ELEC2 - FINALS REVIEWER

Which model serves as the universal framework for every intelligent system?

The Input–Process–Output (IPO) Model

In an intelligent system, the raw material (structured or unstructured) is called?

Input Data

What is the key characteristic of an intelligent system?

It adapts based on data to mimic human-like reasoning

What does the principle of "Garbage in, garbage out" emphasize?

That the quality of output depends on the quality of input data

Which process is used to clean data and handle missing values before training?

Preprocessing

If you are predicting a continuous numeric value, such as a student's final grade percentage, you are using what type of learning?

Regression

Which algorithm category is used to assign inputs to discrete categories such as Pass or Fail?

Classification

For a very small dataset, which design choice helps avoid overfitting?

Choosing a simpler model

Why is Python the industry standard for intelligent system prototyping?

Its readable syntax and extensive ML library ecosystem

Which Python library is used for loading, cleaning, and transforming datasets?

Pandas

What is the primary goal of creating a working prototype?

To establish a baseline and prove the design is functional

In the system integration phase, what does smooth data flow ensure?

Output formats from one stage are valid inputs for the next

A model that performs perfectly on training data but fails on new data is experiencing what problem?

Overfitting

What is the systematic process of identifying and resolving errors in an AI system called?

Debugging

Issues like label noise or insufficient edge case coverage are examples of what?

Data Issues

Which tool helps explain which features are most influential for a specific prediction?

SHAP values

The F1 Score is particularly valuable because it is what?

The harmonic mean of precision and recall

What does Precision specifically measure?

How many positive predictions were actually correct

On a highly imbalanced dataset, which metric can be the most misleading?

Accuracy

Transitioning a model from a notebook to a live environment for real users is called what?

Deployment

What are the two most fundamental performance measures for a classifier?

Accuracy and Error Rate

What is accuracy defined as?

Correct predictions divided by total predictions

What is error rate defined as?

One minus accuracy, or the fraction of predictions that are wrong

Why can accuracy be misleading on imbalanced datasets?

A model predicting the majority class always gets high accuracy without actually learning anything

What does a confusion matrix show?

A breakdown of true positives, false negatives, false positives, and true negatives for a classifier

What does Recall or Sensitivity measure?

How many actual positives the model correctly identified, calculated as TP divided by TP plus FN

What is Specificity?

The proportion of actual negatives correctly identified, calculated as TN divided by TN plus FP

What is a model experiencing when it is too simple, has high training error and high test error, and fails to capture patterns?

Underfitting, also called high bias

What is a model experiencing when it is too complex, has low training error but high test error, and memorizes noise?

Overfitting, also called high variance

What does total prediction error approximately equal in the bias-variance tradeoff?

Bias squared plus Variance plus Irreducible Noise

What does a large gap between low training error and high validation error signal?

Overfitting

What does it signal when both training error and validation error remain high?

Underfitting, meaning more complexity or features should be added

Why is model validation necessary?

Because a model's training accuracy is an optimistic and biased estimate of real-world performance

What is a Hold-Out split strategy?

Dividing data into 80 percent training and 20 percent testing, which is fast but sensitive to how the split falls

What is k-Fold Cross-Validation?

A strategy where data is split into k folds and each fold takes a turn as the validation set, producing a more stable and reliable estimate

What is data leakage in model evaluation?

When preprocessing is fitted on the full dataset before splitting, causing test data to bleed into training

What is the rule of thumb regarding the test set during model development?

Never touch the test set until the model is finalized, as using it for tuning causes data leakage

What are the core components of an intelligent system?

Sensors or Input Layer, Knowledge Base, Inference Engine, and Output or Actuators

What does the Inference Engine do?

It applies logic or learned patterns to input data to derive conclusions or predictions

What does the Knowledge Base store?

Domain knowledge, rules, or trained model parameters that the system uses to reason

What is the IPO model used for?

It is the universal framework for describing how any system transforms raw data into meaningful results through Input, Process, and Output stages

What are the static components of an intelligent system?

The Knowledge Base, User Interface, and Explanation Module

What are the dynamic components of an intelligent system?

Working Memory, Learning Module, and Control Strategy

Which model selection factors favor choosing a simpler model?

Limited training data, high interpretability requirements, real-time inference constraints, low computational budget, and well-structured problem domains

Which model selection factors favor choosing a more complex model?

Large high-dimensional datasets, unstructured data such as images or text, accuracy as the primary metric, offline or batch processing, and sufficient compute resources

What is a Decision Tree best used for?

Classification with interpretable rules, though it risks overfitting on small datasets

What are Neural Networks ideal for?

Complex pattern recognition tasks like image and speech processing, requiring large datasets and significant compute

What is k-Nearest Neighbors best suited for?

Small datasets with well-defined feature spaces, though it is slow at prediction time

What are Rule-Based Systems best for?

Problems where domain knowledge is well-established, requiring explicit IF-THEN logic that is highly interpretable and deterministic

What are the three types of bias that can enter an AI system?

Training Data Bias, Algorithmic Bias, and Feedback Loop Bias

What is Training Data Bias?

When historical data reflects societal prejudices and the model learns and replicates those patterns

What is Algorithmic Bias?

When the design of the algorithm itself favors certain outcomes

What is Feedback Loop Bias?

When biased outputs influence future data, reinforcing the original bias

What are the four definitions of fairness in AI?

Individual Fairness, Group Fairness, Counterfactual Fairness, and Calibration

What is a key insight about satisfying all fairness definitions simultaneously?

It is mathematically impossible to satisfy all fairness definitions at once, so designers must make deliberate and justified trade-offs

What are the five core principles of Responsible AI?

Transparency, Accountability, Safety, Fairness, and Privacy

What is the Black Box problem in AI?

Many powerful AI models, particularly deep neural networks, are opaque and their internal reasoning is difficult or impossible to interpret

What are explainability techniques used in AI?

LIME which stands for Locally Interpretable Model-agnostic Explanations, SHAP which assigns feature importance scores, and Attention Maps which visualize which input features the model focused on

What is the EU AI Act?

The world's first comprehensive AI law, enacted in 2024, which classifies AI systems by risk level with strict requirements for high-risk applications

What are privacy-preserving AI techniques?

Differential Privacy, Federated Learning, and Homomorphic Encryption

What is Differential Privacy?

A technique that adds carefully calibrated mathematical noise to data or model outputs, making it statistically impossible to identify any individual's contribution

What is Federated Learning?

A technique that trains models across decentralized devices without transferring raw data to a central server

What is Homomorphic Encryption?

A technique that allows computations to be performed on encrypted data without decrypting it first

What are the key privacy principles in data protection?

Data Minimization, Informed Consent, Right to Explanation, and Purpose Limitation

True or False Prior knowledge is needed to give AI predictions

True

True or False Logic contains concepts and knowledge bases

True

True or False Semi-structured data has custom tags

True

True or False Unstructured problems have clear interfaces

False

True or False Metadata describes or has data

True

True or False Images and videos are unstructured data

False

True or False Data users inform managers of specific areas

True

True or False Intelligent systems are best used in areas with no data available

False

True or False Structured data has no predefined format

True

True or False Intelligent systems work without human input

False

True or False Semi-structured problems require human judgment and needs

True

True or False Intelligent systems can learn from experience

True