INT428 AI Essentials — Ultimate Exam Prep

🔍 Search Results

🧠 INT428 AI Essentials
Ultimate Exam Prep Package

Complete crash course built from your syllabus, sample paper (60 Qs), CA papers & unit notes (Units 1–4). Units 5–6 intelligently inferred. Optimized for MCQ-based exam.

📚 6 Units❓ 105+ MCQs🃏 40 Flashcards📝 Mock Test📋 Cheat Sheets🏆 Top 100 Points

6

Units Covered

105+

MCQs Generated

40

Flashcards

60

Sample Paper Qs

🎯 Unit Quick Access

UNIT 01

Foundations & Applications

Intelligence, AI types, Expert Systems, ML, DL, NLP, Applications, Toolkits

UNIT 02

Problem Solving & Search

BFS, DFS, A*, Greedy, Heuristics, Water Jug, 8-Puzzle, Minimax, Hill Climbing

UNIT 03

Machine Learning

Supervised/Unsupervised/RL, Bayes, KNN, K-Means, Confusion Matrix, Feature Eng.

UNIT 04

Deep Learning & NLP

Perceptron, MLP, CNN, RNN, Transformer, BERT, GPT, Chatbots, Attention

UNIT 05 ★ INFERRED

Generative AI & Prompts

LLMs, GANs, VAE, Diffusion, Prompt Engineering, Zero-shot, Ethics

UNIT 06 ★ INFERRED

Data, Tools & MLOps

ChatGPT ADA, Tableau, Data pipelines, Cloud, MLOps, Error handling, Python

🔥 Hottest Topics (From Paper Analysis)

🥇 A* Search

f(n)=g(n)+h(n). In nearly every paper. Know formula + example cold.

🥈 Confusion Matrix

Precision & Recall calculations. Numerical MCQs guaranteed in every CA.

🥉 CNN Architecture

Conv→Pool→FC. Role of each layer. Match-the-model questions.

4️⃣ BERT vs GPT

Bidirectional vs Autoregressive. Training objectives. Use cases.

5️⃣ Bayes Theorem

Numerical P(S|W) calculation. Feature independence in Naive Bayes.

6️⃣ GAN Components

Generator creates samples. Discriminator judges. Tested in Q7 sample paper.

📊 Complete Exam Analysis

📋 Exam Pattern (INT428)

60 MCQs × 1 mark = 60 marks. 0.25 negative marking per wrong answer. Time: 3 hours. OMR sheet. All 60 questions compulsory.

🎯 Unit Weightage

Unit	Topic	Est. MCQs	Priority
Unit 1	Foundations, AI types, Applications, Toolkits	8–10	HIGH
Unit 2	Search Algorithms, Heuristics, Game Trees, Logic	10–12	VERY HIGH
Unit 3	ML, Bayesian Networks, Confusion Matrix, Clustering	10–12	VERY HIGH
Unit 4	Deep Learning, CNN, RNN, Transformers, NLP, BERT/GPT	12–14	HIGHEST
Unit 5	GenAI, GANs, LLMs, Prompt Engineering, Ethics	10–12	HIGH
Unit 6	Data Tools, ChatGPT ADA, MLOps, Python pandas	6–8	MEDIUM

🔁 Most Repeated Concepts

🔥1

A* Search: f(n)=g(n)+h(n). Both CA paper and sample paper tested it multiple times.

🔥2

Confusion Matrix: Precision=TP/(TP+FP), Recall=TP/(TP+FN). Numerical MCQs in every CA.

🔥3

Bayes Theorem: P(A|B)=P(B|A)P(A)/P(B). Spam filter or medical example most common.

🔥4

CNN vs RNN vs Transformer: Match-the-model questions appear repeatedly across papers.

🔥5

GAN Components: Generator generates samples, Discriminator evaluates. Tested in sample paper Q7.

🔥6

BERT Bidirectional: MLM + NSP training objectives. Reads L↔R simultaneously.

🔥7

K-Means Clustering: 1-iteration numerical problem guaranteed in every CA paper.

🔥8

Positional Encoding: Why needed in Transformers — parallel processing loses word order.

⚠️ Top MCQ Traps

TRAP 1: Greedy BFS vs A*Greedy = h(n) only. A* = g(n)+h(n). A* guarantees optimal if h admissible. Greedy does NOT guarantee optimality.

TRAP 2: Precision vs RecallPrecision = TP/(TP+FP). Recall = TP/(TP+FN). Fraud detection: misses fraud = LOW RECALL, rarely flags normal = HIGH PRECISION.

TRAP 3: BERT vs GPT DirectionBERT = Bidirectional. GPT = Unidirectional (left→right). BERT: understanding tasks. GPT: generation tasks.

TRAP 4: GAN Generator vs DiscriminatorGenerator CREATES new samples (answer in Q7). Discriminator JUDGES real vs fake. Never confuse these.

TRAP 5: Branching Factor= Number of states generated FROM a given state. NOT paths to goal, NOT heuristic values, NOT depth.

TRAP 6: AI FoundationsThe 6 foundations are Philosophy, Math, Psychology, CS, Neuroscience, Linguistics. Astrology is NOT a foundation (Q27 Sample Paper).

TRAP 7: Vanishing GradientClassic RNN problem for LONG sequences. Transformer solves this via self-attention (no sequential processing).

TRAP 8: Alpha value in Alpha-BetaAlpha = best score for MAXIMIZER. Beta = best score for MINIMIZER. Purpose = reduce nodes in Minimax.

1️⃣ Unit 1 — Foundations & Applications of AI

🧠 What is Intelligence? DEF

The ability to: acquire knowledge and skills, solve problems, and adapt to new situations.

Key Aspects: Reasoning, Problem-solving, Learning, Perception, Decision-making

🤖 What is Artificial Intelligence? EXAM DEF

Branch of computer science that develops machines to simulate human intelligence. Enables machines to think, learn, and solve complex human problems.

💡 AI = Making machines do what requires human intelligence

🏛 Foundations of AI MCQ — NOT ASTROLOGY

Philosophy — Logic and Reasoning
Mathematics — Probability, Statistics, Algorithms
Psychology — Human Behavior
Computer Engineering — Hardware/Software
Neuroscience — Brain Functioning
Linguistics — Language Processing

TRAP (Sample Q27): Astrology is NOT a foundation of AI. → Correct answer: Astrology is the odd one out.

📌 Types of AI HIGH PROB MCQ

Type	Definition	Example	Exists?
Narrow AI	Specific task only, predefined boundaries	Siri, Alexa, Netflix	✅ Yes
General AI	Any intellectual task a human can do	Humanoid robot (theoretical)	❌ Theory
Super AI	Surpasses humans in all domains	Sci-fi AI	❌ Theory

⚙️ Expert Systems DEF MCQ

AI that mimics human expertise using a Knowledge Base + Inference Engine. Applies IF-THEN rules.

Example: MYCIN (Medical Diagnosis). Uses certainty factors for handling uncertainty.

TRAP (Sample Q4): Certainty factors → Rule-based expert system (NOT Bayesian networks, NOT genetic algorithms).

🔑 Key AI Problems LIST

Knowledge representation
Reasoning and inference
Learning from data
Perception and understanding
Natural Language Processing (NLP)
Planning and decision-making
Handling uncertainty
Ethical and social challenges

📱 AI Applications MATCH TYPE

Domain	Example
NLP / Voice	Siri, Alexa, Cortana (Personal Voice Assistants)
Computer Vision	Facial recognition, object detection, self-driving cars
Healthcare	Medical imaging, MYCIN diagnosis system
Finance	Fraud detection, stock prediction
Gaming	Chess AI, game-playing agents

TRAP (Sample Q22): Siri, Alexa, Cortana = Personal voice assistants (NOT machine learning algorithms, NOT web search engines).

📜 Production Systems DEF Q3 Sample

A system of rules for problem-solving. Components: Working Memory + Rule Base (IF-THEN) + Inference Engine.

Chemical Synthesis: Monotonic and Not Partially Commutative (Sample Q32 → Answer d).

🛠 Modern AI Toolkits

TensorFlow — Google; static computation graphs; production-ready
PyTorch — Meta (Facebook); dynamic graphs; research-friendly
scikit-learn — Classical ML in Python (SVM, KNN, Decision Trees)

2️⃣ Unit 2 — Problem Solving & Search Algorithms

🗺 State Space & Problem Formulation CORE

AI problem = Initial State + Goal State + Actions + Path Cost. A problem in AI is defined by an initial and goal state.

Branching factor = number of states generated from a given state A* evaluation: f(n) = g(n) + h(n) g(n) = actual cost from start → current node h(n) = heuristic estimate from current node → goal h*(n) = TRUE optimal cost from current node → goal

TRAP (Sample Q11): Branching factor = number of different states generated FROM a given state. NOT paths to goal, NOT heuristic values, NOT depth.

TRAP (Sample Q60): A problem in AI = "A task defined by an initial and goal state" → Answer (b).

📏 Search Performance Measures

Completeness — Always finds solution if one exists?
Optimality — Guaranteed to find best solution?
Time Complexity — How many nodes expanded?
Space Complexity — Max nodes in memory?

🔍 Uninformed Search Comparison TABLE MCQ

Algorithm	Strategy	Complete	Optimal	Time	Space	Structure
BFS	Level-by-level	✅ Yes	✅ Equal cost	O(b^d)	O(b^d)	Queue (FIFO)
DFS	Deepest first	❌ No	❌ No	O(b^m)	O(bm)	Stack (LIFO)
DLS	DFS + limit L	❌ If L<d	❌ No	O(b^L)	O(bL)	Stack
IDS	DLS + increase L	✅ Yes	✅ Yes	O(b^d)	O(bd)	Stack

TRAP (Sample Q52): DFS preferred over BFS when → memory constraints are strict → Answer (b).

TRAP (Sample Q56): Suitable for infinite state spaces → Iterative Deepening → Answer (c).

🎯 Informed Search MOST TESTED

Greedy Best-First Search

f(n) = h(n) ONLY
Expands closest-to-goal node
Fast but NOT optimal
Can get stuck in loops

A* Search

f(n) = g(n) + h(n)
Optimal IF h is admissible
Complete (finite space)
Best-first + cost-aware

A* Example (Sample Paper): f(B) = g(B) + h(B) = 9 + 4 = 13 f(D) = g(D) + h(D) = 8 + 6 = 14 → Expand B (lowest f value first) 8-Puzzle (Sample Q33): h(n)=5, g(n)=3 → f(n)=8 ← Answer (a)

KEY (Sample Q14, Q44): A* considers BOTH cost and heuristic. GBFS considers ONLY heuristic. A* guarantees optimal (admissible h). GBFS does NOT.

🔺 Admissibility & Heuristics EXAM KEY

Admissible: h(n) ≤ h*(n) [never overestimates] Consistent: h(n) ≤ cost(n,n') + h(n') [triangle inequality] Manhattan distance = |x1-x2| + |y1-y2| [8-puzzle heuristic] Sample Q39: h*(n) = Actual cost from node n to the goal → Answer (b)

🎮 Minimax & Alpha-Beta GAME AI

Minimax: MAX maximizes evaluation, MIN minimizes. Evaluates ALL game states — complete tree exploration.
Alpha-Beta Pruning: Optimization of Minimax. Skips branches that cannot affect the optimal decision.
Alpha (α): Best possible score for the MAXIMIZER along current path.
Beta (β): Best possible score for the MINIMIZER along current path.

TRAP (Sample Q37, Q53): Main purpose Alpha-Beta = Reduce nodes evaluated in Minimax. Minimax evaluates ALL states; Alpha-Beta SKIPS unnecessary branches. Alpha-Beta is an OPTIMIZATION, not a replacement.

⛰ Local Search MCQ

Hill Climbing: Greedy — always moves to better neighbor. Problem: gets stuck at local maxima.
Random Restart: Restarts from multiple random initial states → escapes local maxima (Sample Q49).
Simulated Annealing: Accepts WORSE solutions with some probability. Why? To escape local minima and explore more solution space (Sample Q40 → Answer b).

💧 Water Jug Problem NUMERICAL

Max states formula: (cap1+1) × (cap2+1) Example: 4L + 3L jugs = (4+1)×(3+1) = 5×4 = 20 states ← (Sample Q10 Answer d) Sample Q17: J1=9L, J2=5L, State=(6,2) Step 1: Fill J2 → (6, 5) Step 2: Pour J2→J1: J1 can take 9-6=3L, pour 3L from J2 Result: J1=9, J2=5-3=2 → State = (9,2) ← Answer (b)

📐 Logic & De Morgan's Law LOGIC MCQ

De Morgan's (Sample Q34 → Answer a): ¬(P∨Q) = (¬P)∧(¬Q) ¬(P∧Q) = (¬P)∨(¬Q) Implication: P→Q ≡ ¬P∨Q Biconditional: P↔Q ≡ (P→Q)∧(Q→P) NOT valid connector (Sample Q54): Aggregation ← Answer (c)

3️⃣ Unit 3 — Machine Learning & Probabilistic Reasoning

📊 ML Types Comparison TABLE MCQ

Type	Training Data	Goal	Algorithms	Example
Supervised	Labeled (X, Y)	Predict output	SVM, KNN, LR, DT	Spam filter, image classifier
Unsupervised	Unlabeled (X only)	Find structure	K-Means, DBSCAN, PCA	Customer segmentation
Reinforcement	Rewards/Penalties	Max reward	Q-Learning, Policy Gradient	AlphaGo, game agents
Semi-Supervised	Labeled + Unlabeled	Both	Self-training	Text classification

TRAP (Sample Q51): "Static Learning" is NOT a category of ML — it is the odd one out. Answer: (d) Static Learning.

🎯 Confusion Matrix NUMERICAL — IN EVERY EXAM

Given (CA Paper Q1): TP=40, TN=30, FP=20, FN=10 Recall = TP / (TP + FN) = 40 / (40+10) = 40/50 = 0.80 Precision = TP / (TP + FP) = 40 / (40+20) = 40/60 = 0.667 Accuracy = (TP + TN) / Total = 70/100 = 0.70 F1 Score = 2 × (P × R) / (P + R) Specificity = TN / (TN + FP) Answer CA Q1: Recall=0.80, Precision=0.67 → Option B ✓ CA Q15 (Fraud detection): Misses fraud = LOW RECALL, Rarely flags normal = HIGH PRECISION → Answer: Low recall, high precision → Option (c) ✓

📐 Standard Deviation NUMERICAL

CA Q6: Data = {3, 7, 7, 7, 9} Mean μ = (3+7+7+7+9)/5 = 33/5 = 6.6 Variance = [(3-6.6)²+(7-6.6)²+(7-6.6)²+(7-6.6)²+(9-6.6)²] / 5 = [12.96 + 0.16 + 0.16 + 0.16 + 5.76] / 5 = 19.2 / 5 = 3.84 SD = √3.84 ≈ 1.96 ≈ 2 → Answer: A (2) ✓

🎲 Bayes Theorem NUMERICAL MCQ

P(A|B) = P(B|A) × P(A) / P(B) CA Q7: P(S)=0.3, P(H)=0.7, P(W|S)=0.8, P(W|H)=0.2 P(W) = P(W|S)×P(S) + P(W|H)×P(H) = 0.8×0.3 + 0.2×0.7 = 0.24+0.14 = 0.38 P(S|W) = P(W|S)×P(S) / P(W) = 0.24 / 0.38 ≈ 0.63 → Answer: B ✓

🌐 Bayesian Networks DEF MCQ

Directed Acyclic Graph (DAG) representing probabilistic relationships among variables. Joint probability = Product of conditional probabilities.

TRAP (Sample Q1): Bayesian networks incorporate UNCERTAINTY; rule-based systems rely on deterministic logic. → Answer: (b).

TRAP (CA Q4): Joint probability in Bayesian networks = Product of conditional probabilities (NOT sum, difference, or average). → Answer: (b).

🏷 Naive Bayes KEY ASSUMPTION

Key Assumption: All features are INDEPENDENT given the class label. This is the "naive" assumption.

TRAP (Sample Q16): Naive Bayes assumption = "All features are independent given the class" → Answer: (d). NOT correlated, NOT equal priors.

📍 KNN — K-Nearest Neighbors NUMERICAL

CA Q3: K=3, Points A(1,1)→Red, B(2,2)→Red, C(4,4)→Blue, Classify P(3,3) d(P,A) = √[(3-1)²+(3-1)²] = √8 ≈ 2.83 d(P,B) = √[(3-2)²+(3-2)²] = √2 ≈ 1.41 ← nearest d(P,C) = √[(3-4)²+(3-4)²] = √2 ≈ 1.41 ← nearest 3 nearest: B(Red), C(Blue), A(Red) → 2 Red, 1 Blue → P = Red → Answer: A ✓

📦 K-Means Clustering NUMERICAL EVERY CA

Steps: (1) Choose K centroids. (2) Assign each point to nearest centroid. (3) Update centroids = mean of cluster. (4) Repeat.

CA Part B Q1: k=2, Points A(2,2),B(3,3),C(4,4),D(6,5),E(9,1) Initial Centroids: C1=(2,2), C2=(9,1) Distances (Euclidean): A(2,2): d(C1)=0, d(C2)=√50=7.07 → Cluster1 B(3,3): d(C1)=√2=1.41, d(C2)=√37=6.08 → Cluster1 C(4,4): d(C1)=√8=2.83, d(C2)=√26=5.10 → Cluster1 D(6,5): d(C1)=5, d(C2)=√17=4.12 → Cluster2 E(9,1): d(C1)=√50=7.07, d(C2)=0 → Cluster2 New C1 = ((2+3+4)/3, (2+3+4)/3) = (3, 3) New C2 = ((6+9)/2, (5+1)/2) = (7.5, 3)

✂️ Cross-Validation & Overfitting MCQ

Cross-Validation: Used to REDUCE OVERFITTING (CA Q8 → Answer a). Splits data into k folds.

Overfitting (CA Q9): Good training performance + poor test performance → "Overfitting problem" → Answer (b).

Best Model (CA Q10): Low bias + Low variance → Answer (c).

🎓 Reinforcement Learning — MDP DEF

States (S) — all possible situations the agent can be in
Actions (A) — possible actions at each state
Transition Probability P(s'|s,a) — probability of reaching next state
Reward Function R(s,a) — immediate reward signal
Discount Factor γ — 0 to 1; γ≈1 = future important; γ≈0 = immediate rewards only

4️⃣ Unit 4 — Deep Learning & Natural Language Processing

🧠 Neural Network Basics

Inspired by the human brain. Perceptron: Simplest ANN — single neuron, binary classification, LINEAR classifier (draws one straight line).

Perceptron output = activation(Σ wᵢxᵢ + b) Weight update rule: w = w - η·x (η = learning rate) Limitation: Can't solve non-linear problems (XOR needs MLP)

MLP (Multi-Layer Perceptron): Multiple hidden layers. Solves non-linear problems. Universal Function Approximator. Hidden layers = feature extractors.

🖼 CNN — Convolutional Neural Network MOST TESTED

Designed for grid-like data (images). Preserves spatial relationships between pixels.

Architecture: Input → [Conv → ReLU → Pool] × N → Flatten → FC → Softmax → Output

Convolution Layer: Applies kernels/filters to detect features. Produces feature maps. Kernel slides with stride.
Pooling Layer: Reduces spatial dimensions. Max Pooling: Keeps largest value in each window. Goal: keep important features, reduce computation.
Fully Connected (FC) Layer: Converts features to probability outputs for final classification.

TRAP (CA Q2): Best for image classification = CNN (spatial features). NOT RNN or Transformer or Bayesian.

TRAP (CA Q14): Role of Pooling = Reduce spatial dimensions of feature maps. NOT increase filters, NOT convert to probabilities alone.

🔄 RNN — Recurrent Neural Network COMPARE MCQ

Handles sequential/time-series data. Has a LOOP — output of step t becomes input of step t+1. Hidden state h_t holds memory.

h_t = tanh(W_x·X_t + W_h·h_{t-1} + b) New Memory = Activation(New Input + Previous Memory) Applications: Text, Speech, Stock Prices, Time Series

TRAP (CA Q11): Common RNN problem = Vanishing gradient during LONG sequences. NOT lack of activation functions, NOT spatial resolution issues.

⚡ Transformer Architecture HIGH YIELD

Google Brain 2017. Solves RNN's sequential bottleneck. Reads entire sequence AT ONCE (parallel).

Self-Attention: Q=Query (What I'm looking for?) K=Key (What I can offer?) V=Value (Actual content) Attention(Q,K,V) = softmax(QK^T / √d_k) × V Positional Encoding: Sine/Cosine waves → adds word order info Why needed? Parallel processing → loses word order → must add it back CA Q12: Why positional encoding? → Provide ordered sequence data → Answer (a) ✓

Encoder: Positional Encoding → Multi-Head Attention → Add&Norm → FFN → Rich vector representation.

Decoder: Masked Attention (can't see future words) → Cross-Attention (looks at Encoder output) → Generate output.

Multi-Head Attention: Multiple attention mechanisms in parallel — one head focuses grammar, another vocabulary, another context.

🤖 BERT vs GPT COMPARISON — EXAM FAVOURITE

Feature	BERT	GPT
Direction	Bidirectional (L ↔ R)	Unidirectional (L → R only)
Architecture	Encoder-only	Decoder-only
Training Task	MLM (mask 15%) + NSP	Autoregressive (predict next token)
Best Use	Understanding: QA, Classification, NER, Sentiment	Generation: Text, Code, Dialogue, Summarization
Context Window	Sees past AND future simultaneously	Sees only past tokens (future masked)

TRAP: BERT reads L↔R simultaneously. GPT reads L→R only. GPT GENERATES new text; search engines RETRIEVE existing pages (Sample Q43 → Answer a).

🗣 NLP Core Concepts PHASES MCQ

NLP Pipeline for Chatbot (Sample Q31 → Answer c): Speech Recognition → Text Preprocessing → Intent Recognition → Response Generation Ambiguity types: - Syntactic: "I saw the man with the telescope" (structural) - Semantic: "Bank" = river bank OR financial bank (word meaning)

Tokenization: Splitting text into tokens (words, subwords, characters).

Embedding: Dense vector representation in continuous high-dimensional space. Captures semantic similarity.

Attention: Model focuses on most relevant input parts dynamically. Solves fixed-length bottleneck of Seq2Seq models.

🤝 Chatbots MCQ

Rule-Based Chatbots

Predefined scripts
Keyword matching
Fixed responses
Limited flexibility

AI-Based Chatbots

ML-powered NLP
Understands context
Improves over time
Example: ChatGPT

NLU: Extracts intent + entities from user input ("Book flight to Delhi Friday" → intent: book, entities: Delhi, Friday).

NLG: Converts structured data to human-like response.

NER in Voice Assistants (Sample Q38): Use NLP techniques like NER and Sentiment Analysis to improve Alexa/Siri. → Answer (b).

NLP for spell checking (Sample Q48): Edit Distance → Answer (a). NOT TF-IDF, NOT Tokenization.

5️⃣ Unit 5 — Generative AI & Prompt Engineering ★ Inferred from syllabus + sample papers

🎨 Generative AI — Definition KEY MCQ

AI that creates NEW content (text, images, audio, video, code) by detecting patterns in training data.

KEY (Sample Q19): GenAI CREATES new data; Traditional AI ANALYZES existing data → Answer (b).

KEY (Sample Q45): GenAI "learns" to create → By detecting patterns in training data (NOT memorizing past data) → Answer (b).

🔀 GANs — Generative Adversarial Networks ARCHITECTURE MCQ

Generator

Creates synthetic/fake data
Input: random noise vector z
Goal: fool the Discriminator
GENERATES new samples

Discriminator

Judges real vs fake
Binary classifier
Feedback to Generator
Output: probability real/fake

TRAP (Sample Q7): Component generating NEW samples = Generator (NOT Encoder, Discriminator, or Decoder) → Answer: (d) Generator.

Training: Generator improves until Discriminator can't distinguish real from fake. Applications: Image generation, deepfakes, data augmentation, art creation

📊 VAE — Variational Autoencoders Sample Q41

Generate new data by learning a probabilistic latent space. Encoder maps to distribution (μ, σ²). Sample z. Decoder reconstructs.

Primary Purpose (Sample Q41): Generate new data by learning a probabilistic latent space → Answer: (b).

VAE Loss = Reconstruction Loss + KL Divergence KL Divergence: forces latent space toward N(0,1) → enables random sampling

🔤 Large Language Models (LLMs) MODERN AI

Transformer-based models trained on massive text corpora. Demonstrate emergent abilities at scale.

PaLM vs smaller models (Sample Q46): PaLM has more parameters → understands complex queries better → Answer (c).

Emergent abilities: Zero-shot translation, code writing, chain-of-thought reasoning — not explicitly trained for these tasks.

💬 Prompt Engineering DEFINITIONS

What is a Prompt? (Sample Q30): An input given to a language model → Answer (b).

Best strategy to refine AI responses (Sample Q2): Rewriting and iterating prompts → Answer (a).

Pattern	Description	Example
Instruction-based	Direct command	"Translate this to French:"
Chain-of-Thought	Step-by-step reasoning	"Solve step by step..."
Few-shot	Examples provided first	"Here are 3 examples. Now:"
Role-based	Assign persona	"You are a doctor..."
Zero-shot	No examples given	"Classify this text:"

TRAP (Sample Q58): All three — instruction-based, chain-of-thought, few-shot — ARE prompt patterns → Answer (d) All of the above.

🧭 Zero-Shot vs Few-Shot Learning TRICKY MCQ

TRAP (Sample Q13): Zero-shot = Predict UNSEEN tasks directly (no task-specific training). NOT train only on labeled data. NOT ignore new input. NOT evaluate only known classes → Answer (c).

Few-shot: Learn from a small number of examples (1-shot, 5-shot). Model generalizes from few demonstrations.

🎭 Diffusion Models MODERN

Generate content by reversing a noise process. Forward: add noise step by step. Reverse: denoise to generate.

Examples: DALL-E, Stable Diffusion, Midjourney, Sora (video).

Forward process: x_0 → x_1 → ... → x_T (pure Gaussian noise) Reverse process: x_T → x_{T-1} → ... → x_0 (generated image/content) Model learns: how to remove noise at each step

⚖️ AI Ethics & Responsible AI CO5

Bias: Models inherit biases from training data (gender, racial bias)
Fairness: Equal treatment across demographic groups
Transparency/Explainability (XAI): Understanding why AI makes decisions
Privacy: User data protection; not using training data without consent
Accountability: Humans remain responsible for AI decisions
Hallucination: LLMs confidently generate false information

6️⃣ Unit 6 — Data Analysis, Tools & MLOps ★ Inferred from syllabus + sample papers

🧪 ChatGPT ADA (Advanced Data Analysis) TOOLS MCQ

Key difference (Sample Q50 → Answer c): ADA can analyze datasets and run Python code; standard ChatGPT cannot.

Error identification in ADA (Sample Q18 → Answer a): Using data validation methods to identify discrepancies or unusual patterns in input data.

Without error handling (Sample Q55 → Answer d): Output might include incorrect or incomplete data.

Media files in ADA (Sample Q35 → Answer a): Extract text and metadata using AI models.

📊 Data Visualization MCQ

Categorical data (Sample Q12 → Answer c): Bar chart.

Data Type	Best Chart
Categorical	Bar chart, Pie chart
Distribution	Histogram
Two-variable relationship	Scatter plot
Correlation matrix	Heatmap
Time series / Trend	Line chart

🐍 Python Pandas Quick Reference CODE MCQ

Duplicate rows → df.drop_duplicates() (Sample Q21 → Answer d) Missing values → df.fillna(value) Preview data → df.head() Count non-null → df.count() ZeroDivisionError → dividing by zero (Sample Q15 → Answer b) ValueError → wrong type/value NameError → undefined variable IndexError → index out of range

AI writing assistant (Sample Q23 → Answer b): Grammarly.

Marketing/Social Media AI (Sample Q59 → Answer d): ChatGPT.

