NCA-GENM Premium Exam Engine – Download Free PDF Questions [Q127-Q142]

NEW QUESTION 127
Which of the following techniques can be used to reduce the computational cost and memory footprint of large language models (LLMs) during inference?

Quantization

Knowledge Distillation

Increasing the model size

Pruning

Adding more layers

NEW QUESTION 128
Which of the following is the MOST important factor in ensuring the ‘trustworthiness’ of a multimodal Generative AI model used for a safety-critical application (e.g., medical diagnosis)?

High accuracy on the training dataset.

Explainability and interpretability of the model’s decisions.

Low computational cost for inference.

Use of the latest deep learning architecture.

Ability to generate diverse outputs.

NEW QUESTION 129
You are working on a project to classify images of different types of flowers. You have a relatively small dataset (around 500 images per class). Which of the following techniques would be the MOST effective to improve the performance of your image classifier, considering the limited data?

Train a very deep convolutional neural network from scratch-

Apply aggressive data augmentation techniques, such as random rotations, flips, and crops.

Use a pre-trained convolutional neural network on a large dataset like ImageNet and fine-tune it on your flower dataset

Reduce the image resolution to decrease the number of parameters in the model.

Use a simple linear classifier.

NEW QUESTION 130
You are training a text-to-image diffusion model and observe that the generated images often exhibit a ‘washed-out’ or overly smooth appearance. Which of the following adjustments to the training process would likely improve the image quality and detail?

Decrease the number of diffusion steps used during training.

Increase the weight of the perceptual loss function in the training objective.

Reduce the batch size used during training to minimize memory consumption.

Apply more aggressive data augmentation techniques to the training dataset.

Reduce the learning rate for the U-Net architecture within the diffusion model.

NEW QUESTION 131
Which of the following are potential solutions to mitigate the impact of missing or incomplete data in a multimodal dataset used for training a generative A1 model? (Select all that apply)

Data imputation techniques, such as mean imputation or k-nearest neighbors imputation.

Training the model only on complete data samples, discarding any samples with missing data.

Using a masking strategy during training, where the model learns to predict the missing data based on the available data.

Using a specialized multimodal model designed to handle missing data.

All of the above.

NEW QUESTION 132
Consider the following Python code snippet, which attempts to implement a basic form of cross-validation. What is the primary issue with this code and how would you fix it to prevent data leakage?

The code is correct and doesn’t have any data leakage issues.

Data leakage occurs because feature scaling is applied to the entire dataset before splitting it into training and testing sets. Fix: Apply scaling separately to the training and testing sets within each fold.

Data leakage occurs because the ‘KFold” split is not shuffled. Fix: Set in the “KFold’ constructor.

Data leakage occurs because the model is being reinitialized in each fold. Fix: Move the model initialization outside the loop.

Data leakage occurs because the model is not being evaluated on a hold-out set Fix: Create a separate validation set and evaluate the model on it after each fold.

NEW QUESTION 133
Explain the role of Tensor Cores and mixed-precision training (e.g., using FP16 or bfloat16) in accelerating the training of large generative AI models.

Tensor Cores perform specialized matrix multiplications optimized for lower-precision data types, enabling faster computation and reduced memory footprint.

Mixed-precision training allows using lower precision for forward and backward passes but keeps weights and gradients in higher precision to maintain stability.

Tensor Cores are only useful for inference, not training.

Mixed-precision training guarantees the same convergence behavior as full-precision training.

A and B.

NEW QUESTION 134
Consider this PyTorch code snippet related to processing multimodal dat a. What is the primary purpose of the following code in the context of Generative A1?

To create separate data loaders for images and text.

To concatenate image and text data into a single tensor.

To ensure images and text are processed in the same order during training.

To create a custom dataset class for handling paired image and text data.

To resize all images to the same dimension.

NEW QUESTION 135
You are evaluating a multimodal model that generates descriptions for video clips. You have human ratings for the relevance, fluency, and coherence of the generated descriptions. Which statistical test is MOST appropriate for determining if there is a statistically significant difference in the median ratings for each of these criteria (relevance, fluency, coherence) between two different versions of your model?

