Pearson IT Specialist: Artificial Intelligence (INF-307) — Questions and Answers

Question 1: What is 'perplexity' used to measure in language models?

The number of parameters in the model
How well the model predicts a sample of text; lower is better (Correct answer)
The vocabulary size of the model
The time taken to generate a response

Correct answer: How well the model predicts a sample of text; lower is better

Perplexity is the exponentiated average negative log-likelihood per token; it measures how surprised the model is by the test text, with lower values indicating better predictions.

Question 2: What distinguishes a recurrent neural network (RNN) from a feedforward network?

RNNs have feedback connections that allow information to persist across time steps (Correct answer)
RNNs cannot process sequences
RNNs always use attention mechanisms
RNNs use convolutional filters

Correct answer: RNNs have feedback connections that allow information to persist across time steps

RNNs have recurrent connections that pass hidden states from one time step to the next, enabling them to model sequential and temporal data.

Question 3: What is 'consent' in the context of AI and data privacy?

Freely given, specific, informed, and unambiguous permission from individuals for their data to be collected and used (Correct answer)
The terms of service for using an AI application
The agreement between AI developers and cloud providers
A legal document signed before model training

Correct answer: Freely given, specific, informed, and unambiguous permission from individuals for their data to be collected and used

Meaningful consent ensures individuals know what data is collected, how it is used, and can withdraw permission, forming the ethical and legal foundation of data-driven AI.

Question 4: What does 'fine-tuning' a pre-trained language model involve?

Continuing training on a task-specific labeled dataset to adapt the model to a new task (Correct answer)
Re-training the model entirely from scratch on new data
Converting the model to a different programming language
Pruning the model to reduce its size

Correct answer: Continuing training on a task-specific labeled dataset to adapt the model to a new task

Fine-tuning continues gradient-based training of a pre-trained model on a smaller task-specific dataset, adapting general representations to the target task.

Question 5: What does the subword tokenization algorithm BPE (Byte Pair Encoding) do?

Removes all punctuation from text
Iteratively merges the most frequent character pairs to build a vocabulary of subword units (Correct answer)
Converts all tokens to a fixed embedding size
Splits text only on whitespace

Correct answer: Iteratively merges the most frequent character pairs to build a vocabulary of subword units

BPE starts with individual characters and repeatedly merges the most frequent adjacent pair until a target vocabulary size is reached, balancing between word and character tokenization.

Question 6: What is an 'adversarial example' in the context of AI security?

An input deliberately modified with small perturbations to cause an AI model to make wrong predictions (Correct answer)
A training sample with an incorrect label
An example used to evaluate model robustness
A competitor's AI product

Correct answer: An input deliberately modified with small perturbations to cause an AI model to make wrong predictions

Adversarial examples are inputs crafted by adding carefully chosen, often imperceptible perturbations that reliably fool AI classifiers into producing incorrect outputs.

Question 7: What is zero-shot classification in the context of large language models?

Classifying inputs into categories the model has never explicitly been trained on, using only natural language descriptions (Correct answer)
Removing all zero-value parameters from a model
Running inference without any GPU acceleration
Training a model on zero labeled examples then evaluating on training data

Correct answer: Classifying inputs into categories the model has never explicitly been trained on, using only natural language descriptions

Zero-shot classification leverages a model's pre-trained knowledge to assign labels described in natural language without any task-specific training examples.

Question 8: What is 'value alignment' in AI safety research?

Standardizing AI model file formats across organizations
Calibrating AI confidence scores to match actual accuracy
The challenge of designing AI systems whose goals and behaviors align with human values and intentions (Correct answer)
Ensuring AI models have consistent numerical outputs

Correct answer: The challenge of designing AI systems whose goals and behaviors align with human values and intentions

Value alignment addresses the problem of ensuring that as AI systems become more capable, their objectives remain consistent with human values rather than pursuing misaligned proxy goals.

Question 9: What is 'model fairness' in AI, and why does it matter?

Ensuring the model's predictions are equitable across demographic groups and do not perpetuate systemic discrimination (Correct answer)
Making sure the model is open-source
Ensuring the model runs at the same speed for all users
Ensuring training and test accuracy are similar

Correct answer: Ensuring the model's predictions are equitable across demographic groups and do not perpetuate systemic discrimination

Model fairness ensures that AI decisions do not systematically disadvantage protected groups (e.g., by race, gender, age), promoting equitable outcomes in consequential applications.

Question 10: What is the 'right to explanation' under GDPR in the context of automated AI decisions?

