NVIDIA Generative AI Multimodal Sample Questions:

1. A multimodal dataset consists of video footage of human actions and corresponding wearable sensor data (accelerometer, gyroscope). The goal is to predict the type of action being performed. However, the sensor data is noisy and often misaligned with the video frames. Consider the following code snippet designed to synchronize and clean the sensor data:
What is the primary purpose of the 'resample' function in this code, and what potential issues might arise from using a simple aggregation method during resampling?

A) The 'resample' function filters the sensor data and .mean() only returns the most relevant sensor data
B) The 'resample' function decreases the video framerate to the rate of the sensor. Using .mean()' is only useful if there is no noise in the sensor data
C) The 'resample' function aligns the sensor data to the video frame rate. Using is appropriate as it averages out the noise in the sensor data.
D) The 'resample' function increases the sensor data frequency. Using .mean()' is only useful if there is no noise in the sensor data
E) The 'resample' function aligns the sensor data to the video frame rate. Using '.mean()' might smooth out important peaks and valleys in the sensor data, potentially losing crucial information.

2. Which of the following techniques is LEAST likely to improve the performance of a Generative A1 model tasked with generating realistic images from text descriptions?

A) Increasing the size of the training dataset with high-quality image-text pairs.
B) Implementing classifier-free guidance during the diffusion process.
C) Using a more powerful generative architecture, such as a Transformer-based diffusion model.
D) Reducing the dimensionality of the text embeddings used as input to the image generator.
E) Applying data augmentation techniques to the training images, such as random cropping and rotations.

3. You have trained a multimodal model to generate descriptions for recipes, using images of the finished dish as one modality and a list of ingredients as another. When evaluating the generated descriptions, you notice the descriptions are factually correct in terms of ingredients but often fail to capture the stylistic nuances or tone of professionally written recipes. Which evaluation strategy would provide the most insightful feedback on this aspect of the model's performance?

A) Conduct a human evaluation where individuals rate the generated descriptions based on factors such as informativeness, clarity, and stylistic appropriateness.
B) Use a pre-trained language model to assess the fluency and coherence of the generated descriptions.
C) Calculate the BLEU score between the generated descriptions and a set of reference recipes.
D) Measure the perplexity of the generated descriptions using a language model trained on a corpus of recipe books.
E) Compute the ROUGE score between the generated descriptions and a set of reference recipes.

4. Consider the following code snippet used for evaluating a Generative Adversarial Network (GAN):

What does the code snippet calculate, and what do 'images1' and "images2 represent in the context of GAN evaluation?

A) Calculates the Structural Similarity Index (SSIM); 'images1 and 'images2 represent real and fake images respectively.
B) Calculates the Peak Signal-to-Noise Ratio (PSNR); 'images1 and 'images2 represent real and fake images respectively.
C) Calculates the Kernel Inception Distance (KID); 'images1' and 'images2 represent real and fake images respectively.
D) Calculates the Inception Score; 'images1' and 'images2 represent real and fake images respectively.
E) Calculates the Frechet Inception Distance (FID); 'images1' and 'images2 represent real and fake images respectively.

5. You are developing a system that uses a generative A1 model deployed with Triton Inference Server to create personalized avatars. You want to ensure that the system is robust against malicious inputs designed to generate offensive or harmful content. Which of the following security measures are most critical to implement in conjunction with Triton?

A) Rate limiting the number of requests per user, implementing input validation and sanitization to filter out potentially harmful prompts, and regularly auditing the generated content for offensive material.
B) Implementing input validation to enforce a maximum length for text prompts, using a content moderation API to filter the output generated by the model, and logging all user activity for auditing purposes.
C) All of the above.
D) Encrypting all communication between the client and the Triton server using HTTPS, restricting access to the Triton server to a private network, and using strong passwords for all Triton accounts.
E) Using a web application firewall (WAF) to protect the Triton server from denial-of-service attacks, implementing role-based access control (RBAC) to limit user privileges, and regularly updating the Triton software to patch security vulnerabilities.

Solutions: