NVIDIA Generative AI Multimodal Sample Questions:

1. You're tasked with building a model that can generate recipes from images of food. You decide to use a Variational Autoencoder (VAE) architecture. What would be a suitable loss function combination for this task, considering both reconstruction accuracy and recipe relevance?

A) KL divergence loss + Cosine Similarity loss between the generated image embedding and a text embedding of a random recipe.
B) Reconstruction loss (MSE) between the input image and the decoded image only
C) KL divergence loss only-
D) Reconstruction loss (MSE) + Perceptual loss (based on a pre-trained image classifier) only
E) Reconstruction loss (MSE) + KL divergence loss + Cross-entropy loss between the generated recipe and a plausible recipe given the generated image embedding-

2. Which NVIDIA SDK would be most appropriate for building a real-time, interactive avatar that can respond to voice commands and generate realistic facial expressions?

A) Avatar Cloud Engine (ACE)
B) NeMo
C) Riva
D) Triton Inference Server
E) RAPIDS

3. You are tasked with building a multimodal generative AI model to create marketing content from product images and descriptions. The image encoder uses a pre-trained ResNet50 model, and the text encoder uses a pre-trained BERT model. After initial training, the generated content frequently misinterprets the image. Which of the following strategies is MOST effective in improving the model's ability to correctly interpret the image within the multimodal context?

A) Freeze the weights of both the ResNet50 and BERT models to prevent overfitting.
B) Replace ResNet50 with a simpler image encoder like a shallow CNN to reduce computational complexity.
C) Increase the learning rate for the BERT model to prioritize text-based information.
D) Fine-tune the ResNet50 model with a dataset of images specifically related to the product domain, using a contrastive loss function that encourages representations of images and corresponding text to be close in the embedding space.
E) Decrease the batch size during training.

4. You are developing a multimodal generative A1 model that takes both image and text inputs. The image branch uses a ResNet50 pre- trained on ImageNet, while the text branch uses a BERT model. To effectively combine the features, you need to align their representations. Which of the following techniques is MOST suitable for projecting the image and text features into a common embedding space?

A) Training separate linear projection layers for both ResNet50 and BERT outputs, followed by concatenation.
B) Direct concatenation of ResNet50 and BERT output features.
C) Employing Contrastive Learning with a shared embedding space and using positive and negative pairs of image and text.
D) Using Principal Component Analysis (PCA) to reduce the dimensionality of ResNet50 and BERT features before concatenation.
E) Fine-tuning the entire ResNet50 and BERT models jointly on the multimodal dataset.

5. Which of the following is NOT a typical application or benefit of using U-Net architectures in generative AI, particularly within the context of image generation and manipulation?

A) Image inpainting and super-resolution tasks.
B) Medical image analysis, such as tumor detection.
C) Facilitating efficient feature extraction and upsampling for detailed image generation.
D) Image segmentation and pixel-wise classification.
E) Encoding high-dimensional text data for multimodal embeddings.

Solutions: