NVIDIA NCA-GENM

NVIDIA Generative AI Multimodal 認定 NCA-GENM 試験問題:

1. You are developing a generative A1 model to create music based on textual descriptions of mood and genre. You have a dataset of paired text descriptions and music tracks. When evaluating the generated music, you realize it's difficult to objectively quantify the quality of the music. Which of the following evaluation methods would provide the MOST comprehensive assessment of the generated music's quality and alignment with the text descriptions?

A) Conduct a human evaluation study where participants rate the generated music on various subjective criteria such as pleasantness, mood alignment, and overall quality based on the text description provided.
B) Calculate the Root Mean Square (RMS) energy of the generated music tracks.
C) Measure the file size of the generated music tracks.
D) Calculate the Bit Rate of the generated music tracks.
E) Use a pre-trained music genre classification model to predict the genre of the generated music and compare it to the genre in the text description.

2. You're training a conditional GAN (cGAN) to generate images of handwritten digits conditioned on the digit label. You notice that the generated images are blurry and lack fine details, even after extensive training. Which of the following techniques could you implement to improve the sharpness and realism of the generated images?

A) Increase the learning rate of the generator.
B) Use spectral normalization on both the generator and discriminator.
C) Implement a perceptual loss function in addition to the adversarial loss.
D) Increase the dimensionality of the latent space.
E) Add batch normalization layers to the generator and discriminator.

3. You are building a multi-modal model that combines text and image data for a search application. The goal is to retrieve relevant images given a text query. You have encoded both images and text into embeddings. What's a suitable loss function for training the model to ensure images relevant to a text query are ranked higher than irrelevant ones?

A) Cross-entropy loss
B) KL Divergence
C) Contrastive Loss
D) Triplet Loss
E) Mean Squared Error (MSE)

4. Which of the following are valid techniques for dealing with overfitting in a deep learning model trained on image data?

A) Reducing the amount of training data
B) Using data augmentation techniques.
C) Adding Ll or L2 regularization.
D) Increasing the complexity of the model.
E) Implementing dropout layers.

5. You are building a system to translate spoken language into images. You have a large dataset of audio clips and corresponding images.
Which of the following is the MOST appropriate architecture?

A) A CNN for audio feature extraction, followed by a GAN for generating images conditioned on those features.
B) A transformer-based model that attends to both audio features and a learned visual vocabulary to generate images.
C) A Support Vector Machine (SVM) trained on audio features to classify the type of image to generate.
D) A hidden Markov model (HMM) trained to map audio features to image segments.
E) A sequence-to-sequence model with an LSTM encoder for the audio and an LSTM decoder for generating image pixels directly.

質問と回答：