NCA-GENM
Free NCA-GENM practice questions
10 questions from our NCA-GENM practice tests, each with the correct answer and a written explanation. Read them with the answers shown, or switch to quiz mode and answer first.
Question 1 of 10, Experimentation
An engineer raises the classifier-free guidance scale of a diffusion model from 4 to 15 while keeping the prompt and seed fixed. Which outcome is MOST likely?
- AIdentical images, because the guidance scale only changes how long the noise scheduler runs
- BLower memory use, because high guidance skips the unconditional forward pass entirely
- CGreater sample diversity because the model now samples further from the conditional score
- DStronger adherence to the prompt, with oversaturated colors and less variety across samples Correct answer
Answer and explanation: Higher guidance pushes samples toward the conditional prediction, improving prompt adherence but often producing oversaturated, less diverse images. Diversity decreases rather than increases at high guidance. The scale changes the combined noise prediction, not the number of steps, and classifier-free guidance still needs both the conditional and unconditional predictions.
Question 2 of 10, Experimentation
A team fine-tunes a diffusion model with DreamBooth on a pet's photos. Generated images look right but always show the same pose and background as the training photos. Which change is MOST likely to improve variety?
- ATrain on the same photos for many more steps so that the identifier token becomes stronger
- BTrain on photos with varied poses and backgrounds and stop at an earlier checkpoint Correct answer
- CReplace the rare identifier token with the word 'dog' in every training caption
- DLower the guidance scale to 1 so the prompt has almost no effect on the outputs
Answer and explanation: Repeated pose and background means the model memorized limited training photos; varied inputs and earlier stopping reduce overfitting. More steps worsen memorization, guidance near 1 weakens prompt control including the subject, and replacing the identifier with 'dog' overwrites the general class concept.
Question 3 of 10, Experimentation
A captioning experiment uses beam search with beam widths of 1, 3, and 10. Captions at width 10 become generic and repetitive. Which adjustment is MOST appropriate?
- AIncrease the beam width to 50 so more candidates are explored
- BRemove the end-of-sequence token so captions are never cut off
- CUse a moderate beam with a length or repetition penalty Correct answer
- DSwitch to greedy decoding at temperature 2.0 for more variety
Answer and explanation: Large beams favor safe, generic high-probability outputs; moderate widths with length or repetition penalties balance quality and specificity. Wider beams intensify the effect, greedy decoding ignores temperature (and high-temperature sampling produces errors), and removing EOS creates runaway captions.
Question 4 of 10, Core Machine Learning and AI Knowledge
Why does a VAE use the reparameterization trick during training?
- ATo force the encoder to output discrete codes that come from a learned codebook of vectors
- BTo replace the KL divergence term with an adversarial loss from a separate discriminator
- CTo let gradients flow through the sampling step by writing the latent as mean plus scaled noise Correct answer
- DTo make the decoder output deterministic so that reconstructions never contain any noise
Answer and explanation: Sampling is not differentiable, so the latent is written as mu + sigma * epsilon with epsilon drawn from a standard normal, letting gradients reach the encoder. It does not make decoding noise-free, does not replace the KL term with an adversarial loss, and discrete codebooks are a VQ-VAE design.
Question 5 of 10, Core Machine Learning and AI Knowledge
A GAN produces sharp faces, but nearly every sample looks like the same few people regardless of the input noise. What is this failure called?
- APosterior collapse
- BExposure bias
- CMode collapse Correct answer
- DVanishing gradients
Answer and explanation: Mode collapse is when the generator covers only a few modes of the data distribution, giving low diversity. Vanishing gradients slow learning, posterior collapse is a VAE failure where the decoder ignores the latent, and exposure bias affects autoregressive sequence models.
Question 6 of 10, Multimodal Data
A team needs thousands of labeled images of rare factory defects that are hard to capture with real cameras. Which NVIDIA approach is designed for this?
- AGenerate synthetic images with Omniverse Replicator using randomized scenes and automatic labels Correct answer
- BUse TensorRT to quantize the detector so that it needs fewer labeled images for training
- CUse Triton Inference Server to serve the defect model and collect predictions as new labels
- DUse NVIDIA DALI to decode existing JPEG files faster on the GPU during each training epoch
Answer and explanation: Omniverse Replicator renders physically based synthetic scenes with domain randomization and produces ground-truth annotations automatically. DALI speeds data loading, Triton serves models, and TensorRT optimizes inference; none of them creates new labeled training data, and pseudo-labels from an undertrained model are unreliable.
Question 7 of 10, Multimodal Data
A multimodal dataset pairs dashcam video with separately recorded audio. Model accuracy is poor and inspection shows speech is offset by about 400 ms from lip movement. What should be done FIRST?
- AAlign the streams using timestamps or cross-correlation before extracting features Correct answer
- BIncrease the video frame rate so that more frames fall inside each audio analysis window
- CApply heavier data augmentation to the audio so the model ignores the offset
- DTrain separate audio and video models and average their output probabilities
Answer and explanation: The streams are misaligned, so they must be synchronized using timestamps or audio-visual cross-correlation before features are extracted. A higher frame rate does not fix the offset, augmentation hides rather than corrects it, and late fusion avoids learning the cross-modal timing the task depends on.
Question 8 of 10, Software Development
A multimodal training script works on one GPU. The team wants to run it on eight GPUs in one node with data parallelism and minimal changes. Which approach is MOST appropriate in PyTorch?
- AWrap the model in DataParallel and run it from a single Python process on one node
- BLaunch with torchrun and wrap the model in DistributedDataParallel plus a DistributedSampler Correct answer
- CStart eight copies of the script by hand, each loading the full dataset in order
- DMove half of the layers to GPU 1 and spread the remaining layers over the other GPUs in a fixed sequence
Answer and explanation: DDP launched with torchrun runs one process per GPU with efficient gradient all-reduce, and a DistributedSampler splits data. DataParallel is single-process and slower, uncoordinated copies do not share gradients, and manual layer splitting is model parallelism, not data parallelism.
Question 9 of 10, Data Analysis
A UMAP plot of CLIP embeddings for a product catalog shows one tight cluster containing images from many categories. Inspection shows all of them have a large white watermark. What does this indicate?
- AThe categories in that cluster are mislabeled and should all be merged into one single shared product category
- BThe embeddings encode the watermark strongly, which can bias retrieval and should be fixed in the data Correct answer
- CThe watermark helps retrieval quality, so more images should get the same watermark added
- DUMAP failed to run correctly, because CLIP embeddings never form clusters in two dimensions
Answer and explanation: A shared visual artifact dominating the embedding means retrieval may match watermarks instead of products, so the data should be cleaned or the artifact removed. CLIP embeddings do cluster, the categories are not mislabeled, and adding watermarks would worsen the shortcut.
Question 10 of 10, Performance Optimization
A team trains a video model whose activations exceed GPU memory even at batch size 1, while weights fit easily. Which technique targets this MOST directly?
- AOptimizer state sharding with ZeRO stage 1, which partitions only the Adam moment tensors
- BActivation checkpointing, optionally with sequence or context parallelism Correct answer
- CWeight quantization to INT4 for both the forward and backward passes
- DRaising the learning rate so the model needs fewer total training steps
Answer and explanation: Activation memory is the problem, so recomputing activations or splitting the long sequence across GPUs targets it directly. ZeRO stage 1 shards optimizer state, which is not the bottleneck, INT4 weight training is not a standard fix for activations, and learning rate does not change per-step memory.