NVIDIA NCA-GENM 問題集

質問 1：
You are training a multimodal generative A1 model that takes text and images as input to generate videos. During experimentation, you observe that the model performs well on common scenarios (e.g., 'a dog playing in the park') but struggles to generate coherent videos for less frequent or abstract scenarios (e.g., 'the concept of time flowing'). What is the MOST effective strategy to improve the model's performance on these challenging scenarios, focusing on test data quality?
A. Train the model for a significantly longer duration on the existing training data.
B. Implement data augmentation techniques on the existing training data, focusing on color adjustments and minor image transformations.
C. Increase the size of the training dataset by duplicating existing common scenario examples.
D. Reduce the complexity of the model architecture to prevent overfitting on the common scenarios.
E. Curate a new test dataset specifically containing challenging scenarios and use it to evaluate and fine-tune the model. Ensure the new test data includes diverse interpretations and variations of the abstract concepts.
正解：E
解説: (Topexam メンバーにのみ表示されます)

質問 2：
You're training a Generative Adversarial Network (GAN) to generate realistic images of faces. After several epochs, you notice that the generator is producing very similar faces, lacking diversity. Which of the following techniques could BEST address this mode collapse issue?
A. Use a simpler generator architecture.
B. Reduce the noise injected into the generator's input.
C. Increase the batch size used for training the generator
D. Decrease the learning rate of the discriminator.
E. Implement Minibatch Discrimination in the discriminator.
正解：E
解説: (Topexam メンバーにのみ表示されます)

質問 3：
You're training a VQA (Visual Question Answering) model. During evaluation, you notice the model performs well on common object recognition tasks but struggles with questions requiring reasoning about object relationships or scene understanding. What are the MOST effective strategies to improve the model's performance on these complex reasoning tasks? (Choose two)
A. Train the model on a larger dataset with more diverse and complex questions/answers.
B. Decrease the learning rate.
C. Increase the size of the image embedding.
D. Replace the LSTM with a simpler RNN in the question encoder.
E. Use a more sophisticated attention mechanism that attends to relevant image regions based on the question.
正解：A,E
解説: (Topexam メンバーにのみ表示されます)

質問 4：
You're working with a text-to-image generation model. After training, you notice the generated images lack fine-grained details and appear blurry. Which hyperparameter tuning strategy would be MOST effective in improving the visual quality of the generated images, considering the computational cost?
A. Optimizing the learning rate schedule.
B. Decreasing the batch size.
C. Adding more layers to the discriminator network (if using GANs).
D. Increasing the number of training epochs.
E. Switching to a different model architecture (e.g., from VAE to GAN).
正解：A
解説: (Topexam メンバーにのみ表示されます)

質問 5：
You're building a generative A1 model that can create realistic 3D models from text descriptions. You have a dataset of text descriptions and corresponding 3D models, but the alignment between the text and the 3D models is weak. The model sometimes generates 3D shapes that don't accurately reflect the text. Which of the following techniques could improve the alignment between the text descriptions and the generated 3D models?
A. Applying data augmentation techniques to the 3D models (e.g., random rotations and scaling).
B. Training the model with a larger batch size.
C. Using a pre-trained text encoder (e.g., BERT or CLIP) to extract meaningful features from the text descriptions.
D. Increasing the number of vertices and faces in the 3D models.
E. Using a contrastive loss function that encourages the model to generate 3D models that are semantically similar to the corresponding text descriptions.
正解：C,E
解説: (Topexam メンバーにのみ表示されます)

質問 6：
You are working with a large dataset of images to train a Generative A1 model. You suspect that some images are corrupted or of poor quality, which could negatively impact training. Which of the following methods would be the MOST effective in identifying and removing these problematic images?
A. Calculate the average pixel intensity for each image and remove those with very low or very high average intensity.
B. Compute the image sharpness (e.g., using Laplacian variance) and remove images with low sharpness values.
C. Manually inspect each image and remove those that appear to be corrupted or low quality.
D. Use a pre-trained image quality assessment model (e.g., BRISQUE, NIQE) to score each image and remove those with low scores.
E. Check for file corruption errors during image loading and remove those files.
正解：B,D,E
解説: (Topexam メンバーにのみ表示されます)

質問 7：
Which data augmentation techniques are MOST suitable for improving the robustness of a multimodal model that uses images and text?
A. Rescaling images and changing the font of the text.
B. Adding Gaussian noise to images and randomly deleting words in the text.
C. Changing the image resolution and increasing the text size.
D. Randomly cropping images and translating text to different languages.
E. Rotating images and back-translating text (translating to another language and back to the original).
正解：B,E
解説: (Topexam メンバーにのみ表示されます)