T-test

ANOVA

Mann-Whitney U test

Kruskal-Wallis test

Friedman Test

NEW QUESTION 136
Consider a scenario where you are evaluating the performance of a multimodal A1 model that generates descriptions for images. However, the generated descriptions tend to be repetitive and lack diversity. Which of the following techniques can be employed to address this issue and encourage more diverse and creative outputs from the model? (Select TWO)

Using a beam search decoding strategy with a small beam width.

Employing a nucleus sampling (top-p sampling) decoding strategy.

Increasing the model’s training data size.

Utilizing a temperature scaling parameter during decoding and increasing its value.

Using a greedy decoding strategy.

NEW QUESTION 137
You are tasked with evaluating a text-to-video generation model. Which of the following metrics would be MOST appropriate for assessing the temporal coherence and smoothness of the generated videos?

Inception Score (IS)

Frchet Inception Distance (FID)

Learned Perceptual Image Patch Similarity (LPIPS)

Frchet Video Distance (FVD)

BLEU score

NEW QUESTION 138
You are tasked with evaluating the trustworthiness of a multimodal A1 model that predicts diagnoses based on medical images and patient history text. Which of the following evaluation metrics or techniques are MOST relevant for assessing the model’s trustworthiness in this critical application?

Accuracy and F1-score on a held-out test set.

Calibration error, measuring the alignment between predicted probabilities and actual outcomes.

Attribution methods (e.g., Grad-CAM) to visualize which parts of the image and text the model focuses on.

Robustness testing by introducing adversarial perturbations to the input data.

Measuring inference throughput (samples per second).

NEW QUESTION 139
You are fine-tuning a pre-trained language model for a specific task. You notice that the model performs well on the training data but poorly on the validation dat a. Which of the following techniques can help mitigate this overfitting problem? (Select TWO)

Increase the learning rate.

Apply weight decay (L2 regularization).

Use dropout regularization.

Decrease the batch size.

Increase the size of the training data.

NEW QUESTION 140
A financial institution aims to detect fraudulent transactions by analyzing transaction history (time-series), customer profiles (text and numerical data), and network activity (graph data). The system must identify fraudulent patterns in real-time. Which of the following architectural patterns is MOST suitable for building this multimodal fraud detection system, considering both accuracy and latency requirements?

A batch processing pipeline that aggregates data daily and trains a machine learning model to identify fraudulent transactions.

A real-time stream processing architecture that uses a combination of rule-based systems and machine learning models to detect fraudulent transactions as they occur.

A hybrid approach that combines real-time stream processing for initial detection and batch processing for periodic model retraining and refinement.

Train three seperate models for each data set, transaction history, customer profile, and network activitiy.

An architecture based solely on rule-based systems to minimize latency, ignoring the potential for machine learning to detect complex fraud patterns.

NEW QUESTION 141
You’re training a Generative Adversarial Network (GAN) to generate images from text descriptions. After a few epochs, you notice the generator is producing nearly identical images regardless of the text input (mode collapse). Which of the following strategies could help mitigate this issue?

Decrease the learning rate of the discriminator.

Increase the capacity (number of layers/neurons) of the discriminator.

Use a larger batch size for the generator.

Implement mini-batch discrimination in the discriminator.

Apply weight decay regularization to the generator

NEW QUESTION 142
You are building a retrieval-augmented generation (RAG) system that utilizes a knowledge graph to enhance the responses generated by a large language model. The knowledge graph contains information about entities and their relationships extracted from both text documents and image metadat a. However, you observe that the system often retrieves irrelevant or outdated information from the knowledge graph, leading to inaccurate or misleading responses. Which of the following strategies would be MOST effective in addressing this issue?

Increase the size of the knowledge graph.

Implement a mechanism to filter and rank the retrieved information based on relevance and recency, using both semantic similarity and temporal information.

Use a simpler language model for the generative component of the RAG pipeline.

Reduce the number of entities in the knowledge graph.

None of the above.