Multimodal Data
Free NVIDIA-Certified Associate: Generative AI Multimodal practice — 6 questions on Multimodal Data, with explanations. No sign-up. Full 12-question mixed test →
A team is building a CLIP-style vision-language model. Their corpus contains 50 million images but only 2 million have paired captions; the rest are uncaptioned. Which strategy BEST maximizes use of the full dataset while preserving alignment quality for contrastive training?
Self-training/pseudo-labeling with a captioning model trained on the paired subset lets the team leverage the full 50M images while generating usable, if noisy, supervision signal for contrastive fine-tuning — a standard technique for scaling multimodal alignment beyond the paired subset.
A ViT-B/16 model processes a 224x224x3 RGB image using non-overlapping patches. What is the resulting patch sequence length fed into the transformer encoder, excluding the class token?
With patch size 16, each spatial dimension yields 224/16 = 14 patches, giving 14x14 = 196 total patches.
When extracting mel-spectrogram features for an audio-to-text pipeline, increasing the FFT hop length while keeping the window (frame) size constant primarily results in which trade-off?
Hop length controls the stride between successive analysis windows, so a larger hop length produces fewer, more widely spaced frames (coarser time resolution); frequency resolution is governed by the window/FFT size, not the hop length.
A multimodal dataset for text-to-speech alignment has 500,000 text-only samples but only 8,000 samples with paired audio recordings. The team wants to train a joint text-audio embedding without simply duplicating the scarce audio-paired examples. Which approach is the BEST practice?
Combining weighted sampling of the paired subset with modality-specific augmentation increases exposure and effective diversity of the scarce paired data without introducing pure duplicates or fabricated signal, which is standard practice for imbalanced multimodal training.
Which statement correctly distinguishes paired from unpaired multimodal training data?
This matches the technical definitions used across the field: CLIP-style contrastive models require paired image-text data, while approaches like CycleGAN use cycle-consistency losses specifically because their image-to-image datasets are unpaired.
A user submits a 120-word descriptive prompt to a text-to-image diffusion pipeline whose text encoder (a CLIP ViT-L/14 text tower) has a maximum context length of 77 tokens. The tokenized prompt exceeds this limit. What is the standard behavior in most production pipelines?
The CLIP text encoder has a fixed context length of 77 tokens (including BOS/EOS special tokens); standard tokenizer behavior truncates excess tokens rather than erroring, which is a well-known prompt engineering consideration.
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 →