NVIDIA NCA-GENM 問題集

質問 1：
You are tasked with fine-tuning a pre-trained multimodal model for a new task involving image and text inputs. The pre-trained model was trained on a large dataset of image-caption pairs. Which of the following strategies would be MOST effective for transfer learning in this scenario, considering computational efficiency and performance?
A. Fine-tune all layers of the pre-trained model with a very small learning rate.
B. Fine-tune a subset of layers, specifically those responsible for feature extraction from both image and text modalities, while keeping the lower layers frozen.
C. Fine-tune only the classification head (output layer) while freezing all other layers of the pre-trained model.
D. Train a new model from scratch on the new task's dataset.
E. Use knowledge distillation to transfer knowledge from the pre-trained model to a smaller, more efficient model.
正解：B
解説: (Topexam メンバーにのみ表示されます)

質問 2：
You're building a system that generates images from text descriptions, incorporating spatial relationships. For instance, the text 'A red ball is to the left of a blue cube' should result in an image where the red ball is actually positioned to the left of the blue cube. Which of the following approaches would be MOST suitable for encoding and utilizing spatial information in this text-to-image generation process?
A. Augmenting the text encoder with explicit spatial relation embeddings that represent the relative positions between objects. Use these embeddings to modulate the image generation process (e.g., through attention mechanisms).
B. Applying a pre-trained object detector to the generated image and penalizing the model if the spatial relationships are incorrect
C. Relying solely on the image decoder to learn spatial relationships implicitly from the text description during training.
D. Using a bag-of-words representation for the text, ignoring word order and spatial relationships.
E. Using a standard Transformer architecture for text encoding without any specific spatial awareness mechanisms
正解：A
解説: (Topexam メンバーにのみ表示されます)

質問 3：
You're developing a system to generate realistic 3D models from text descriptions. You're using a diffusion model-based approach and find that the generated models often lack fine details and exhibit artifacts. Which of the following techniques would likely lead to the MOST significant improvement in the quality of the generated 3D models?
A. Implement classifier-free guidance with a higher guidance scale.
B. Increase the number of diffusion steps during the reverse diffusion process.
C. Train the diffusion model on a larger dataset of text-3D model pairs.
D. Use a larger IJ-Net architecture for the denoising process.
E. All of the above
正解：E
解説: (Topexam メンバーにのみ表示されます)

質問 4：
You are evaluating a multimodal model that generates descriptions for video clips. You have human ratings for the relevance, fluency, and coherence of the generated descriptions. Which statistical test is MOST appropriate for determining if there is a statistically significant difference in the median ratings for each of these criteria (relevance, fluency, coherence) between two different versions of your model?
A. Kruskal-Wallis test
B. ANOVA
C. Friedman Test
D. T-test
E. Mann-Whitney U test
正解：E
解説: (Topexam メンバーにのみ表示されます)

質問 5：
You are building a system that uses audio and video to detect emotional states of a user. What are the challenges to this system?
A. Subjectivity in emotional expression across cultures and individuals.
B. Differences in lighting conditions influencing facial expression recognition.
C. Variations in background noise affecting audio quality.
D. Synchronization issues between audio and video streams.
E. All of the above.
正解：E
解説: (Topexam メンバーにのみ表示されます)

質問 6：
You are building a multimodal generative A1 model that creates realistic indoor scenes by combining textual descriptions, floor plans (geospatial data), and object libraries. The goal is to generate high-quality 3D models of the scenes. However, the model often produces scenes with physically implausible object arrangements (e.g., objects floating in the air, overlapping furniture). How can you MOST effectively integrate physical constraints into the generation process to ensure more realistic scene compositions?
A. Use a physics engine (e.g., NVIDIA PhysX) as a post-processing step to simulate the generated scene and correct any physically implausible object placements.
B. Increase the size of the training dataset with more examples of realistic indoor scenes.
C. Implement a rule-based system that enforces basic physical constraints (e.g., objects must be supported by a surface, no object interpenetration) during the generation process.
D. Train a separate discriminator network that evaluates the physical plausibility of generated scenes and penalizes implausible configurations during training.
E. Force the model to generate only scenes that exist within the training set.
正解：A,C,D
解説: (Topexam メンバーにのみ表示されます)

質問 7：
You are tasked with optimizing a multimodal model that combines audio and text data for speech recognition. The model currently struggles with noisy audio environments. Which data augmentation technique would be MOST effective in improving the model's robustness to noise?
A. Translating the text into different languages and back.
B. Randomly masking parts of the text input.
C. Adding Gaussian noise to the audio data.
D. Normalizing the text data to lowercase.
E. Rotating the images used for visual context.
正解：C
解説: (Topexam メンバーにのみ表示されます)