I85PREP
⌘ K
NVIDIA NCA-AIIO study materials
NVIDIA NCA-AIIO study resources

NVIDIA NCA-AIIO practice questions

One question at a time. Build your NVIDIA NCA-AIIO understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Essential AI Knowledge

In supervised learning, what must the training data provide for each example?

  1. A reward signal from an environment
  2. Input features paired with target labels Correct answer.
  3. A simulator that generates episodes
  4. Input features without any labels

Explanation

Supervised learning fits a mapping from inputs to known targets, so every training example needs both features and a label. Unlabeled data is used in unsupervised or self-supervised settings, not standard supervised training.

Question 2Essential AI Knowledge

Why are GPUs well suited to deep learning workloads?

  1. They execute thousands of arithmetic operations in parallel. Correct answer.
  2. They run a single instruction stream faster than any CPU core.
  3. They store the entire dataset in on-chip memory.
  4. They replace the need for floating-point arithmetic.

Explanation

Neural network math is dominated by large matrix and tensor operations that expose massive parallelism. GPUs contain many simple cores designed to execute those independent operations concurrently.

Question 3Essential AI Knowledge

What distinguishes inference from training?

  1. Inference requires labeled data to compute gradients.
  2. Inference runs forward passes with frozen weights. Correct answer.
  3. Inference updates weights using backpropagation.
  4. Inference always runs on CPUs rather than GPUs.

Explanation

Training adjusts parameters by computing gradients and applying an optimizer. Inference uses the learned weights unchanged and only performs forward computation to produce predictions.

Question 4Essential AI Knowledge

In model training, one epoch is defined as which of the following?

  1. One evaluation run on the validation set
  2. One forward and backward pass on a single sample
  3. One update of the model weights
  4. One full pass over the training dataset Correct answer.

Explanation

An epoch counts a complete traversal of the training set, regardless of how many parameter updates occur within it. The number of updates per epoch depends on batch size.

Question 5Essential AI Knowledge

Overfitting is best described as which condition?

  1. Low training error but poor performance on new data Correct answer.
  2. Training error that falls slower than validation error
  3. A model too small to fit the training set
  4. High training error and equally high validation error

Explanation

Overfitting occurs when a model captures noise specific to the training set, so training loss stays low while generalization suffers. The gap between training and held-out performance is the key signal.

Question 6Essential AI Knowledge

Activation functions give a neural network which essential property?

  1. Nonlinearity Correct answer.
  2. Zero gradients
  3. Deterministic outputs
  4. Weight symmetry

Explanation

Stacked linear layers collapse mathematically into a single linear transform. Nonlinear activations between layers let the network represent curved decision boundaries and complex functions.

Question 7Essential AI Knowledge

NVIDIA Tensor Cores are specialized hardware for which operation?

  1. Host-to-device memory copy operations
  2. Scalar integer division with rounding
  3. Disk I/O request scheduling
  4. Mixed-precision matrix multiply Correct answer.

Explanation

Tensor Cores perform matrix multiply-accumulate operations, often mixing lower-precision inputs with higher-precision accumulation. This makes them central to accelerating deep learning training and inference.

Question 8Essential AI Knowledge

A validation set is used during model development to do what?

  1. Store the labels that will be used at inference time
  2. Provide additional gradient updates during training
  3. Tune hyperparameters and select models Correct answer.
  4. Measure final performance after all tuning is complete

Explanation

The validation set provides unbiased feedback for choices such as learning rate, architecture, and early stopping. The test set is reserved for the final estimate after those choices are frozen.

Question 9Essential AI Knowledge

A team raises the minibatch size from 32 to 256 while keeping other settings fixed. What typically happens to the gradient estimate?

  1. Gradients become more noisy across steps
  2. Gradients become less noisy across steps Correct answer.
  3. Backpropagation is no longer required
  4. The loss surface becomes convex

Explanation

A minibatch gradient averages the per-sample gradients, so larger batches reduce variance in the estimate. This often permits a larger learning rate and better hardware utilization.

Question 10Essential AI Knowledge

During training, the loss oscillates and then diverges within a few steps. Which cause is most likely?

  1. Learning rate set too high Correct answer.
  2. Batch size set to one
  3. Too many epochs planned
  4. Learning rate set too low

Explanation

An excessively large learning rate makes each update overshoot the minimum, producing oscillation or divergence. Reducing the learning rate or adding warmup usually stabilizes training.

Question 11Essential AI Knowledge

What does dropout do during the training phase?

  1. Quantizes all activations to 8-bit integers
  2. Randomly zeroes activations each step Correct answer.
  3. Skips whole layers on alternating steps
  4. Permanently prunes the smallest weight values

Explanation

Dropout randomly deactivates a fraction of units on each training step, which prevents units from relying on specific neighbors. At inference, dropout is disabled and outputs are scaled appropriately.

Question 12Essential AI Knowledge

Why do convolutional layers use far fewer parameters than fully connected layers on image inputs?

  1. They process one image at a time.
  2. Filters share weights across spatial positions. Correct answer.
  3. They discard most of the input pixels.
  4. They use lower-precision arithmetic.

Explanation

A convolution reuses the same small filter at every spatial location, so it learns local patterns with a small parameter set. A fully connected layer instead needs a separate weight for every input-output pair.

Question 13Essential AI Knowledge

In a transformer, self-attention builds each token representation from what?

  1. A weighted mix of all tokens in the sequence Correct answer.
  2. Only the immediately preceding token
  3. The final hidden state of the encoder
  4. A fixed window of eight neighboring tokens

Explanation

Self-attention computes attention weights between every pair of tokens, so each position aggregates information from the whole sequence. This is what allows long-range dependencies to be modeled directly.

Question 14Essential AI Knowledge

Converting an FP32 model to INT8 for inference mainly delivers which benefit?

  1. Guaranteed accuracy improvement on all inputs
  2. Lower memory use and higher throughput Correct answer.
  3. Faster training convergence from scratch
  4. Higher numerical precision for each weight

Explanation

INT8 representations are one quarter the size of FP32, reducing memory footprint and enabling faster integer math on supported hardware. Accuracy is usually preserved closely but must be validated.

Question 15Essential AI Knowledge

Fine-tuning a pretrained model is most appropriate when which condition holds?

  1. No pretrained checkpoints exist
  2. The target dataset is huge and unrelated
  3. Random initialization is required
  4. The target dataset is small but related Correct answer.

Explanation

Pretrained features transfer well when the new task resembles the pretraining domain, and small datasets benefit most from that head start. Training from scratch is preferred when data is abundant and domains differ greatly.

Question 16Essential AI Knowledge

Random crops and flips during image training help primarily by doing what?

  1. Removing label noise from the annotations
  2. Speeding up every forward and backward pass
  3. Adding data diversity to reduce overfitting Correct answer.
  4. Making the effective dataset much smaller

Explanation

Augmentation creates plausible variations of each example, so the model sees a richer distribution without collecting new data. This regularizes the model and improves generalization.

Question 17Essential AI Knowledge

Mixed-precision training typically uses which numerical strategy?

  1. Rounds all gradients to integers
  2. Uses FP16 math with FP32 master weights Correct answer.
  3. Uses FP32 math and FP16 optimizer states
  4. Removes the loss scaling requirement

Explanation

Most matrix operations run in FP16 for speed while an FP32 copy of the weights preserves small updates. Loss scaling is used to keep gradients representable in the lower-precision range.

Question 18Essential AI Knowledge

A fraud detector reports 99.5% accuracy on data where only 0.5% of cases are fraud. Which metric best exposes performance on the fraud class?

  1. Overall accuracy
  2. Training loss
  3. Recall for the fraud class Correct answer.
  4. Throughput in samples per second

Explanation

With heavy class imbalance, predicting the majority class alone yields high accuracy. Recall, precision, or F1 for the positive class measures how well the minority fraud cases are actually detected.

Question 19Essential AI Knowledge

CUDA is best described as which of the following?

  1. A curated repository of pretrained model weights
  2. A parallel computing platform for NVIDIA GPUs Correct answer.
  3. An inference serving framework for model containers
  4. A dataset labeling and annotation tool

Explanation

CUDA provides a programming model and runtime for executing parallel code on NVIDIA GPUs. Higher-level libraries and frameworks are commonly built on top of it.

Question 20Essential AI Knowledge

NVIDIA TensorRT is used to accomplish which goal?

  1. Optimize trained models for fast inference Correct answer.
  2. Schedule multi-node training jobs
  3. Author neural network architectures from scratch
  4. Label raw training data automatically

Explanation

TensorRT applies optimizations such as layer fusion, precision calibration, and kernel selection to a trained model. The result is lower latency and higher throughput at deployment time.

Question 21Essential AI Knowledge

Gradient accumulation allows a model to do what?

  1. Increase the available GPU memory bandwidth
  2. Skip the backward pass entirely
  3. Avoid the need for a learning rate
  4. Simulate larger batches on limited memory Correct answer.

Explanation

Accumulation sums gradients over several small microbatches before applying one optimizer step. This approximates the effect of a larger batch when memory cannot hold it.

Question 22Essential AI Knowledge

A three-class classifier outputs softmax probabilities [0.7, 0.2, 0.1] and the true class is index 0. Using natural log, what is the cross-entropy loss?

  1. 1.10
  2. 2.30
  3. 0.36 Correct answer.
  4. 0.10

Explanation

Categorical cross-entropy is the negative log of the probability assigned to the true class. Here -ln(0.7) is approximately 0.357, which rounds to 0.36.

Question 23Essential AI Knowledge

Profiling shows a GPU kernel achieving far below peak FLOPs with very high memory traffic per arithmetic operation. How should this kernel be classified?

  1. Compute-bound
  2. Thermally throttled
  3. Latency-bound by launch overhead
  4. Memory-bandwidth-bound Correct answer.

Explanation

When arithmetic intensity is low, performance is limited by how fast data can be moved rather than by math throughput. Such kernels benefit from fusion, caching, or reduced data movement.

Question 24Essential AI Knowledge

In data-parallel training, what does an all-reduce operation accomplish?

  1. Sums gradients across devices Correct answer.
  2. Compresses activations before the forward pass
  3. Splits the model layers across devices
  4. Broadcasts the input batch to every device

Explanation

Each worker computes gradients on its own data shard, and all-reduce combines those gradients so every worker applies the same update. This keeps model replicas synchronized.

Question 25Essential AI Knowledge

Why do plain recurrent networks struggle to learn long-range dependencies?

  1. Gradients shrink or explode through many time steps Correct answer.
  2. They process entire sequences in parallel
  3. Recurrent weights cannot be updated by backpropagation
  4. They lack an embedding layer

Explanation

Backpropagation through time multiplies many Jacobians, so gradient magnitudes can decay toward zero or grow without bound. Gating mechanisms such as LSTM and GRU cells were designed to mitigate this.

Question 26Essential AI Knowledge

What is the primary role of an optimizer during neural network training?

  1. Store the training examples in GPU memory
  2. Convert the trained model into an inference engine
  3. Adjust model parameters to reduce the loss Correct answer.
  4. Measure final accuracy on a held-out test set

Explanation

An optimizer applies an update rule that uses gradients to change model parameters so the loss tends to decrease. It does not handle inference conversion, data storage, or final test measurement.

Question 27Essential AI Knowledge

In supervised learning, what is a label?

  1. The target output paired with an input example Correct answer.
  2. A numeric feature scaled to a fixed range
  3. A randomly initialized embedding vector
  4. The weight assigned to a single neuron

Explanation

A label is the supervision signal paired with an input; it defines the target the model is trained to predict. Features and weights are inputs and parameters, not targets.

Question 28Essential AI Knowledge

What does one epoch refer to in model training?

  1. One full pass over the training dataset Correct answer.
  2. One parameter update on a single example
  3. One forward pass through the network
  4. One evaluation pass over the test set

Explanation

An epoch counts one complete traversal of the training dataset, regardless of how many parameter updates occur within it. A single update is a step, and an evaluation pass is not an epoch.

Question 29Essential AI Knowledge

Which loss function is standard for multi-class classification with mutually exclusive classes?

  1. Hinge loss applied per class independently
  2. Binary cross-entropy applied once
  3. Categorical cross-entropy after softmax Correct answer.
  4. Mean squared error on raw logits

Explanation

Categorical cross-entropy compares a softmax probability distribution with a one-hot target and is the standard objective for mutually exclusive classes. Binary cross-entropy suits independent binary labels, and hinge losses are used for margin-based classifiers.

Question 30Essential AI Knowledge

What is inference in a machine learning workflow?

  1. Using a trained model to produce predictions Correct answer.
  2. Updating weights from a batch of gradients
  3. Splitting a dataset into training and test parts
  4. Searching hyperparameters with a validation set

Explanation

Inference is the forward-only use of a trained model to score new inputs. It does not modify weights or split data.

Question 31Essential AI Knowledge

What does the learning rate control during gradient-based training?

  1. The size of each parameter update step Correct answer.
  2. The fraction of data used for validation
  3. The number of examples in each class
  4. The number of layers in the model

Explanation

The learning rate scales the gradient step, so it controls how far parameters move per update. It does not set architecture, validation fraction, or class counts.

Question 32Essential AI Knowledge

What is an embedding vector in a machine learning model?

  1. A copy of all model weights saved at one checkpoint
  2. A histogram of pixel values in an image
  3. A dense learned representation of a discrete item Correct answer.
  4. A list of raw text tokens before any cleaning step

Explanation

An embedding maps discrete items such as words or IDs to dense vectors whose values are learned. It is not raw text, a pixel histogram, or a checkpoint.

Question 33Essential AI Knowledge

What does the bias term in a neuron allow it to do?

  1. Reduce the number of input features
  2. Prevent gradients from flowing backward
  3. Shift activation independent of inputs Correct answer.
  4. Normalize each input to zero mean

Explanation

The bias adds a constant offset before the nonlinearity, letting the neuron shift its activation even when inputs are zero. It does not change feature count, normalize inputs, or block gradients.

Question 34Essential AI Knowledge

Backpropagation computes which quantities for training?

  1. The memory required by each GPU kernel
  2. The optimal number of training epochs
  3. Final predictions on the validation dataset
  4. Gradients of the loss with respect to parameters Correct answer.

Explanation

Backpropagation applies the chain rule through the computational graph to obtain gradients of the loss with respect to each parameter. Those gradients are then used by an optimizer.

Question 35Essential AI Knowledge

What is the main effect of L2 weight regularization?

  1. It increases the learning rate over time
  2. It removes a random fraction of neurons
  3. It penalizes large weights and shrinks them Correct answer.
  4. It duplicates examples from the minority class

Explanation

L2 regularization adds a penalty proportional to the squared magnitude of weights, which discourages very large values. It does not drop neurons, change learning rate, or resample data.

Question 36Essential AI Knowledge

During inference, batch normalization typically uses which statistics?

  1. Statistics computed from the current batch only
  2. Statistics recomputed from the test labels
  3. Running estimates accumulated during training Correct answer.
  4. Random values drawn from a normal distribution

Explanation

At inference batch normalization uses running mean and variance estimated during training so outputs do not depend on the current batch. Using batch statistics at inference makes predictions vary with batch composition.

Question 37Essential AI Knowledge

Adam is an optimizer that combines which two ideas?

  1. Dropout regularization combined with early stopping
  2. Weight pruning followed by post-training quantization
  3. Momentum and adaptive per-parameter scaling Correct answer.
  4. Grid search over hyperparameters and random restarts

Explanation

Adam maintains momentum-like first moments and adaptive second moments that scale updates per parameter. It does not perform hyperparameter search, compression, or regularization.

Question 38Essential AI Knowledge

Early stopping ends training when which condition is met?

  1. Validation loss stops improving for several epochs Correct answer.
  2. Training accuracy reaches exactly one hundred percent
  3. The learning rate scheduler reaches its final step
  4. GPU memory usage rises above a configured threshold

Explanation

Early stopping tracks a validation metric and halts when it has not improved for a patience window. Training accuracy reaching 100% is not the criterion, and memory or scheduler thresholds are unrelated.

Question 39Essential AI Knowledge

In a binary classifier, lowering the decision threshold typically has what effect?

  1. Recall increases and precision decreases Correct answer.
  2. Both precision and recall increase
  3. Both precision and recall stay unchanged
  4. Recall decreases and precision increases

Explanation

Lowering the threshold classifies more examples as positive, which usually raises recall but also raises false positives and lowers precision. The two metrics generally move in opposite directions at a fixed model score distribution.

Question 40Essential AI Knowledge

For a categorical feature with thousands of unique values, what does an embedding layer provide?

  1. A sorted list of the most frequent categories
  2. A one-hot vector with thousands of dimensions
  3. A hash table mapping values to raw strings
  4. A dense low-dimensional learnable representation Correct answer.

Explanation

An embedding layer learns a dense vector per category, giving a compact representation with shared statistical strength. One-hot encoding would be very high-dimensional, and frequency lists or string tables are not learned representations.