The right to request a copy of your personal data
The right to opt out of all AI recommendations
The right of individuals to receive meaningful information about the logic behind automated decisions that significantly affect them (Correct answer)
The right to delete your data from AI training sets

Correct answer: The right of individuals to receive meaningful information about the logic behind automated decisions that significantly affect them

GDPR Article 22 grants individuals the right not to be subject to solely automated decisions with significant effects, and to obtain meaningful explanations of the decision logic.

Question 11: What is 'image augmentation' with random cropping designed to help the model learn?

To convert color images to grayscale
To resize all images to a uniform size
To make predictions invariant to the object's position and scale within the image (Correct answer)
To add text labels to training images

Correct answer: To make predictions invariant to the object's position and scale within the image

Random cropping exposes the model to objects appearing at different positions and scales, building spatial invariance into the learned features.

Question 12: What is the role of an activation function in a neural network?

Calculate the loss
Introduce non-linearity into the model (Correct answer)
Normalize the input data
Initialize the weights

Correct answer: Introduce non-linearity into the model

Activation functions add non-linearity, allowing neural networks to learn complex patterns beyond linear mappings.

Question 13: Which technique allows a network to locate the image regions most responsible for a classification decision?

Max pooling
Batch normalization
Dropout
Grad-CAM (Gradient-weighted Class Activation Mapping) (Correct answer)

Correct answer: Grad-CAM (Gradient-weighted Class Activation Mapping)

Grad-CAM uses gradients of the target class score flowing into the final convolutional layer to produce a heatmap highlighting discriminative image regions.

Question 14: What does PCA (Principal Component Analysis) accomplish?

Removes outliers from a dataset
Reduces dimensionality by projecting data onto principal components (Correct answer)
Classifies data into clusters
Generates synthetic data samples

Correct answer: Reduces dimensionality by projecting data onto principal components

PCA finds orthogonal axes of maximum variance and projects data onto fewer dimensions while preserving variance.

Question 15: Which technique adapts a large pre-trained model to a specific task using a smaller labeled dataset?

Feature hashing
Fine-tuning (Correct answer)
Random initialization
Data augmentation

Correct answer: Fine-tuning

Fine-tuning continues training a pre-trained model on task-specific data, leveraging transfer learning.

Question 16: Which NLP task determines whether the relationship between two sentences is entailment, contradiction, or neutral?

Coreference resolution
Question answering
Natural language inference (Correct answer)
Semantic role labeling

Correct answer: Natural language inference

Natural language inference (NLI) classifies the logical relationship between a premise and a hypothesis sentence as entailment, contradiction, or neutral.

Question 17: What is the term for a model that performs well on training data but poorly on new data?

Generalization
Underfitting
Regularization
Overfitting (Correct answer)

Correct answer: Overfitting

Overfitting occurs when a model memorizes training data instead of learning generalizable patterns.

Question 18: Which layer type in a convolutional neural network applies a learned filter to detect local patterns in an image?

Pooling layer
Fully connected layer
Convolutional layer (Correct answer)
Dropout layer

Correct answer: Convolutional layer

A convolutional layer slides learned filters over the input to produce feature maps capturing local spatial patterns like edges and textures.

Question 19: Which word embedding model learns vector representations by predicting surrounding words (skip-gram) or predicting a word from context (CBOW)?

Word2Vec (Correct answer)
FastText
GloVe
BERT

Correct answer: Word2Vec

Word2Vec offers two architectures: skip-gram predicts context words from a target, and CBOW predicts a target word from its context window.

Question 20: What does 'transparency' require of AI developers and deployers?

Ensuring the AI is available 24/7 without downtime
Being open about how AI systems work, what data they use, their limitations, and potential risks (Correct answer)
Making all model weights publicly available
Publishing all training datasets online

Correct answer: Being open about how AI systems work, what data they use, their limitations, and potential risks

Transparency requires that AI developers disclose how systems are built, what they can and cannot do, and where they may fail, enabling informed use and external scrutiny.

Question 21: What is 'data minimization' as it relates to responsible AI?

Reducing the number of training epochs
Collecting and retaining only the personal data strictly necessary for the intended AI application (Correct answer)
Limiting the size of input images
Using the smallest possible neural network

Correct answer: Collecting and retaining only the personal data strictly necessary for the intended AI application

Data minimization reduces privacy risk by limiting data collection and retention to what is genuinely needed, aligning with GDPR and privacy-by-design principles.

Question 22: The Turing Test is designed to evaluate what aspect of a machine?

Its computational speed
Its memory capacity
Its ability to solve math proofs
Its ability to exhibit human-indistinguishable conversation (Correct answer)

Correct answer: Its ability to exhibit human-indistinguishable conversation

Alan Turing proposed the test to judge whether a machine's conversational responses are indistinguishable from a human's.

Question 23: What is 'data privacy' in the context of AI model training?

Ensuring that personal data used in training is handled, stored, and used in accordance with privacy laws (Correct answer)
Encrypting model weights
Removing duplicate records from training data
Using only public datasets

Correct answer: Ensuring that personal data used in training is handled, stored, and used in accordance with privacy laws

Data privacy in AI means protecting individuals' personal information used during model training from unauthorized access or misuse.

Question 24: What does the activation function in a neural network primarily provide?

Data normalization
Weight initialization
Non-linearity (Correct answer)
Faster training speed

Correct answer: Non-linearity

Activation functions like ReLU introduce non-linearity, letting networks model complex relationships.

Question 25: What does data augmentation in computer vision typically involve?

Collecting additional labeled images from the internet
Applying random transformations like flips, rotations, and crops to training images to improve generalization (Correct answer)
Converting images to grayscale before training
Compressing training images to reduce disk usage

Correct answer: Applying random transformations like flips, rotations, and crops to training images to improve generalization

Data augmentation artificially expands training data by applying label-preserving transformations, helping models become robust to variations in orientation, scale, and lighting.

Question 26: In a seq2seq model for machine translation, what is the role of the encoder?

To score candidate translations
To generate the translated output token by token
To compress the source sentence into a context representation (Correct answer)
To perform beam search over possible outputs

Correct answer: To compress the source sentence into a context representation

The encoder processes the source sequence and produces a fixed-size context vector (or sequence of hidden states) that summarizes its meaning for the decoder.

Question 27: What is the purpose of a dropout layer in a neural network?

To normalize activations across a batch
To speed up forward passes
To increase the number of parameters
To randomly deactivate neurons during training to reduce overfitting (Correct answer)

Correct answer: To randomly deactivate neurons during training to reduce overfitting

Dropout randomly sets a fraction of neuron activations to zero during each training step, acting as an ensemble regularization technique.

Question 28: In machine learning, what does 'hyperparameter tuning' refer to?

Selecting model parameters that are set before training (Correct answer)
Adjusting weights during backpropagation
Normalizing the training dataset
Increasing the number of training samples

Correct answer: Selecting model parameters that are set before training

Hyperparameter tuning searches for the best configuration values (e.g., learning rate, depth) that are set before training begins.

Question 29: What is the primary difference between extractive and abstractive text summarization?

Extractive selects existing sentences from the source; abstractive generates new sentences (Correct answer)
Extractive summarization works only on emails; abstractive on articles
Extractive uses neural networks; abstractive uses rule-based systems
Extractive is always shorter than abstractive summaries

Correct answer: Extractive selects existing sentences from the source; abstractive generates new sentences

Extractive summarization picks and ranks existing sentences from the document, while abstractive summarization generates novel sentences that may not appear verbatim in the source.

Question 30: What is 'hallucination' in large language models?

The model generating images during text tasks
A training instability that causes random outputs
Generating confident but factually incorrect or fabricated information (Correct answer)
Repeating the same token endlessly

Correct answer: Generating confident but factually incorrect or fabricated information

LLM hallucination occurs when the model produces plausible-sounding but false or invented information with unwarranted confidence.

Question 31: What does regularization in machine learning primarily address?

Underfitting
Data normalization
Overfitting (Correct answer)
Class imbalance

Correct answer: Overfitting

Regularization adds a penalty to the loss function to reduce model complexity and prevent overfitting.

Question 32: In machine learning, what problem occurs when a model performs well on training data but poorly on unseen data?

Data leakage
Overfitting (Correct answer)
Underfitting
Vanishing gradients

Correct answer: Overfitting

Overfitting means the model memorized training-specific noise instead of learning generalizable patterns.

Question 33: In forward chaining, an AI system:

Randomly selects rules to apply until a solution is found
Uses neural network layers to propagate activations forward
Starts from a goal and works backward to find supporting facts
Starts from available facts and applies rules to derive new conclusions until the goal is reached (Correct answer)

Correct answer: Starts from available facts and applies rules to derive new conclusions until the goal is reached

Forward chaining (data-driven reasoning) begins with known facts and fires applicable rules to generate new facts, continuing until the target goal is derived or no more rules can fire.

Question 34: Which principle of responsible AI states that AI systems should cause minimal harm and consider the well-being of all stakeholders?

Non-maleficence (Correct answer)
Fairness
Accountability
Transparency

Correct answer: Non-maleficence

Non-maleficence ('do no harm') requires AI systems to avoid causing physical, psychological, financial, or social harm to individuals or society.

Question 35: What architecture is BERT (Bidirectional Encoder Representations from Transformers) based on?

Transformer decoder
Transformer encoder (Correct answer)
Convolutional neural network
Recurrent neural network

Correct answer: Transformer encoder

BERT uses only the encoder portion of the Transformer and is pre-trained with bidirectional context, reading the entire sentence at once.

