NVIDIA Certified Associate: Generative AI Multimodal — Free Practice Questions
15 free NVIDIA Certified Associate: Generative AI Multimodal practice questions with every answer explained. Covers all exam domains, no signup needed.
15 questions · every answer explained · free to practise
8 of the 15 questions in this set, with the correct answer marked and every option explained.
1. A team is combining a text encoder and an image encoder into a single classifier. They concatenate the raw feature vectors from both encoders immediately after the input layer and train the rest of the network on the joined representation. Which fusion strategy is this?
✓Early fusion
Early fusion combines the modalities at or near the input stage, before most of the model's processing, so the network learns joint representations across modalities throughout its depth. Concatenating raw features immediately after the input layer is the defining example.
✗Late fusion
Late fusion keeps each modality in its own separate model all the way through and combines only the final predictions or scores. The combination here happens at the input, not at the output.
✗Intermediate fusion
Intermediate fusion merges modality-specific representations partway through the network, after each branch has done some independent processing. Here the merge happens immediately, with no independent processing first.
✗Cross-validation fusion
This is not a fusion strategy. Cross-validation is an evaluation technique for estimating model performance; it has nothing to do with how modalities are combined architecturally.
2. What does CLIP-style contrastive language-image pretraining actually learn?
✓A shared embedding space in which the representation of an image and the representation of its matching caption are close together, and mismatched pairs are pushed apart.
That contrastive objective — pull matched image-text pairs together, push mismatched pairs apart, within one shared space — is the whole idea. It is what enables zero-shot classification by comparing an image embedding against embeddings of candidate text labels.
✗A generative model that produces images directly from noise without any text conditioning.
That describes an unconditional diffusion or GAN model. CLIP is not a generator at all; it is an encoder pair that scores how well an image and a text match.
✗A translation model that converts captions from one language to another using images as an interlingua.
CLIP is not a translation system. It aligns images and text in a shared space, but it does not generate text output in another language.
✗A supervised classifier trained on a fixed taxonomy of object labels.
The opposite is the point. CLIP trains on loosely paired image-caption data rather than a fixed label taxonomy, which is precisely why it generalizes to categories it was never explicitly trained to classify.
3. In a denoising diffusion model, what does the model learn to predict during training?
✓The noise that was added to a sample at a given timestep, so that it can be subtracted to move one step back toward a clean image.
Training corrupts real data with a known amount of noise at a random timestep and asks the network to predict that noise. At generation time the model repeatedly predicts and removes noise, walking a pure-noise sample back to a coherent image.
✗The final clean image in a single forward pass, with no iterative steps.
Single-pass generation describes a GAN or a VAE decoder. The defining characteristic of diffusion is that generation is iterative, taking many denoising steps.
✗The discriminator's confidence that a sample is real rather than generated.
That is a GAN concept. Diffusion models have no adversarial discriminator; they are trained with a straightforward regression loss against the known noise.
✗A discrete token sequence representing the image, decoded by a language model.
That describes autoregressive image generation over discrete visual tokens, a different family. Diffusion operates in continuous space over noise levels.
4. A model must accept audio, images and text in the same pipeline. What must happen to each of these before a neural network can process them together?
✓Each modality must be converted into numeric tensors in a consistent representation the network can consume, such as spectrograms for audio, pixel tensors or patch embeddings for images, and token embeddings for text.
Networks consume numbers. Every modality needs a modality-appropriate encoding into tensors before fusion is even possible, which is why multimodal pipelines always begin with per-modality preprocessing.
✗Every modality must be converted to text, since transformers can only process language.
Transformers operate on sequences of embeddings, not on language specifically. Vision transformers consume image patches and audio transformers consume spectrogram frames without any conversion to text.
✗All three must be downsampled to the same resolution and duration.
Modalities have incompatible natural units — resolution means nothing for text and duration means nothing for a still image. Each is encoded appropriately for its own type; alignment happens in embedding space, not in raw units.
