NVIDIA Generative AI Multimodal Sample Questions:

1. You are building a multimodal generative A1 system that creates 3D models from text descriptions. The system produces accurate shapes but struggles to generate realistic textures and surface details. What approach would BEST address this limitation?

A) Increase the number of parameters in the text encoder.
B) Increase the batch size during the 3D model generation phase.
C) Reduce the resolution of the generated 3D models to simplify the texture generation process.
D) Add more layers to the shape decoder.
E) Train a separate texture generation network conditioned on the generated 3D shape.

2. You are training a multimodal model that combines text and images. You observe that the model is heavily biased towards the text modality and largely ignores the image data. Which of the following strategies could you use to address this modality imbalance? (Select all that apply)

A) Decrease the learning rate for the text-related parameters of the model.
B) Use a modality-specific loss weighting scheme, assigning a higher weight to the loss component derived from the image data.
C) Oversample the image data during training.
D) Increase the learning rate for the image-related parameters of the model.
E) Reduce the dimensionality of the image features to match the dimensionality of the text embeddings.

3. You are optimizing a multimodal model for deployment on an edge device with limited memory and computational resources. The model takes video frames and audio as input to predict human actions. Which of the following optimization techniques would provide the MOST significant reduction in model size and computational complexity without drastically sacrificing accuracy?

A) Removing batch normalization layers.
B) Increasing the number of layers in the neural network.
C) Increasing the sampling rate of the audio input.
D) Applying knowledge distillation to train a smaller student model.
E) Using larger kernel sizes in convolutional layers.

4. You're training a model to generate code snippets from natural language descriptions. You are using a Transformer architecture and a large dataset of code examples. You notice the model frequently generates syntactically correct code, but the code doesn't accurately implement the described functionality (i.e., it's semantically incorrect). Select TWO methods which could improve the semantic correctness of the generated code.

A) Apply techniques like beam search during decoding to generate more diverse code snippets.
B) Increase the number of Transformer layers.
C) Fine-tune the model on a subset of the data where each code snippet is accompanied by unit tests, and train the model to also generate these tests.
D) Use a reinforcement learning approach where the reward is based on whether the generated code passes the unit tests associated with the descriptiom
E) Increase the size of the vocabulary used by the model.

5. Consider the following code snippet which aims to create a custom prompt for a Stable Diffusion model using the 'diffusers* library. The goal is to generate an image of 'a cat wearing a hat sitting on a chair'. Which of the following modifications would MOST effectively improve the quality and coherence of the generated image?

A) Using a smaller image resolution to reduce computational cost.
B) Removing the word 'sitting' from the prompt to make it more general.
C) Decreasing the number of inference steps to speed up the generation process.
D) Increasing the 'guidance_scale' and adding a negative prompt such as 'blurry, distorted'.
E) Adding more unrelated keywords to the prompt to increase diversity.

Solutions: