NCA-GENL Core Machine Learning and AI Knowledge — 20 Questions

Question 1

Which TWO of the following statements regarding Gradient Descent variants are correct?

Accepted Answer

Batch Gradient Descent calculates the gradient using the entire dataset, ensuring a stable path to the global minimum.. Understanding optimization algorithms is vital for efficient LLM training on NVIDIA hardware. Stochastic Gradient Descent (SGD) introduces noise that can help escape local minima, while Adam combines adaptive learning rates with momentum. Selecting the appropriate optimizer significantly impacts convergence speed and the final quality of the model, especially when dealing with high-dimensional loss landscapes common in modern deep learning architectures.

Answer

Stochastic Gradient Descent always converges faster than Adam in all neural network architectures.

Answer

Momentum-based optimization has no effect on the speed of convergence in deep learning.

Answer

Mini-batch SGD is inherently less efficient than using the full dataset for every update.

Question 2

Which of the following activation functions is most commonly used in modern transformer-based LLMs to avoid the vanishing gradient problem?

Accepted Answer

GeLU. Choosing the right activation function is foundational for deep learning success. In deep architectures like transformers, standard sigmoid or tanh functions often suffer from saturation, leading to vanishing gradients. Modern variants like GeLU (Gaussian Error Linear Unit) provide a non-linear, smooth gradient that promotes better convergence and training stability, making them the industry standard for LLMs and deep neural networks on NVIDIA hardware.

Answer

Sigmoid

Answer

ReLU

Answer

Softmax

Question 3

Which THREE of the following are primary benefits of using Mixed Precision Training on NVIDIA GPUs?

Accepted Answer

Reduced memory usage for model weights and activations.. Mixed precision training is an optimization technique that utilizes both FP16 and FP32 to accelerate training. It is critical for maximizing the utilization of NVIDIA Tensor Cores. By reducing the memory footprint, it allows for larger models and batch sizes while maintaining numerical stability through techniques like loss scaling, thus significantly decreasing overall training time without sacrificing model accuracy or convergence quality.

Answer

Improved model accuracy by eliminating all quantization noise.

Answer

Automatic conversion of all operations to 8-bit integers.

Question 4

Which TWO of the following are true regarding the transformer's Multi-Head Attention (MHA) mechanism?

Accepted Answer

It allows the model to jointly attend to information from different representation subspaces.. Multi-head attention allows the model to simultaneously focus on different parts of the input sequence from different representation subspaces. By running multiple attention heads in parallel, the transformer can capture diverse relationships—such as syntactic, semantic, or long-range dependencies—that a single head might miss. This architectural design is a cornerstone of current LLM performance, providing the necessary richness to understand complex, context-heavy language input.

Answer

It replaces the feed-forward network entirely in the transformer block.

Answer

It ensures that the model can only attend to the immediate previous token.

Answer

It is computationally slower than an RNN for sequential tasks.

Question 5

A machine learning engineer is training a large transformer model on a multi-GPU cluster. The training process periodically hangs due to gradient synchronization delays during backpropagation. Which technique should be implemented to mitigate this bottleneck?

Accepted Answer

Implement gradient accumulation to reduce synchronization frequency.. Gradient synchronization often causes bottlenecks in distributed training due to high communication overhead between nodes. Gradient accumulation effectively increases the effective batch size while reducing the frequency of synchronization steps. This approach optimizes bandwidth usage and stabilizes the training process by allowing the model to perform multiple forward and backward passes before updating weights, directly addressing the synchronization latency inherent in large-scale distributed GPU training workflows.

Answer

Increase the learning rate linearly with the number of GPUs.

Answer

Transition to synchronous Stochastic Gradient Descent with all-reduce.

Answer

Switch from FP16 to FP32 precision to improve numerical stability.

Question 6

Refer to the exhibit. An engineer is deploying a PyTorch DistributedDataParallel job on an NVIDIA DGX cluster. Based on the logs, what is the root cause of the training failure?

Accepted Answer

NCCL backend initialization failed due to node rank mismatch.. The logs indicate a mismatch between the expected infrastructure configuration and the actual resources discovered by the NCCL backend. NCCL is highly sensitive to network topology and rank assignment in multi-node clusters. The error confirms that the distribution logic expected a specific multi-GPU parity that the current node allocation failed to provide, resulting in a system error when attempting to initialize the collective communication primitives.

Answer

The model weights are corrupted in the checkpoint file.

Answer

Insufficient VRAM on Rank 0 causing OOM killer.

Answer

Incompatible CUDA toolkit version across the cluster.

Question 7

When designing a high-performance inference pipeline using NVIDIA TensorRT, which TWO practices are essential for maximizing throughput and reducing latency on A100 GPUs?

