NVIDIA NCA-AMAD exam · Official objectives · Original practice content · Last reviewed October 16, 2026

Free NCA-AMAD Models and Model Types Practice Questions

The Models and Model Types domain is 11% of the NVIDIA NCA-AMAD exam and covers LLM mechanics, small and reasoning models, mixture-of-experts, embedding and reranking models, computer vision tasks and Nemotron. Review each answer and explanation as you practice.

The NCA-AMAD exam has 50-60 multiple-choice questions plus a hands-on lab in 1 hour. It costs $125, results are pass or fail, and the certification is valid for 2 years.

Take the 10-question NCA-AMAD quick-start test · NCA-AMAD cheat sheet · Create a free account

What Models and Model Types covers on the NCA-AMAD exam

This domain is 11% of the NCA-AMAD exam (about 7 questions in a 60-question mock). NVIDIA's objectives include:

How to study Models and Model Types

NCA-AMAD Models and Model Types sample questions with answers

Models and Model Types

Sample question 1

What most distinguishes a deep neural network from a classical machine learning model such as logistic regression?

  1. It learns features from raw data through many stacked layers
  2. It usually beats classical models on small tabular datasets as well
  3. It learns a classification task without needing any labeled examples
  4. It makes predictions from fixed rules rather than learned weights
Show correct answer and explanation

Correct answer: A. It learns features from raw data through many stacked layers

A is correct because deep learning stacks layers of nonlinear transformations that learn features such as edges and shapes directly from raw data, so the features do not have to be engineered by hand first. B is incorrect because classical models such as gradient-boosted trees often match or beat neural networks on small tabular datasets. C is incorrect because neural networks are commonly trained on labeled data; whether a task is supervised is a choice of problem, not what separates them from classical models. D is incorrect because a neural network's behavior comes from weights adjusted during training, not from fixed hand-written rules.

Models and Model Types

Sample question 2

How does a neural network reduce its prediction error during training?

  1. The network adds a hidden layer whenever the training error goes up
  2. A domain expert writes a new rule after each batch to correct the model's outputs
  3. A new decision tree is fitted to the remaining errors after each step
  4. Backpropagation passes the error back through the layers to adjust weights
Show correct answer and explanation

Correct answer: D. Backpropagation passes the error back through the layers to adjust weights

D is correct because with backpropagation inside gradient descent, the error between prediction and target flows back through the layers and each weight is adjusted to reduce it, repeated over many iterations. A is incorrect because the architecture is fixed before training; training changes the weights, not the number of layers. B is incorrect because hand-written rules describe rule-based systems; a neural network learns by updating its weights from data. C is incorrect because fitting trees to residual errors describes gradient boosting, a classical ensemble method, not neural network training.

Models and Model Types

Sample question 3

A retail bank has a table of 5 million labeled card transactions with 40 columns and wants a fraud model. The team has not yet decided which model family to use. What is the BEST first modeling step?

  1. Tune one gradient-boosted tree model exhaustively before trying any other family
  2. Compare fast baselines from linear, boosted-tree and neural network families
  3. Start with a large transformer model, since deep networks tend to win on big tables
  4. Convert each row into an image so that a convolutional network can classify it
Show correct answer and explanation

Correct answer: B. Compare fast baselines from linear, boosted-tree and neural network families

B is correct because diverse baselines across linear models, gradient-boosted trees and neural networks show early which family fits the data best and set a minimum bar to beat, and GPU libraries such as cuML make this fast. A is incorrect because a single baseline says little about the data; heavy tuning of one family before comparing others can waste effort on the wrong path. C is incorrect because no model family reliably wins on tabular data; committing to one large model first skips the comparison that guides later work. D is incorrect because convolutional networks exploit spatial structure in images, and tabular rows have no such structure to exploit.

Models and Model Types

Sample question 4

A plant has 3,000 labeled rows of tabular sensor readings and wants to predict equipment failures. In cross-validation, a gradient-boosted tree model and a deep neural network reach the same score, and the tree model trains in seconds. Which model is the MOST sensible choice?

  1. The neural network, since deep models tend to generalize better to new data
  2. The tree model, since it matches accuracy at far lower training cost here
  3. Neither, since 3,000 rows is too few for any machine learning model
  4. The neural network, since it removes the need to clean the sensor data
Show correct answer and explanation

Correct answer: B. The tree model, since it matches accuracy at far lower training cost here

B is correct because on a small structured dataset, classical models such as gradient-boosted trees are strong baselines, and when accuracy is equal the cheaper model that is faster to retrain is the better choice. A is incorrect because neural networks do not reliably generalize better, and equal validation scores give no reason to prefer the costlier model. C is incorrect because 3,000 labeled rows is enough for classical models, and both models here already validate well. D is incorrect because deep learning can learn features, but it still depends on clean, reliable data; it does not make data cleaning unnecessary.

Models and Model Types

Sample question 5

A manufacturer has 200,000 labeled photos of circuit boards and wants to flag boards with solder defects. Which approach is MOST appropriate?

