Core Machine Learning and AI Knowledge
Free NVIDIA-Certified Associate: Generative AI Multimodal practice — 6 questions on Core Machine Learning and AI Knowledge, with explanations. No sign-up. Full 12-question mixed test →
A research team is building a text-to-image generation system and must choose between a GAN and a diffusion model. Their primary requirement is stable training and avoidance of collapsed output diversity across varied prompts. Which model family BEST meets these requirements?
Diffusion models learn to reverse a gradual noising process via a denoising objective, avoiding the adversarial min-max dynamics that cause GAN training instability, and they reliably produce diverse, high-fidelity samples across prompts.
In the original Transformer architecture, sinusoidal positional encodings are added to token embeddings instead of using learned positional embeddings. What is the primary advantage of this design choice, as commonly tested?
Sinusoidal functions are parameter-free and their fixed periodicity lets the model extrapolate to positions beyond the maximum length observed during training, unlike learned positional embeddings which are only defined up to the training length.
CLIP (Contrastive Language-Image Pretraining) learns a joint embedding space for images and text without using manually assigned class labels. Which learning paradigm does CLIP's training objective represent?
CLIP uses a contrastive loss that treats matched image-text pairs as positives and mismatched pairs as negatives, deriving the supervisory signal from the inherent pairing structure of the data itself, which defines self-supervised learning.
A team has trained a diffusion-based image generation model and needs to quantitatively evaluate both the visual fidelity and the diversity of generated images relative to real images. Which evaluation metric is MOST appropriate for this scenario?
FID compares the statistics (mean and covariance) of Inception-network feature embeddings between generated and real image sets, capturing both fidelity and diversity of the generative distribution.
A multimodal model generates text captions conditioned on image content: a frozen pretrained vision encoder extracts image features, and a language decoder attends directly to those features while generating each output token. Which multimodal fusion pattern does this describe, as used in models like BLIP-2 and Flamingo?
BLIP-2 and Flamingo insert cross-attention layers into the language decoder that let it attend to visual encoder features during generation, fusing modalities at an intermediate representational level rather than at input or output.
When generating images with a diffusion model using classifier-free guidance, a practitioner increases the guidance scale far beyond the recommended range for a given prompt. What is the MOST likely effect on the output?
Excessively high guidance scale pushes the denoising trajectory far from the unconditional score estimate, over-emphasizing prompt conditioning at the cost of naturalness, typically producing oversaturated, artifact-heavy, less diverse images.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$99.99 $34.99 with code FREETEST33 — valid through September 9.
Get my $34.99 deal →