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

1. When training a Variational Autoencoder (VAE) for generating new data points, which of the following objectives does the VAE optimize?

A) Only A and B.
B) Maximizing the similarity between the input data and the reconstructed data.
C) Maximizing the likelihood of the input data given the latent representation.
D) Minimizing the Kullback-Leibler (KL) divergence between the learned latent distribution and a prior distribution (e.g., a Gaussian distribution).
E) All of the above.

2. You are working with a dataset containing text descriptions of products and corresponding product images. You want to train a model that can retrieve the most relevant image for a given text description. Which of the following loss functions is MOST appropriate for this task?

A) Binary Cross-entropy loss
B) Mean Squared Error (MSE) loss
C) Triplet loss
D) Huber Loss
E) Cross-entropy loss

3. You're training a multimodal model to generate 3D models from text descriptions. The models are evaluated using Intersection over Union (IOU) between the generated and ground truth 3D models. During evaluation, you observe perfect IOU scores on some samples, but visual inspection reveals significant discrepancies. What is the MOST likely cause for this, and what can be done to correct the process?

A) The IOU calculation is being performed incorrectly, or there is a bug in the evaluation code. Verify the IOU implementation.
B) IOU is an inherently flawed metric for evaluating 3D models and needs to be replaced by Chamfer distance.
C) The model is overfitting, resulting in near-perfect reconstruction of a subset of training samples. Reduce the model's capacity.
D) The text descriptions are too simple. Use more complex text prompts to prevent overfitting.
E) There is a data leakage issue, where some of the test data is present in the training data. Ensure that training and test data are completely disjoint.

4. Which of the following techniques can be used to reduce the computational cost and memory footprint of large language models (LLMs) during inference?

A) Quantization
B) Knowledge Distillation
C) Increasing the model size
D) Adding more layers
E) Pruning

5. You have a text-to-image model deployed using Triton Inference Server. You want to monitor the GPU utilization and inference latency to ensure optimal performance. Which of the following methods is the MOST effective way to achieve this?

A) Using the Triton Inference Server client API to measure inference latency from the client-side.
B) Relying solely on the operating system's resource monitor to track GPU usage.
C) Using 'nvidia-smi' to periodically check GPU utilization and manually calculate latency.
D) Using Triton's built-in Prometheus metrics endpoint and Grafana for visualization.
E) Writing custom scripts to parse Triton's log files and extract performance metrics.

質問と回答：