A) Decrease the learning rate.
B) Increase the weight decay (L2 regularization).
C) Increase the training data size by collecting more data.
D) Increase the learning rate.
E) Implement early stopping based on the validation loss.

4. You are training a conditional GAN (cGAN) to generate images of animals based on text descriptions. Which of the following is the most crucial difference in the training process of a cGAN compared to a regular GAN?

A) The generator in a cGAN receives both random noise and the condition (text description) as input.
B) The discriminator in a cGAN does not receive generated images.
C) The cGAN does not require a discriminator.
D) The discriminator in a cGAN only receives real images as input.
E) The cGAN uses a different loss function than a regular GAN.

5. You are building a real-time image captioning system using a Transformer model. You observe significant latency issues when generating captions for high-resolution images. Which optimization strategies would be most effective in reducing the latency without significantly sacrificing caption quality? (Select all that apply)

A) Use mixed precision training and inference, leveraging Tensor Cores on NVIDIA GPUs.
B) Decrease the batch size during inference.
C) Increase the number of Transformer layers to improve caption accuracy, even if it increases latency.
D) Implement knowledge distillation, training a smaller, faster student model to mimic the behavior of the larger Transformer model.
E) Employ model quantization techniques (e.g., INT8 quantization) to reduce model size and memory bandwidth requirements.

質問と回答：