NVIDIA Generative AI Multimodal Sample Questions:

1. You are working on a project that involves generating music from video. The approach uses a pre-trained video encoder and a pre- trained music decoder. You find that the generated music often lacks a clear connection to the visual content of the video. To improve the coherence between the video and the generated music, which of the following steps would be the MOST effective? (Select TWO)

A) Train the video encoder and music decoder separately on larger datasets.
B) Remove the video encoder and generate music directly from random noise.
C) Fine-tune the entire system end-to-end with a loss function that encourages temporal alignment between video and music features.
D) Introduce a cross-modal attention mechanism to allow the music decoder to attend to relevant visual features during music generation.
E) Only use videos that are shorter than 5 seconds.

2. You are tasked with building a multimodal generative A1 model that takes both image and text as input to generate a coherent video. Which of the following architectures is MOST suitable for this task, considering the need to fuse information from different modalities and generate sequential data?

A) A Generative Adversarial Network (GAN) trained solely on image data and later fine-tuned with text embeddings.
B) A Support Vector Machine (SVM) classifier trained to predict the next frame based on image and text features.
C) A standard Convolutional Neural Network (CNN) followed by a fully connected layer.
D) A simple recurrent neural network (RNN) that concatenates image feature vectors and text embeddings as input at each time step.
E) A Transformer-based architecture with separate encoders for image and text, followed by a decoder that generates video frames.

3. You are fine-tuning a pre-trained large language model (LLM) for a specific text generation task using LoRA (Low-Rank Adaptation).
Which of the following statements accurately describes the benefits and limitations of using LoRA?

A) LoRA is not compatible with model parallelism techniques.
B) LoRA can improve the accuracy of the fine-tuned model compared to full fine-tuning by preventing overfitting.
C) A and B.
D) LoRA reduces the number of trainable parameters by inserting low-rank matrices into the original model layers, making fine-tuning more memory-efficient.
E) LoRA allows for efficient task switching by only storing and loading the small LoRA parameters for different tasks, while keeping the original LLM weights frozen.

4. You're working with a client to develop a generative A1 model for creating personalized marketing content. During requirements acquisition, the client expresses a desire for 'highly creative' and 'unique' outputs. However, they struggle to articulate specific aesthetic preferences. How would you best approach translating these subjective requirements into concrete model training and prompt engineering strategies?

A) Implement a system for interactive prompt refinement, allowing the client to iteratively modify prompts and observe the resulting outputs in real-time, facilitating a collaborative exploration of the model's creative potential.
B) Focus solely on quantitative metrics like perplexity and FID score to ensure the model generates diverse and high-quality content, assuming that 'creative' and 'unique' will naturally emerge.
C) B and D
D) Conduct extensive A/B testing with a large user group, presenting them with various model outputs and gathering feedback on which content they perceive as most 'creative' and 'unique'. Use this feedback to refine the model and prompts.
E) Use a pre-trained style transfer model to apply different artistic styles to the generated content, offering the client a diverse range of options to choose from and identify their preferred aesthetic.

5. Consider a scenario where you are developing a system for automatically generating product descriptions based on images and specifications. The system needs to generate diverse and creative descriptions. Which of the following techniques would be MOST helpful in achieving this?

A) Using a rule-based system to extract keywords from the image and specifications and then assemble them into sentences.
B) Training a recurrent neural network (RNN) from scratch to generate the descriptions.
C) Fine-tuning a pre-trained language model (e.g., GPT-3) on a dataset of product descriptions, using the image features as a conditional input.
D) Using a simple template-based approach with predefined sentence structures.
E) Employing a denoising autoencoder to clean the images before feeding them into the description generation model.

Solutions: