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

1. You are experimenting with different architectures for a text-to-speech (TTS) model. You have implemented a Tacotron 2 model and a FastSpeech 2 model. Which of the following statements accurately describes the key differences between these two architectures and their implications?

A) Tacotron 2 uses an attention mechanism for aligning text and speech, while FastSpeech 2 relies on a fixed alignment, resulting in faster training and inference for FastSpeech 2 but potentially lower quality.
B) Both Tacotron 2 and FastSpeech 2 use attention mechanisms, but FastSpeech 2 incorporates length regulator and variance adaptor modules to address the one-to-many mapping problem, leading to more stable and controllable synthesis.
C) Tacotron 2 is an autoregressive model, while FastSpeech 2 is a non-autoregressive model. This allows FastSpeech 2 to generate speech in parallel, resulting in significantly faster inference speeds.
D) Both Tacotron 2 and FastSpeech 2 rely on fixed alignments, but FastSpeech 2 uses a more complex decoder architecture, leading to higher quality but slower inference.
E) FastSpeech 2 uses an attention mechanism for aligning text and speech, while Tacotron 2 relies on a fixed alignment, resulting in faster training and inference for Tacotron 2 but potentially lower quality.

2. You are working on a project that involves generating realistic images from text descriptions using a diffusion model. You want to reduce the inference time of the model, which currently takes several minutes to generate a single image. Which of the following techniques would be MOST effective for accelerating inference without significantly compromising image quality?

A) Switching to a CPU-based inference engine.
B) Using a smaller batch size during inference.
C) Employing techniques like DDIM (Denoising Diffusion Implicit Models) or progressive distillation to reduce the number of sampling steps required.
D) Increasing the number of diffusion steps.
E) Training the diffusion model with a larger dataset.

3. You are working on a multimodal model for video captioning, where the model needs to generate captions describing the actions and events happening in a video. You notice that the model tends to focus only on the most salient objects in the scene and ignores subtle but important actions. Which of the following techniques can help the model attend to these subtle actions and generate more comprehensive captions?

A) Using a larger batch size.
B) Increasing the learning rate during training.
C) Adding more layers to the LSTM or GRIJ used for sequence modeling.
D) Decreasing the regularization strength.
E) Implementing a hierarchical attention mechanism that first attends to relevant time steps and then to relevant regions within those time steps.

4. Consider the following Python code snippet used for evaluating a generative model. What potential issue exists with this code, and how would you rectify it to ensure a robust evaluation?

A) The 'inception_score' function is deprecated and should be replaced with a newer metric like FID.
B) The number of generated images Cn_generated_imageS) is too small for a reliable Inception Score calculation and should be increased to at least 50,000.
C) The code is fundamentally correct and does not have any issues.
D) The Inception Score is not an appropriate metric for evaluating generative models, and other methods should be used instead.
E) The code does not account for the stochastic nature of generative models and needs to be run multiple times with different random seeds to get a stable estimate of the Inception Score.

5. Consider a generative AI model that combines text and audio inputs to produce a musical composition. The text input is a description of the desired mood and style, while the audio input is a short melody. Which of the following loss functions would be MOST appropriate for training this model?

A) Wasserstein loss (Earth Mover's Distance)
B) A combination of perceptual loss (based on audio features like pitch and timbre) and a style loss (based on the text description's mood).
C) Cross-entropy loss
D) Mean Squared Error (MSE) loss
E) Hinge Loss

質問と回答：