Question 41Essential AI Knowledge

Which output activation is appropriate for multi-label classification where labels are independent?

  1. ReLU applied to the final logits
  2. Sigmoid applied independently per output Correct answer.
  3. Linear activation with no squashing
  4. Softmax applied across all output units

Explanation

Multi-label targets are independent, so each output needs its own binary probability from a sigmoid. Softmax would force the outputs to compete and sum to one.

Question 42Essential AI Knowledge

In a binary confusion matrix, what is a false positive?

  1. An example predicted positive that is actually negative Correct answer.
  2. An example predicted negative that is actually positive
  3. An example correctly predicted as negative
  4. An example correctly predicted as positive

Explanation

A false positive is a negative example incorrectly flagged as positive. The reverse error is a false negative, and correct predictions are true positives or true negatives.

Question 43Essential AI Knowledge

Which practice introduces data leakage into a model evaluation?

  1. Fitting a scaler on all data before splitting Correct answer.
  2. Evaluating once on a held-out test set
  3. Splitting data before any preprocessing is fit
  4. Using a separate validation set for tuning

Explanation

Leakage occurs when information from the evaluation data influences preprocessing or training. Fitting a scaler on all data before splitting lets test statistics leak into training.

Question 44Essential AI Knowledge

In transfer learning by feature extraction, what is typically done?

  1. Increase the input resolution and retrain from scratch
  2. Retrain all pretrained layers from random weights
  3. Freeze pretrained weights and train a new output head Correct answer.
  4. Delete the pretrained backbone and train a new one

Explanation

Feature extraction keeps the pretrained backbone frozen and trains a new classifier on its features. Retraining all layers or replacing the backbone is closer to full fine-tuning or training from scratch.

Question 45Essential AI Knowledge

A model shows high training accuracy but much lower test accuracy. What does this indicate?

  1. The test set is larger than the training set
  2. The model is overfitting the training data Correct answer.
  3. The learning rate was set too low
  4. The model is underfitting the training data

Explanation

A large gap between training and test performance is the classic sign of overfitting, where the model memorizes training-specific patterns. Underfitting would show poor training performance as well.

Question 46Essential AI Knowledge

Residual connections primarily help deep networks by doing what?

  1. Removing the need for any activation functions
  2. Reducing the number of parameters in each layer
  3. Providing a shortcut path for gradient flow Correct answer.
  4. Guaranteeing that the loss is convex

Explanation

Residual connections add identity shortcuts so gradients can flow directly across layers, easing optimization of very deep networks. They do not reduce parameter count or remove activations.

Question 47Essential AI Knowledge

When training on imbalanced data, what does assigning higher class weights to the rare class do?

  1. Guarantees balanced precision and recall
  2. Raises the loss weight on rare examples Correct answer.
  3. Makes the model ignore the majority class
  4. Removes rare examples from the loss

Explanation

Class weights multiply the loss for each class, so rare examples contribute more and the model pays more attention to them. They do not delete examples or guarantee a particular precision-recall balance.

Question 48Essential AI Knowledge

What does increasing the temperature of a softmax distribution do?

  1. It makes the distribution more peaked
  2. It forces every probability to be exactly identical
  3. It flattens the distribution toward uniform Correct answer.
  4. It changes the number of output classes

Explanation

Higher softmax temperature divides logits by a larger number, producing a flatter distribution closer to uniform. Lower temperature sharpens the distribution; temperature does not change class count.

Question 49Essential AI Knowledge

Why is layer normalization often preferred over batch normalization in transformers?

  1. It removes the need for residual connections
  2. It always increases model accuracy on test data
  3. It normalizes per example, not per batch Correct answer.
  4. It requires larger batches to estimate statistics

Explanation

Layer normalization computes statistics across features within each example, so it works with small or variable batch sizes common in transformer training. Batch normalization relies on batch statistics and can be unstable when batches are small.

Question 50Essential AI Knowledge

What problem does gradient clipping address during training?

  1. Vanishing gradients that shrink to exactly zero
  2. Slow data loading from disk to GPU
  3. Exploding gradients with very large updates Correct answer.
  4. Overfitting caused by too many parameters

Explanation

Gradient clipping rescales gradients whose norm exceeds a threshold, limiting the damage from exploding gradients. It does not restore vanished gradients, reduce parameter count, or speed data loading.