NVIDIA NCA-GENM Practice Questions with Explanations
Free NVIDIA NCA-GENM practice questions. 14 of them, each with the correct answer, a full explanation, and the reason every other option is wrong.
NCA-GENM practice questions
Question 1
You are tasked with developing an image processing model using machine learning. You need to classify thousands of labeled images of cats and dogs.
Which algorithm is commonly used for image classification?
- A. Decision Trees
- B. K-Means Clustering
- C. Convolutional Neural Networks (CNN)
- D. Linear Regression
Show answer and explanation ▾
Correct answer: C
Convolutional neural networks learn visual features such as edges and shapes from labeled images, which makes them the usual choice for classifying pictures such as cats and dogs.
Why the other options are wrong:
- A. Decision trees can classify tabular data but are not the common choice for raw image pixels.
- B. K-Means is an unsupervised clustering method and does not use labels.
- D. Linear regression predicts a continuous value rather than a class.
Question 2
What is the correct order of steps in an ML project?
- A. Data preprocessing, Data collection, Model training, Model evaluation
- B. Data collection, Data preprocessing, Model training, Model evaluation
- C. Model evaluation, Data preprocessing, Model training, Data collection
- D. Model evaluation, Data collection, Data preprocessing, Model training
Show answer and explanation ▾
Correct answer: B
A project starts by collecting data, then preprocessing it, then training a model, and finally evaluating it.
Why the other options are wrong:
- A. Preprocessing cannot come before the data has been collected.
- C. Evaluation cannot come before training, and training needs the data first.
- D. Starting with evaluation makes no sense, since there is no model or data yet.
Question 3
In a Generative Adversarial Network (GAN), what is the role of the discriminator?
- A. To generate new data based on the training set.
- B. To distinguish between real and generated data.
- C. To optimize the training process.
- D. To calculate the loss function and update the generator.
Show answer and explanation ▾
Correct answer: B
In a GAN the discriminator is a classifier that judges whether a sample is real or produced by the generator, and its feedback trains the generator.
Why the other options are wrong:
- A. Generating new data is the generator's role.
- C. Optimizing the training process is not the discriminator's role.
- D. Loss calculation and generator updates belong to the training procedure, not the discriminator itself.
Question 4
What is the significance of using a U-Net like architecture in denoising diffusion probabilistic models?
- A. To classify input images as noisy or clean.
- B. To generate new images from pure noise.
- C. To segment noisy patches in input images.
- D. To detect noisy objects in input images.
Show answer and explanation ▾
Correct answer: B
A U-Net is used in a denoising diffusion model to predict and remove noise step by step, which is how new images are generated starting from pure noise.
Why the other options are wrong:
- A. The model does not classify inputs as noisy or clean.
- C. Segmenting noisy patches is a segmentation task, not the role in diffusion.
- D. Detecting noisy objects is an object detection task, not the role in diffusion.
Question 5
In ML applications, which machine learning algorithm is commonly used for creating new data based on existing data?
- A. Decision tree
- B. Support vector machine (SVM)
- C. K-means clustering
- D. Generative adversarial network (GAN)
Show answer and explanation ▾
Correct answer: D
A generative adversarial network learns the pattern of existing data and creates new samples that look like it.
Why the other options are wrong:
- A. A decision tree is a predictive model for classification or regression, not a data generator.
- B. An SVM separates classes and does not create new data.
- C. K-means groups existing data and does not create new data.
Question 6
In large-language models, what is the purpose of the attention mechanism?
- A. To measure the importance of the words in the output sequence.
- B. To assign weights to each word in the input sequence.
- C. To determine the order in which words are generated.
- D. To capture the order of the words in the input sequence.
Show answer and explanation ▾
Correct answer: B
Self-attention gives a weight to each part of the input while the model processes it. The weight shows how important that input token is in the context of the rest of the input, so the model can focus on the parts that matter.
Why the other options are wrong:
- A. Attention weights the tokens of the input sequence, not a ranking of words in the output.
- C. Attention does not decide the order in which tokens are generated.
- D. Word order is captured by positional encoding, not by the attention weights themselves.
Question 7
What is the purpose of a kernel in a Convolutional Neural Network (CNN)?
- A. To normalize the input data.
- B. To classify the data into different categories.
- C. To perform convolution operations on input data.
- D. To calculate the loss function.
Show answer and explanation ▾
Correct answer: C
In a CNN, the convolution operation combines the input data or feature map with a convolution kernel, also called a filter, to produce a transformed feature map. The filter values are learned during training so the layer extracts useful features.
Why the other options are wrong:
- A. Normalizing data is the job of preprocessing or normalization layers, not of a kernel.
- B. Classification is done by the final fully connected layer, not by a convolution kernel.
- D. The loss function is computed from the model output and the labels, not by a kernel.
Question 8
What characteristic of autoencoders makes them suitable for anomaly detection?
- A. Their capacity to learn a compressed representation of the data.
- B. Their ability to classify images with high accuracy.
- C. Their capability to predict future outcomes based on past data.
- D. Their function in enhancing the quality of images.
Show answer and explanation ▾
Correct answer: A
An autoencoder's encoder compresses data into a lower-dimensional latent representation and the decoder rebuilds the input from it. Normal data is rebuilt with lower reconstruction error than anomalies, so a high error flags an anomaly.