Question 36: What is the 'attention mechanism' in neural NLP models?

A way to speed up tokenization
A technique for removing irrelevant tokens
A rule for choosing the correct part of speech
A method that allows the model to weight the importance of different input tokens when producing an output (Correct answer)

Correct answer: A method that allows the model to weight the importance of different input tokens when producing an output

Attention lets the model dynamically focus on relevant parts of the input sequence when generating each output token, enabling better handling of long-range dependencies.

Question 37: What is backpropagation in neural network training?

Normalizing layer activations
Using the chain rule to compute gradients and update weights (Correct answer)
Feeding data forward through the network
Randomly initializing network weights

Correct answer: Using the chain rule to compute gradients and update weights

Backpropagation computes gradients of the loss with respect to all weights by applying the chain rule backwards through the network.

Question 38: What is 'differential privacy' used for in AI and machine learning?

Adding mathematically calibrated noise to data or outputs to protect individual privacy while allowing aggregate analysis (Correct answer)
Detecting adversarial examples in model inputs
Balancing class distributions in training data
Training models faster on distributed hardware

Correct answer: Adding mathematically calibrated noise to data or outputs to protect individual privacy while allowing aggregate analysis

Differential privacy provides a formal guarantee that the inclusion of any individual's data in training has a bounded, negligible effect on model outputs, protecting personal information.

Question 39: What is named entity recognition (NER) in NLP?

Generating summaries of long documents
Identifying and classifying real-world entities such as people, organizations, and locations in text (Correct answer)
Detecting the language of a text
Counting the frequency of words in a corpus

Correct answer: Identifying and classifying real-world entities such as people, organizations, and locations in text

NER identifies spans of text that refer to specific entity categories (persons, organizations, dates, etc.) and tags them accordingly.

Question 40: An AI chatbot confidently states a false fact that sounds plausible. What is this phenomenon called?

Tokenization
Hallucination (Correct answer)
Quantization
Regularization

Correct answer: Hallucination

Hallucination refers to language models generating fluent but factually incorrect or fabricated content.

Question 41: Which technique allows a language model to answer questions about a document it wasn't trained on by providing that document as context?

Model distillation
Prompt tuning
Retrieval-augmented generation (Correct answer)
Knowledge distillation

Correct answer: Retrieval-augmented generation

Retrieval-augmented generation (RAG) retrieves relevant documents at inference time and includes them in the model's context, enabling factual answers beyond the model's training knowledge.

Question 42: What is the study of computer algorithms that improve themselves over time?

Ability
Formulation
Compilation
Machine learning (Correct answer)

Correct answer: Machine learning

Machine learning is a subfield of artificial intelligence focused on developing algorithms that allow computers to learn from data without being explicitly programmed. These algorithms identify patterns, make predictions, and improve their performance over time as they are exposed to more data. This iterative improvement is a core characteristic that distinguishes machine learning from traditional programming.

Question 43: Which metric is most appropriate when classes in a dataset are heavily imbalanced?

R-squared
Mean Squared Error
Accuracy
F1 Score (Correct answer)

Correct answer: F1 Score

F1 Score balances precision and recall, making it more informative than accuracy on imbalanced datasets.

Question 44: What does precision measure in a classification model?

The fraction of positive predictions that were correct (Correct answer)
The speed of inference
The fraction of actual positives that were found
The total accuracy across all classes

Correct answer: The fraction of positive predictions that were correct

Precision is true positives divided by all positive predictions, measuring prediction correctness.

Question 45: What is RLHF (Reinforcement Learning from Human Feedback) used for?

Fine-tuning vision models on human-labeled data
Evaluating RL agents in human-designed environments
Training robots to walk using human demonstration
Aligning LLM outputs with human preferences using reward signals from human raters (Correct answer)

Correct answer: Aligning LLM outputs with human preferences using reward signals from human raters

RLHF trains a reward model from human preference comparisons, then uses RL to fine-tune the LLM to produce higher-rated outputs.

Pearson IT Specialist: Artificial Intelligence (INF-307)

The Pearson IT Specialist Artificial Intelligence exam (INF-307) validates foundational knowledge and practical skills in AI, covering problem definition, data engineering, AI algorithms and models, application deployment, and monitoring AI systems in production.

Questions

50 min

Time Limit

70%

To Pass

Checking for saved progress...

Exam Rules

You can skip questions and return to them later
Flag questions for review before submitting
No feedback shown until you submit the entire exam
Unanswered questions count as wrong — answer everything
10 pretest questions are mixed in and don't affect your score
Timer auto-submits when time runs out
Your progress is auto-saved every 30 seconds