✗Audio and images must be discarded if a text description is available, to avoid redundancy.
This defeats the purpose of multimodal modelling. The value of these systems is precisely that different modalities carry complementary information that a caption alone does not.
5. A team has a small labelled dataset for a specialized visual inspection task and limited compute budget. Which approach best fits these constraints?
✓Transfer learning — start from a model pretrained on a large general dataset and fine-tune it on the small task-specific dataset.
Transfer learning is the standard answer to 'small data, limited compute'. The pretrained backbone already encodes general visual features, so only comparatively little task-specific data and training time is needed to adapt it.
✗Train a large model from random initialization on the small dataset.
With little data this overfits badly, and training from scratch is exactly the compute expense the constraint rules out.
✗Increase model capacity until training loss reaches zero.
Driving training loss to zero on a small dataset is memorization, not learning. Validation performance would degrade even as training loss improves.
✗Collect no additional data and rely entirely on prompt engineering of a text-only LLM.
A text-only model cannot perform visual inspection. Prompting is not a substitute for a model that can actually see the images.
6. A generated image consistently misses a detail requested in the prompt. The team wants more control over the output without retraining the model. Which lever should they reach for first?
✓Refine the context embeddings and prompt conditioning, and tune guidance strength, testing systematically against a fixed set of prompts.
Prompt conditioning and guidance are the intended control surface for a trained diffusion model, and the blueprint explicitly frames this as experimental work: test and refine the context embeddings to achieve the desired output. No retraining required.
✗Retrain the model from scratch with a larger dataset.
That is the most expensive possible response to a controllability problem and is explicitly ruled out by 'without retraining'. Conditioning should be exhausted first.
✗Reduce the number of denoising steps to force the model to commit earlier.
Fewer steps generally reduces fidelity rather than improving adherence to the prompt. It trades quality for speed, which is not what is being asked for.
✗Switch from GPU to CPU inference for more deterministic results.
Hardware choice does not change what the model attends to in the prompt. It would only make generation dramatically slower.
7. When evaluating two versions of a multimodal model, a team compares them using only the single best sample produced by each. Why is this methodologically unsound?
✓Cherry-picking a single best sample measures the luck of one draw rather than the distribution of outputs the model actually produces, so it cannot support a claim that one version is better.
Generative models are stochastic. A fair comparison needs many samples per model, ideally with fixed seeds and a defined metric or structured human evaluation, so that the comparison reflects typical behaviour instead of the extreme tail.
✗It is sound, provided the same prompt is used for both models.
Holding the prompt constant removes one source of variation but not the sampling randomness, which is the dominant one here. A single draw from each model is still a sample size of one.
✗It is unsound only if the two models have different parameter counts.
Parameter count is irrelevant to the flaw. The problem is the sample size and the selection bias of choosing the best output, which applies regardless of model size.
✗It is unsound because generated images cannot be evaluated quantitatively at all.
They can — FID, CLIP score and human preference studies are all established quantitative approaches. The flaw is the methodology, not the impossibility of measurement.
8. In an end-to-end conversational AI pipeline built with NVIDIA Riva, what is the correct order of the three core stages?
✓Automatic speech recognition, then natural language processing, then text-to-speech.
Speech comes in and must first be transcribed (ASR), the resulting text is then understood and a response formed (NLP), and finally the response is spoken back (TTS). This ASR → NLP → TTS ordering is exactly how the Riva pipeline is presented in the course path for this exam.
✗Text-to-speech, then automatic speech recognition, then natural language processing.
This begins by speaking before anything has been heard or understood, which is not a coherent pipeline.
✗Natural language processing, then automatic speech recognition, then text-to-speech.
NLP cannot run first because there is no text yet — the audio has not been transcribed. ASR must precede it.
✗Automatic speech recognition, then text-to-speech, then natural language processing.
This would speak the raw transcription back before any understanding or response generation, making the NLP stage pointless.
7 more questions in the app
Practise the full 15-question set with a timer, scoring and progress tracking.