Why the other options are wrong:
- B. Autoencoders are not image classifiers; anomaly detection relies on reconstruction error, not classification accuracy.
- C. Forecasting future outcomes from past data is not the property that makes autoencoders useful for anomaly detection.
- D. Improving image quality is not the reason autoencoders can detect anomalies.
Question 9
How is CTC loss utilized to optimize Large Language Models (LLMs) for sequence modeling tasks?
- A. CTC loss is employed to enhance the translation accuracy of LLMs between different languages.
- B. It is mainly used to align the lengths of the input and output sequences in models performing tasks like speech recognition.
- C. CTC loss is utilized to reduce the vocabulary size of LLMs to make training more efficient.
- D. CTC loss is applied in LLMs to improve sentence generation by ensuring grammatical correctness.
Show answer and explanation ▾
Correct answer: B
CTC loss is used to train speech recognition models, where there are many more audio frames than output units. It allows every possible alignment between frames and text and uses a blank unit for frames with no output, so no per-frame timing labels are needed.
Why the other options are wrong:
- A. CTC is not a technique for improving translation between languages.
- C. CTC does not reduce the vocabulary size of a model.
- D. CTC does not check or enforce grammatical correctness of generated sentences.
Question 10
Which NLP task involves token classification?
- A. Sentinment analysis
- B. Tokenization
- C. Text classification
- D. Part-of-speech tagging
Show answer and explanation ▾
Correct answer: D
Token classification assigns a label to each token in the text. Part-of-speech tagging is a token level classification task, like named entity recognition, because every word gets its own grammatical tag.
Why the other options are wrong:
- A. Sentiment analysis labels a whole sentence or document, not each token.
- B. Tokenization splits text into tokens; it does not classify them.
- C. Text classification assigns one label to an entire text rather than to each token.
Question 11
Which technique involves leveraging pre-trained models to achieve efficient results with less data and computation?
- A. Neural network integration
- B. Prompt engineering
- C. State management and composition
- D. Transfer learning
Show answer and explanation ▾
Correct answer: D
Transfer learning reuses features learned by a pretrained model, retraining it for a new task instead of starting from scratch. It needs less data to train an accurate model, which also cuts overall training time and cost.
Why the other options are wrong:
- A. Neural network integration is not the technique of retraining a pretrained model for a new task.
- B. Prompt engineering shapes the input to a model and does not retrain it on new data.
- C. State management and composition is not a model training technique.
Question 12
What could be a way to handle overfitting in deep learning?
- A. Increase the complexity of the model to fit the training dataset better.
- B. Add regularization techniques such as dropout or L1/L2 regularization.
- C. Increase the training time to fit the training dataset better.
- D. Decrease the size of the training dataset.
Show answer and explanation ▾
Correct answer: B
Adding regularization such as dropout is a standard remedy for overfitting. Other remedies include using more and more varied training data and training for fewer epochs, which helps the model generalize to new inputs.
Why the other options are wrong:
- A. Fitting the training data even more closely is the cause of overfitting, not a remedy.
- C. Training longer increases overfitting; the guidance is to reduce training epochs.
- D. A smaller training set makes overfitting more likely; the guidance is to increase dataset size.
Question 13
Which of the following best describes hyperparameter tuning in ML experimentation?
- A. Hyperparameter tuning refers to the process of generating new training data to improve model accuracy.
- B. Hyperparameter tuning involves finding the optimal values for the hyperparameters of a machine learning model.
- C. Hyperparameter tuning refers to the process of adjusting the weights of a neural network to optimize model performance.
- D. Hyperparameter tuning is the process of selecting the best model architecture for a given task.
Show answer and explanation ▾
Correct answer: B
Hyperparameter tuning searches for the best values of the model settings that are chosen for training, such as the depth of a decision tree or the learning rate, in order to find the most accurate model for the problem.
Why the other options are wrong:
- A. Generating new training data is data collection or augmentation, not hyperparameter tuning.
- C. Network weights are learned during training; hyperparameters are the settings chosen for training.
- D. Choosing a model architecture is model selection, not tuning the hyperparameter values of a model.
Question 14
You need to evaluate the realism of a generative AI model.
Which of the following metrics can be used to measure how realistic the output images of a generative AI model are?
- A. Accuracy
- B. F1 Score
- C. FID
- D. MSE
Show answer and explanation ▾
Correct answer: C
Frechet Inception Distance (FID) evaluates the quality of generated images. It measures the distance between the distribution of real images and the distribution of generated images, using features extracted by a pretrained Inception model.
Why the other options are wrong:
- A. Accuracy measures correct predictions in classification and does not compare generated images with real ones.
- B. F1 score balances precision and recall for classifiers and does not measure image realism.
- D. Mean squared error compares values one by one and does not measure how realistic a set of generated images is.
That was 14 of 60.
The full NVIDIA NCA-GENM pack has all 60 questions, each with the answer, the explanation and why the other options are wrong, plus a questions-only copy for timed runs. $39.00 USD, paid once, with free monthly updates and a pass-or-your-money-back guarantee.
Get the full packNot ready yet? Get the first 14 free as one PDF by email. Get the Half Pack →
About the NCA-GENM exam
These 14 NVIDIA NCA-GENM questions are free here, and as a free PDF by email.
They are drawn from the same bank as the full NCA-GENM pack, which has 60 questions in total.