CCNA Aio Ai Concepts Techniques Questions

Perplexity measures language model confidence, but BLEU measures translation quality. The deeper model may overfit to the training data, reducing perplexity but hurting generalization to validation translations. Overfitting causes high confidence (low perplexity) but poor translation diversity or exact matches.

Providing a few-shot example with concise summaries (Option A) directly demonstrates the desired output format to the model, leveraging in-context learning to bias generation toward brevity and relevance. This is the most effective technique for controlling output style without altering the model's underlying parameters.

Diffusion models are the best choice because they iteratively denoise random noise to produce high-quality, photorealistic images with diverse styles. Unlike GANs, they avoid mode collapse and training instability, and they generate more detailed and varied outputs than VAEs, making them ideal for furniture catalog images in different room settings.

Reinforcement learning learns via trial-and-error using rewards and penalties, ideal for game-playing agents. Supervised learning requires labeled data; unsupervised learning finds patterns without rewards; semi-supervised uses a mix.

Recommender systems are a classic application of supervised learning (if using regression or classification to predict ratings) or unsupervised learning (collaborative filtering). Among the options, supervised learning with regression is appropriate for predicting purchase likelihood.

Chain-of-thought prompting encourages the model to show intermediate reasoning steps, improving accuracy on multi-step problems. Zero-shot gives no examples; few-shot provides examples but not necessarily step-by-step; temperature controls randomness.

Input sanitization removes special characters or patterns; output filtering checks responses for sensitive data; system prompts with separation instructions can reduce injection risk. Restricting temperature only affects randomness; using a smaller model does not prevent injection.

Decision trees inherently provide interpretable decision rules by splitting data based on feature thresholds at each node. The entire model can be visualized as a tree structure, allowing compliance teams to trace the exact path and logic behind each loan approval or rejection, which directly satisfies explainability requirements.

Option C is correct because the company has no pre-labeled data, which means supervised learning (which requires labeled examples) is not feasible. Unsupervised clustering, such as K-means, groups data points into clusters based on feature similarity without needing any labels, making it ideal for automatically discovering categories like billing, technical, or account from raw ticket text.

Clustering partitions data into groups where intra-cluster similarity is high. Classification is supervised; anomaly detection finds outliers; regression predicts continuous values.

Option B is correct because fine-tuning on a curated domain-specific corpus directly aligns the model with the factual patterns and terminology of the target domain. This supervised learning process adjusts the model's weights to reduce the probability of generating incorrect or hallucinated content by reinforcing ground-truth examples from the domain.

A large gap between training and validation loss indicates overfitting. Increasing dropout (a regularisation technique) reduces overfitting by preventing co-adaptation of neurons. Increasing LSTM units or epochs would worsen overfitting, and reducing sequence length may lose important temporal patterns.

L2 regularisation (Ridge) adds a penalty term equal to the sum of the squared coefficients to the loss function, which shrinks coefficient magnitudes without forcing them to zero. This reduces variance and overfitting by making the model less sensitive to individual features, which is ideal when the model performs well on training data but poorly on test data due to high variance.

Precision (A) is correct because it measures the proportion of true positive predictions among all positive predictions, which is critical in imbalanced medical diagnosis where false positives can lead to unnecessary stress or procedures. In a dataset with only 5% positive cases, a model that predicts all negatives would achieve high accuracy but zero precision, so precision helps assess the cost of false alarms.

Recall (true positive rate) minimises false negatives, which is critical when missing a positive case is dangerous.

Accuracy is high but precision and recall are notably lower, indicating class imbalance where the model biases toward the majority class, inflating accuracy.

The training loss decreasing slowly and accuracy improving marginally each epoch indicates that the learning rate is too small, causing the optimizer to take very small steps toward the minimum of the loss function. Increasing the learning rate allows the optimizer to take larger steps per update, which accelerates convergence. Option C is correct because adjusting the learning rate directly addresses the step size in gradient descent.

A learning rate of 0.001 is very low for many logistic regression implementations, causing the gradient descent algorithm to take extremely small steps toward the minimum of the loss function. This results in a slow decrease in loss because each weight update is minimal. Increasing the learning rate to 0.01 allows larger steps per iteration, accelerating convergence without typically causing divergence in well-scaled data.

AUC-ROC is the most appropriate metric because it evaluates the model's ability to distinguish between the minority fraud class (0.1%) and the majority legitimate class across all classification thresholds, without being biased by the extreme class imbalance. Unlike accuracy, AUC-ROC remains robust when the dataset is 99.9% legitimate, as it measures the true positive rate against the false positive rate, providing a comprehensive view of model performance for rare event detection.

Decision trees are inherently interpretable, showing the decision rules. Random forests and gradient boosting are ensembles that sacrifice interpretability for accuracy. Neural networks are black-box models.

WordPiece tokenisation is the component of BERT that handles out-of-vocabulary (OOV) words by breaking them into subword units (e.g., 'playing' → 'play' + '##ing'). This allows the model to represent any word, even unseen ones, as a sequence of known subword tokens, ensuring no word is truly out of vocabulary.

Random forests provide feature importance, handle mixed data types, and require little preprocessing.

Fine-tuning a pre-trained BERT model is most appropriate because BERT is a transformer-based model pre-trained on a large corpus and can be fine-tuned on the 50,000 labeled reviews to achieve high accuracy with relatively little data. It captures bidirectional context, which is crucial for sentiment classification, and avoids the need for training from scratch.

Supervised learning uses labeled data to train a model to map inputs to outputs. Fraud detection with historical labels is a classic binary classification problem.

Variational Autoencoders (VAEs) are known for producing blurry outputs because their loss function (ELBO) encourages pixel-wise averaging, which smooths out fine details. Diffusion models, by contrast, iteratively denoise a random field, learning to reconstruct high-frequency details through a multi-step reverse process, directly addressing the blurriness issue.

Retrieval-Augmented Generation (RAG) is the most appropriate approach because it allows the chatbot to answer questions based on the latest policy documents without retraining the model. By indexing the documents in a vector store and retrieving relevant chunks at query time, RAG provides up-to-date, contextually accurate answers while keeping the underlying LLM static, which avoids the cost and complexity of monthly retraining.

Reinforcement learning (RL) is the correct paradigm because it explicitly involves an agent learning a policy through trial-and-error interactions with an environment, receiving scalar reward signals (positive or negative) to maximize cumulative reward. This matches the question's description of making decisions based on rewards or penalties, which is the defining characteristic of RL, as opposed to learning from labeled data or discovering hidden patterns without feedback.

Retrieval-Augmented Generation (RAG) is the most appropriate approach because it allows the chatbot to answer questions by retrieving relevant chunks from the policy documents stored in a vector store, without requiring model retraining when documents are updated monthly. The retrieval component dynamically fetches the latest content, while the generation component uses a pre-trained LLM to produce answers, making it cost-effective and scalable for frequently changing knowledge bases.

RAG grounds answers in retrieved evidence, fine-tuning can align with medical domain, and prompt engineering can enforce accuracy and safety.

Option D is correct because the described pattern—decreasing training perplexity alongside increasing validation perplexity—is the classic signature of overfitting. The model is memorizing the training data rather than learning generalizable patterns, causing its performance on unseen validation data to degrade after a certain point in training.

Input validation and sanitisation (A) prevent malicious user inputs from being interpreted as instructions by the LLM, directly mitigating prompt injection by stripping or escaping special characters and control sequences. A strict system prompt (B) defines the chatbot's role and boundaries, reducing the attack surface by making it harder for injected prompts to override the intended behavior.

Semi-supervised learning uses the small labeled set to guide learning from the large unlabeled set. Self-supervised learning would require a pretext task; fine-tuning a pre-trained model is also valid but semi-supervised directly addresses the labeled-unlabeled mix.

A is correct because a system-level prompt establishes a foundational instruction set that defines the model's allowed behavior and boundaries. This acts as a first line of defense by explicitly instructing the model to ignore any user attempts to override its core directives, thereby reducing the risk of prompt injection attacks.

Dropout randomly drops neurons during training, preventing co-adaptation and overfitting. L1 and L2 add penalties to weights but are less common for transformers; L1 induces sparsity, L2 reduces weight magnitude. However, dropout is the standard regularization in BERT-like models.

Logistic regression is interpretable (coefficients show feature impact) and performs well on binary classification with a large dataset. Decision trees are interpretable but may overfit; random forests and gradient boosting are less interpretable.

With 99.9% legitimate transactions and only 0.1% fraudulent ones, accuracy would be misleadingly high (99.9%) even if the model never predicts fraud. The F1-score is the harmonic mean of precision and recall, making it robust to class imbalance by penalizing both false positives and false negatives. This makes it the most appropriate metric for evaluating fraud detection performance.

GANs (Generative Adversarial Networks) consist of a generator and discriminator that compete, producing highly realistic images. Diffusion models are also good but GANs are historically the go-to for image generation. VAEs produce blurrier images; LLMs are for text.

With high class imbalance (90% negatives), accuracy is misleading. F1 score balances precision and recall, giving a better picture of performance on the minority class. AUC-ROC is also good but F1 directly optimizes for positive class.

Named entity recognition (NER) identifies predefined entities like names, dates, or product names. For extracting key phrases, keyword extraction or topic modeling (e.g., TF-IDF, RAKE) is more appropriate. Sentiment analysis gives sentiment scores, and text classification assigns categories.

A high-variance loss plateau after rapid initial convergence typically indicates that the learning rate is too large, causing the optimizer to overshoot the minima and oscillate. Decreasing the learning rate allows smaller, more stable weight updates, reducing variance and enabling smoother convergence.

AUC-ROC is robust to class imbalance because it measures the trade-off between true positive rate and false positive rate across all thresholds. Accuracy is misleading when classes are imbalanced. Precision and recall focus on one class but are threshold-dependent.

Train/validation/test split is standard; cross-validation gives more robust estimates. Shuffling before split prevents ordering bias.

L2 regularization (also known as weight decay) adds a penalty proportional to the square of the magnitude of the weights to the loss function. This discourages the model from learning overly complex patterns that fit the training data noise, effectively reducing overfitting by keeping weights small and the decision boundary simpler.

Generative Adversarial Networks (GANs) are widely used for image generation, and diffusion models (like Stable Diffusion) have achieved state-of-the-art results in image synthesis. Variational autoencoders (VAEs) can generate images but often produce blurrier outputs. GPT is for text, and BERT is for understanding.

Fine-tuning the existing model on a corpus of code-switched text adapts the model to the new language pattern efficiently.

Collaborative filtering recommends based on user similarities. K-Nearest Neighbors can find similar users or items. Both are suitable for recommendation.

Option B is correct because stratified train-test splitting ensures that the class distribution (80% positive, 20% negative) is preserved in both training and test sets. This prevents the model from being evaluated on a test set that has a different class ratio, which could give a misleading impression of performance, especially in imbalanced datasets.

Early stopping halts training when validation loss stops improving, preventing overfitting. Increasing learning rate would worsen divergence; adding more layers increases capacity and overfitting; reducing batch size may help optimization but not directly address overfitting.

Gradient clipping directly addresses the problem of exploding gradients, which can cause large, destabilizing weight updates during fine-tuning of large models like BERT. By capping the gradient norm (e.g., to a value like 1.0), it prevents a single batch from drastically altering the model's parameters, thus smoothing out validation performance fluctuations across epochs.

Fine-tuning the pre-trained model on the small labeled dataset is standard for transfer learning with BERT. Training from scratch would require massive data; feature extraction with a linear classifier is possible but less effective with limited data; data augmentation alone does not adapt the model.

Data augmentation (A) is effective for preventing overfitting because it artificially expands the training dataset by applying random transformations (e.g., rotation, flipping, cropping, color jitter) to existing images. This exposes the CNN to a wider variety of input patterns, reducing the model's tendency to memorize noise or specific details and improving generalization to unseen data.

Collaborative filtering uses user-item interactions; content-based filtering uses item features. Association rule mining is for basket analysis; regression is not typically used for recommendations; anomaly detection is for outliers.

Narrow AI specializes in one task, while general AI would possess human-like cognitive abilities across domains.

Normalizing pixel values to zero mean and unit variance (A) ensures that input features have similar scales, which prevents certain weights from updating disproportionately and stabilizes gradient descent. This is especially important for CNNs processing satellite imagery, where raw pixel intensities can vary widely across bands and scenes, leading to poor convergence.

If the generator loss drops too fast and discriminator loss rises, the generator is overpowering the discriminator. The typical remedy is to train the discriminator more often (e.g., 5 steps per generator step) or adjust the learning rates.

Narrow AI, also known as Weak AI, is designed and trained to perform a single, specific task with high proficiency, such as language translation, image recognition, or playing chess. It cannot generalize its intelligence to other domains, which distinguishes it from Artificial General Intelligence (AGI). Option C correctly captures this fundamental characteristic.

The F1 score is appropriate because it balances precision and recall, making it robust for binary classification when class distribution may be imbalanced. AUC-ROC measures the model's ability to distinguish between positive and negative classes across all classification thresholds, providing a threshold-independent evaluation of discriminative performance.

Overfitting occurs when the model learns the training data too well, including noise and irrelevant patterns, causing it to memorize rather than generalize. This is evidenced by the loss decreasing on the training set while increasing on the validation set after a few epochs, as the model's performance on unseen data degrades.

L2 regularization (option B) penalizes large weights by adding a squared magnitude term to the loss function, which discourages the model from fitting noise in the training data. This directly reduces overfitting while maintaining high precision and recall by keeping the decision boundary smooth and generalizable.

Lowering temperature makes the model more deterministic and less likely to produce random outputs. Top-p and top-k can also help but are secondary; temperature directly controls randomness.

Random Forest handles irrelevant features well via feature importance. Lasso (L1) regression performs automatic feature selection by shrinking coefficients to zero.

This is a binary classification problem because the target variable has exactly two discrete outcomes: 'yes' (churn) or 'no' (no churn). Classification algorithms such as logistic regression, decision trees, or support vector machines are used to assign input features to one of these two predefined classes. The output is a categorical label, not a continuous value or a reward signal.

CNNs are ideal for image feature extraction; RNNs/LSTMs are designed for sequence modelling to track temporal patterns.

RAG retrieves relevant documents from the knowledge base at inference time, grounding the model's responses in facts. Prompt engineering alone can't ensure factual accuracy; fine-tuning may still hallucinate; temperature reduction only reduces randomness.

Convolutional Neural Networks (CNNs) are the standard architecture for image classification because they use convolutional layers to automatically learn spatial hierarchies of features (edges, textures, shapes) from pixel data. They are highly effective for large labeled image datasets due to their parameter efficiency and translation invariance.

Option B is correct because distribution shift (data drift) causes the model's training distribution to differ from the real-world distribution, degrading performance despite high test accuracy. Implementing a monitoring system to detect drift and retraining with fresh data directly addresses this by ensuring the model adapts to the current data distribution, which is the most critical action for maintaining performance in production.

Unsupervised anomaly detection is the most appropriate approach because the system must learn 'normal' traffic patterns from unlabeled data and then flag deviations without requiring pre-labeled examples of attacks. This aligns with the core requirement of detecting unknown or novel security breaches, which supervised methods cannot handle due to the lack of labeled attack data.

Unsupervised anomaly detection is the most suitable approach because the startup has a large dataset of normal traffic but very few labeled attacks. This technique learns the baseline of normal behavior from unlabeled data and flags deviations as potential anomalies, which is ideal for detecting unknown or rare attack patterns without requiring labeled attack samples.

Artificial General Intelligence (AGI) is the concept of a machine that understands or learns any intellectual task that a human being can.

Increasing the temperature parameter raises the randomness of token selection, encouraging the model to explore less probable word sequences and produce more varied, creative, and longer outputs. A low temperature (e.g., 0.1) makes the model deterministic and repetitive, often yielding short, generic responses. By increasing temperature (e.g., to 0.8 or 1.0), the model is more likely to generate diverse and expansive text, directly addressing the user's request for longer, more creative poems.

RNNs (including LSTMs) are designed for sequential data with temporal dependencies. CNNs excel at spatial data; transformers are also used but RNNs are the classic answer.

Recurrent Neural Networks (RNNs) are specifically designed for sequential data because they maintain a hidden state that is updated at each time step, allowing information about previous inputs to persist and influence current and future outputs. This feedback loop makes them ideal for tasks like time series forecasting, natural language processing, and speech recognition, where order and context matter.

LSTM (Long Short-Term Memory) is a type of RNN designed with gating mechanisms to prevent vanishing gradients in long sequences. CNNs are for spatial data; vanilla RNNs suffer from vanishing gradients; transformers use attention but are not specifically designed to mitigate vanishing gradients (they use residual connections).