NVIDIA NCA-GENM 問題集

質問 1：
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. Replace ResNet50 with a simpler image encoder like a shallow CNN to reduce computational complexity.
B. Freeze the weights of both the ResNet50 and BERT models to prevent overfitting.
C. 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.
D. Decrease the batch size during training.
E. Increase the learning rate for the BERT model to prioritize text-based information.
正解：C
解説: (Topexam メンバーにのみ表示されます)

質問 2：
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)

質問 3：
You are building a multimodal model to classify news articles using both text and images. The text data is processed using spaCy, and image data is processed using Keras. You've noticed that the model is heavily biased towards the text dat a. Which of the following techniques would be MOST effective in addressing this modality imbalance?
A. Reducing the dimensionality of the image feature vectors using Principal Component Analysis (PCA).
B. Normalizing the length of text sequences to a fixed size before feeding into the model.
C. Using data augmentation techniques on the image dataset, such as random rotations and flips.
D. Applying TF-IDF to the text data to reduce the impact of common words.
E. Implementing modality-specific weighting in the loss function, giving a higher weight to the image loss.
正解：E
解説: (Topexam メンバーにのみ表示されます)

質問 4：
You are tasked with analyzing a large dataset of images used for training a generative A1 model. The dataset contains noisy labels and varying image quality. Which of the following preprocessing steps are MOST crucial for improving the performance of your model?
A. Using a pre-trained image quality assessment model to filter out low-quality images.
B. Resizing all images to a fixed resolution (e.g., 256x256).
C. Implementing a label smoothing technique to mitigate the impact of noisy labels.
D. Converting all images to grayscale to reduce computational complexity.
E. Applying aggressive data augmentation techniques like random rotations and flips.
正解：A,C
解説: (Topexam メンバーにのみ表示されます)

質問 5：
You are tasked with generating realistic images of human faces using a GAN. However, you notice that the generated images often contain artifacts, such as distorted facial features or unrealistic textures. Which of the following techniques would be most effective in improving the realism and quality of the generated faces?
A. Employing a StyleGAN architecture with adaptive instance normalization (AdalN) and mapping network.
B. Applying L1 regularization to the generator's weights.
C. Training the GAN for fewer epochs.
D. Using a smaller batch size.
E. Using a simpler discriminator architecture.
正解：A
解説: (Topexam メンバーにのみ表示されます)

質問 6：
You are tasked with building a system that can answer questions based on both an image and a corresponding text description. The image is represented as a feature vector from a CNN, and the text is represented as a sequence of word embeddings from a pre-trained language model. Which architecture would be most suitable for this task?
A. A recurrent neural network (RNN) that processes the text and then uses the final hidden state to attend to the image features.
B. A Transformer-based architecture with cross-attention mechanisms that allow the model to attend to both the image and text features simultaneously.
C. Two separate models, one for processing images and another for processing text, with the final answer being chosen based on the higher confidence score.
D. A combination of a CNN and an LSTM, where the CNN processes the image and the LSTM processes the text independently.
E. A simple feedforward neural network that concatenates the image and text feature vectors.
正解：B
解説: (Topexam メンバーにのみ表示されます)

質問 7：
You are developing a multimodal model that combines text and tabular data for predicting customer churn. The text data consists of customer reviews, and the tabular data includes demographics and transaction history. You've preprocessed both datasets. Which of the following approaches would be the MOST effective for integrating these modalities?
A. Concatenate the raw text and tabular data into a single feature vector.
B. Convert the text data into numerical features using techniques like TF-IDF, then concatenate these features with the tabular data.
C. Use a Transformer-based model to encode the text and a separate neural network for the tabular data, then fuse the embeddings.
D. Train separate models for text and tabular data, then average their predictions.
E. All of the above.
正解：B,C
解説: (Topexam メンバーにのみ表示されます)