NVIDIA Generative AI Multimodal 認定 NCA-GENM 試験問題:
1. You're training a conditional GAN to generate images of birds based on text descriptions. The GAN generates images, but they lack fine- grained details and often have artifacts. Which of the following techniques are MOST likely to improve the quality and realism of the generated images? (Select TWO)
A) Reducing the size of the input noise vector to the generator.
B) Using a more powerful discriminator architecture (e.g., with attention mechanisms).
C) Implementing spectral normalization in both the generator and discriminator.
D) Using a deeper and wider generator network (e.g., with more layers and channels).
E) Using a simple Multi-Layer Perceptron (MLP) as the generator.
2. You are fine-tuning a pre-trained multimodal model for a new task. You have limited computational resources. Which of the following fine-tuning strategies would be the MOST computationally efficient while still achieving good performance?
A) Freeze all layers except the classification head and fine-tune only the classification head.
B) Freeze the lower layers of the model and fine-tune the upper layers and the classification head.
C) Train a new random model from scratch for the task, which will avoid the need to load the pre-trained model.
D) Fine-tune all the layers of the model.
E) Randomize the model to train, if it improves the training rate.
3. Which of the following techniques is most appropriate for mitigating the vanishing gradient problem in very deep neural networks, particularly when training generative models?
A) Data augmentation
B) Weight decay
C) Dropout
D) Early stopping
E) Residual connections (skip connections)
4. Consider the following Python code snippet using PyTorch Lightning and a Hugging Face Transformers model for multimodal classification. Which of the following code snippets is MOST appropriate to perform gradient accumulation in this context, assuming you want to accumulate gradients over 4 batches?
A) 
B) 
C) 
D) 
5. You are tasked with building a system that generates realistic images based on both textual descriptions and a semantic segmentation map. The segmentation map provides spatial information about the objects present in the scene. Which of the following generative architectures is MOST appropriate for this multimodal task?
A) Variational Autoencoder (VAE)
B) Diffusion model without conditioning
C) Autoregressive model like PixelCNN
D) Vanilla Generative Adversarial Network (GAN)
E) Conditional Generative Adversarial Network (cGAN) with both text and segmentation map as conditions.
質問と回答：
| 質問 # 1 正解： C、D | 質問 # 2 正解： B | 質問 # 3 正解： E | 質問 # 4 正解： C | 質問 # 5 正解： E |
1190 お客様のコメント
品質保証JPexamはIT認定試験のシラバスに従って、試験問題の範囲を正確に絞って、的中率が99％の最新問題集を捧げます。
1年間の無料更新サービスJPexamは1年以内に問題集の無料更新サービスを提供し、お客様がいつでも最新版の問題集を持つことを保証いたします。もし試験の内容が変更されたら、弊社は直ちにお客様にお知らせします。それに、弊社の問題集が更新されたら、早速メールで最新バージョンを送付いたします。
全額返金JPexamの問題集を利用すると、短時間で勉強しても試験に合格できるのを保証いたします。試験に不合格になってしまった場合、弊社は全額返金いたします。(
ご購入前のお試しJPexamは問題集のサンプルを無料で提供いたします。ご購入前にサンプルを試用して製品の品質を確認することができます。ご遠慮なく利用してください。