Accepted Answer

Utilize layer fusion to combine redundant operations.. Optimizing TensorRT pipelines requires leveraging hardware-specific features like kernel autotuning and precision calibration. By performing layer fusion and utilizing Tensor Cores via FP16 or INT8 quantization, the engine minimizes memory bandwidth bottlenecks. These practices ensure the model executes efficiently by reducing the number of memory operations and maximizing the computational density provided by the NVIDIA Ampere architecture, which is critical for real-time production inference.

Answer

Enable dynamic shape support for all input layers.

Answer

Always prioritize FP32 precision to maintain maximum accuracy.

Answer

Bypass the build phase and use direct PyTorch execution.

Question 8

When preparing a dataset for training large-scale foundation models, which THREE data-cleaning steps are recommended to ensure high model quality?

Accepted Answer

Global deduplication of training sequences.. Data quality is the most significant factor in foundation model performance. Filtering low-quality text, deduplicating training samples, and enforcing safety guidelines are standard industry practices. These steps prevent the model from learning noise or harmful patterns, which is critical when training on massive, uncurated web-scale datasets. Maintaining high data integrity is essential for avoiding catastrophic forgetting and ensuring consistent output quality across diverse tasks in the LLM.

Answer

Upsampling all datasets to have an equal number of samples.

Answer

Manually labeling all 10 trillion tokens in the corpus.

Question 9

Which THREE of the following are primary benefits of utilizing NVIDIA NeMo for LLM development?

Accepted Answer

Native support for large-scale distributed training.. NVIDIA NeMo is an integrated toolkit for training, fine-tuning, and deploying large models efficiently. By providing built-in support for distributed training (FSDP/Megatron), hardware-aware optimizations, and data processing pipelines, it simplifies the development cycle significantly. Its modular design allows developers to scale from small experiments to multi-node clusters seamlessly, making it a critical asset for researchers and engineers aiming to move from research prototypes to production-grade foundation models on NVIDIA hardware.

Answer

Automatic translation of model code into C++.

Answer

Direct access to proprietary model weights.

Question 10

Which TWO of the following statements correctly describe the role of gradient descent in training modern deep learning models?

Accepted Answer

Stochastic Gradient Descent (SGD) introduces noise that can help escape local minima.. Gradient descent is the foundational optimization algorithm used to minimize the loss function by iteratively updating model parameters. It relies on calculating partial derivatives with respect to weights. In modern architectures, variations like Adam are preferred because they adapt learning rates per parameter, helping navigate complex, high-dimensional loss landscapes more effectively than standard batch gradient descent, ensuring faster convergence and better performance on sparse data.

Answer

Batch gradient descent is highly efficient for datasets with billions of parameters.

Answer

Gradient descent is only applicable to convex loss functions.

Answer

Learning rate is not a hyperparameter for gradient descent.

Question 11

Which term describes the process of fine-tuning a model on a smaller, task-specific dataset after it has been pre-trained on a large, general-purpose dataset?

Accepted Answer

Transfer Learning. Transfer Learning is a powerful strategy that leverages the features learned by a model on massive datasets. By using these learned weights as a starting point, the model converges faster and requires significantly less data to perform well on a new, specialized task. This approach is standard practice in generative AI, enabling high-quality results in domains where labeled data might be scarce, by building upon a foundation of general knowledge.

Answer

Active Learning

Answer

Feature Engineering

Answer

Reinforcement Learning

Question 12

A machine learning engineer is training a large language model and observes the training loss plateauing prematurely. Which technique is most appropriate to address this issue while maintaining computational efficiency?

Accepted Answer

Implementing a learning rate scheduler with warm-up. Premature plateauing often indicates a sub-optimal learning rate or vanishing gradients. Using a learning rate scheduler, such as cosine annealing with warm-up, allows the model to traverse the loss landscape more effectively. This ensures the model avoids local minima early in the training process, leading to better convergence. Proper optimization strategies are critical in large-scale AI projects where computational costs of re-training models are significant and hardware resources must be maximized.

Answer

Increasing the batch size infinitely

Answer

Decreasing the model architecture size

Answer

Switching from FP16 to FP64 precision

Question 13

Which TWO of the following are potential issues when training deep learning models on highly imbalanced datasets?

Accepted Answer

The model may become biased toward the majority class. Imbalanced datasets cause models to become biased toward the majority class, as the loss function is dominated by these frequent examples. This leads to poor performance on the minority class, which is often the most critical for business applications (e.g., fraud detection). Recognizing these issues allows engineers to apply techniques like oversampling or class weighting, ensuring the model provides utility across the entire data spectrum, not just the majority.

Answer

The model will always exhibit higher variance

Answer

The training will always lead to gradient explosion

Answer

The model will fail to utilize GPU acceleration

Question 14

A data scientist is building a text classifier and wants to evaluate how well the trained model will perform on completely unseen data. They have a dataset of 10,000 labeled examples. Which evaluation approach provides the most reliable estimate of the model's generalization performance?

Accepted Answer

Split the data into training and test sets, train on the training set, and evaluate on the test set.. The most reliable estimate of generalization comes from evaluating on data the model has never seen during training or hyperparameter tuning. A train-test split ensures the test set is independent, so its accuracy reflects real-world performance. Other methods either reuse training data or involve tuning, which can bias the estimate.

Answer

Evaluate the model on a separate validation set that was used for hyperparameter tuning.

Answer

Train the model on all 10,000 examples and report the training accuracy.

Answer

Use k-fold cross-validation on the entire dataset and report the average validation accuracy.

Question 15

A research team is pretraining a transformer-based language model on a large corpus. They want to ensure the model learns bidirectional context, unlike traditional left-to-right models. Which TWO pretraining objectives are designed to achieve this? (Choose two.)

Accepted Answer

Permutation Language Modeling (PLM). Masked Language Modeling and Permutation Language Modeling are explicitly designed to capture bidirectional context. MLM masks tokens and predicts them using both sides, while PLM permutes the factorization order to allow bidirectional dependencies. Causal LM is unidirectional, and NSP and span corruption serve different purposes, so they do not fulfill the bidirectional requirement.

Answer

Span Corruption

Answer

Causal Language Modeling (CLM)

Answer

Next Sentence Prediction (NSP)

Question 16

A research team is pre-training a large language model on a massive corpus of text. They want to ensure that the model learns useful representations and generalizes well to downstream tasks. Which TWO of the following techniques are commonly used during pre-training to improve the model's ability to capture contextual relationships and avoid overfitting? (Choose two.)

Accepted Answer

Masked language modeling (MLM). Masked language modeling and causal language modeling are two fundamental pre-training objectives for large language models. MLM enables bidirectional context learning, while CLM enables autoregressive generation. Both are essential for learning rich representations that generalize to downstream tasks, and they inherently prevent overfitting by exposing the model to diverse contexts.

Answer

Batch normalization

Answer

Weight initialization with small random values

Answer

Dropout on input embeddings

Question 17

A researcher is training a transformer-based language model and wants to prevent the model from attending to future tokens during training. Which mechanism should be implemented in the self-attention layer?

Accepted Answer

Add a causal mask to the attention scores before the softmax operation.. To prevent attending to future tokens, a causal mask must be applied to the attention scores before softmax. This sets future positions to negative infinity, resulting in zero attention probability. Other techniques like layer normalization, dropout, or more heads do not enforce causality and are unrelated to masking future tokens.

Answer

Use a higher dropout rate on the attention weights.

Answer

Apply layer normalization before the self-attention computation.

Answer

Increase the number of attention heads.

Question 18

A team is deploying a large language model for real-time inference on NVIDIA GPUs. They want to reduce latency and increase throughput. Which TWO techniques are most effective for achieving these goals? (Choose two.)

Accepted Answer

Using NVIDIA TensorRT to optimize the model graph.. Quantization and TensorRT optimization are both highly effective at reducing latency and increasing throughput for LLM inference on NVIDIA GPUs. Quantization lowers precision to speed up computation and reduce memory bandwidth, while TensorRT optimizes the model graph for the specific GPU. Together, they provide substantial performance gains without altering the model's core architecture, making them the most suitable choices for real-time deployment.

Answer

Increasing the batch size beyond the GPU's memory capacity.

Answer

Enabling dynamic batching to group incoming requests.

Answer

Implementing a larger number of transformer layers.

Question 19

A team is deploying a large language model for real-time chatbot inference on NVIDIA GPUs. They need to reduce latency without sacrificing output quality. Which two techniques should they implement? (Choose two.)

Accepted Answer

Apply kernel fusion to combine multiple operations into a single GPU kernel.. Quantization to INT8 reduces memory bandwidth and leverages faster integer tensor core operations, while kernel fusion minimizes overhead and memory traffic. Both techniques are widely used in NVIDIA inference stacks to lower latency without retraining or degrading output quality. They directly address the real-time constraint by making each forward pass faster and more efficient.

Answer

Increase the batch size for each inference request.

Answer

Switch from FP16 to FP32 for all computations.

Answer

Increase the number of attention heads in each transformer layer.

Question 20

A researcher is training a large language model and notices the training loss plateaus early while validation loss increases. What is the most likely cause, and which action should be taken?

Accepted Answer

The model is overfitting; apply dropout or weight decay.. The symptoms described clearly indicate overfitting, where the model captures noise in the training set rather than generalizing to unseen data. In the context of large language models, this is a critical challenge. Implementing regularization techniques such as weight decay or dropout helps constrain model complexity, forcing it to learn more robust features rather than memorizing specific patterns, thereby improving overall model performance and generalizability.

Answer

The model is underfitting; increase the number of hidden layers.

Answer

The learning rate is too low; increase it to accelerate convergence.

Answer

The dataset is too small; reduce the batch size.