  1. Fit a linear regression that predicts defects from each image's file size
  2. Cluster the photos with k-means based on each image's average color
  3. Train a convolutional neural network that learns features from the pixels
  4. Fit a logistic regression on raw pixel values flattened into table columns
Show correct answer and explanation

Correct answer: C. Train a convolutional neural network that learns features from the pixels

C is correct because convolutional neural networks learn filters that extract the visual features that matter for the task directly from images, which is why they are the standard model for image recognition and visual inspection. A is incorrect because file size is a weak proxy unrelated to the shape or location of a defect, and regression does not suit a defect label. B is incorrect because k-means is unsupervised and average color discards the fine detail that marks a defect, ignoring the labels the team already has. D is incorrect because a linear model on raw pixels cannot capture shapes and local patterns; classical models need hand-crafted image features to work well.

Models and Model Types

Sample question 6

A churn model built with gradient-boosted trees on a customer table has stopped improving. The table has eight categorical columns, such as plan, region and device type. Which step is MOST likely to uncover new predictive signal?

  1. Train on a smaller random sample so each experiment finishes faster
  2. Create new features that combine pairs of the categorical columns
  3. Replace the tree model with a recurrent neural network over the rows
  4. Drop the categorical columns so the model trains only on numeric columns
Show correct answer and explanation

Correct answer: B. Create new features that combine pairs of the categorical columns

B is correct because combining pairs of categorical columns creates interaction features (28 from eight columns) that capture relationships the original columns do not show, a proven way to add signal to tabular models. A is incorrect because sampling speeds iteration but reduces the data available; it does not add any new signal by itself. C is incorrect because recurrent networks model ordered sequences; customer rows are not a sequence, so the change adds complexity without new signal. D is incorrect because removing the categorical columns discards information rather than revealing new patterns.

Models and Model Types

Sample question 7

On the same tabular dataset, a team's best gradient-boosted tree model and best neural network reach similar validation scores, but they make mistakes on different records. GPU time is available. Which approach is MOST likely to raise accuracy further?

  1. Keep only the neural network, since deep models win on most tabular data
  2. Stack them, training a second-level model on their out-of-fold predictions
  3. Compare the two models' training losses and keep the one with the lower value
  4. Retrain the tree model with more trees until its training error reaches zero
Show correct answer and explanation

Correct answer: B. Stack them, training a second-level model on their out-of-fold predictions

B is correct because when models capture different patterns, stacking learns how best to combine their outputs and often beats any single model; GPU acceleration makes the extra model fits across folds practical. A is incorrect because neither model is better here, and neural networks do not reliably beat boosted trees on tabular data; dropping one discards its complementary strengths. C is incorrect because picking one model by training loss ignores that combining them can correct errors each makes alone. D is incorrect because driving training error to zero tends to overfit and does not use the network's different errors.

Models and Model Types

Sample question 8

A card issuer finds that fraud often shows up only as an unusual order of purchases within a few minutes, not in any single transaction. It has millions of labeled transaction sequences. Which model approach BEST fits this pattern?

  1. A gradient-boosted tree model scored on each transaction's own fields only
  2. A convolutional neural network trained on photos of each customer's payment cards
  3. A recurrent neural network that reads each customer's transactions in order
  4. A k-means model that groups customers by their average monthly spending
Show correct answer and explanation

Correct answer: C. A recurrent neural network that reads each customer's transactions in order

C is correct because recurrent neural networks are built for sequenced data and are good at judging what should come next, so they can flag an unusual series of purchases; fraud detection is a cited use. A is incorrect because boosted trees are strong on tabular data, but scoring each transaction alone drops the ordering where this fraud signal lives. B is incorrect because card images carry no information about the sequence of purchases that reveals this fraud. D is incorrect because unsupervised clustering on monthly averages ignores both the labels and the short-term purchase order.

Models and Model Types

Sample question 9

Which computer vision task locates each object in an image by drawing a bounding box around it?

  1. Image classification
  2. Optical character recognition
  3. Object detection
  4. Semantic segmentation
Show correct answer and explanation

Correct answer: C. Object detection

C is correct because detection localizes where objects are, typically by placing a rectangular bounding box that fully contains each object along with its class. A is incorrect because classification assigns a label to the whole image without saying where the object is. B is incorrect because OCR reads text characters in an image; it does not locate general objects. D is incorrect because segmentation labels individual pixels by category rather than drawing boxes.

Models and Model Types

Sample question 10

What does an optical character recognition (OCR) model produce from a scanned page?

  1. A single label describing what kind of document the page shows
  2. A pixel mask that separates the page from the scanner background
  3. Bounding boxes for the people and vehicles visible in the image
  4. Machine-readable text for the characters it finds in the image
Show correct answer and explanation

Correct answer: D. Machine-readable text for the characters it finds in the image

D is correct because OCR, often paired with character detection that finds the text regions, converts the characters in an image into text that software can search, index and process. A is incorrect because assigning one label to the whole page is image classification. B is incorrect because separating regions at the pixel level is segmentation; it does not read any text. C is incorrect because locating general objects is object detection, not character recognition.

Keep practicing

NCA-AMAD practice test hub · Mixed quick start · NCA-GENL practice test

FlashGenius is an independent study resource and is not affiliated with or endorsed by NVIDIA. NVIDIA and NCA-AMAD are trademarks of NVIDIA Corporation. 600 original practice questions.