☁️ Cloud AI & MLOps DEF

MLOps: ML + DevOps — automates training, deployment, monitoring of ML models
Cloud Services: AWS SageMaker, GCP Vertex AI, Azure ML — for scalable training/hosting
Edge Deployment: Run models on local devices (low latency, offline capable)
CI/CD Pipelines: Continuous integration/delivery for ML workflows
Model Monitoring: Track accuracy drift, data drift over time in production

📋 Knowledge Representation & Reasoning MCQ

Knowledge Representation approaches (Sample Q26 → Answer a): Frames, Semantic Networks, Logical Representation. Neural networks are NOT a KR approach.

Odd one out (Sample Q25 → Answer a): Heuristic Search is NOT a knowledge representation technique.

Backward chaining advantage (Sample Q42 → Answer b): It is goal-driven (starts from goal, works backward to find supporting facts).

Dempster-Shafer (Sample Q47 → Answer a): Belief functions = degree of confidence in a hypothesis.

📋 Cheat Sheets

🔍 Search Algorithms

BFSQueue, Level-by-level, O(b^d), Complete, Optimal (equal cost)

DFSStack, Depth-first, O(b^m), Incomplete, Not Optimal

IDSDLS+increasing L, Complete, Optimal, O(b^d), O(bd) space

Greedyf(n)=h(n) only, Fast, NOT Optimal

A*f(n)=g(n)+h(n), Complete, Optimal (admissible h)

MinimaxMAX maximizes, MIN minimizes, evaluates ALL states

Alpha-BetaPrunes Minimax, α=best MAX, β=best MIN

Sim. Anneal.Accepts worse solutions to escape local minima

📊 ML Metrics

PrecisionTP/(TP+FP) — of predicted +ve, how many correct?

RecallTP/(TP+FN) — of actual +ve, how many caught?

Accuracy(TP+TN)/(TP+TN+FP+FN)

F1 Score2×P×R/(P+R) — harmonic mean

Low RecallMisses many positives (high FN)

Low PrecisionMany false alarms (high FP)

Best ModelLow bias + Low variance

Cross-valUsed to REDUCE OVERFITTING

🤖 Neural Networks

PerceptronSingle neuron, linear classifier, binary output

MLPHidden layers, solves XOR, Universal Approx.

CNNImages, spatial features, Conv+Pool+FC

RNNSequences, hidden state h_t, vanishing gradient

LSTMSolves vanishing gradient, forget/input/output gates

TransformerParallel, self-attention (Q,K,V), positional enc.

BERTBidirectional, MLM+NSP, understanding tasks

GPTAutoregressive, next-token pred., generation tasks

🎨 Generative AI

GANGenerator (creates) + Discriminator (judges)

VAEEncoder→latent distribution→Decoder, probabilistic

DiffusionAdd noise (forward) → Remove noise (reverse)

Zero-shotPredict UNSEEN tasks without specific training

Few-shotLearn from small number of examples

PromptInput given to a language model

CoTChain-of-thought: step-by-step reasoning

LLM ScalingMore params + more data → emergent abilities

🧮 Probability & Logic

BayesP(A|B) = P(B|A)·P(A) / P(B)

Joint PProduct of conditional probabilities (Bayesian nets)

Naive BayesFeatures INDEPENDENT given class

Bayesian NetDAG, probabilistic relationships, uncertainty

De Morgan¬(P∨Q)=(¬P)∧(¬Q)

ImplicationP→Q ≡ ¬P∨Q

D-S TheoryBelief = degree of confidence in hypothesis

💻 Python/Tools

Drop dupesdf.drop_duplicates()

Fill NAdf.fillna(value)

Div by zeroZeroDivisionError (NOT ValueError)

CategoricalBar chart visualization

NLP spellcheckEdit Distance (NOT TF-IDF)

Writing AIGrammarly

Marketing AIChatGPT

ADAChatGPT ADA can run Python code on datasets

🌙 One-Night-Before Exam — 20 Must-Know Points

1. AI foundations: Philosophy, Math, Psychology, CS, Neuroscience, Linguistics — NOT Astrology

2. Narrow AI=Siri/Alexa (EXISTS). General AI=theoretical. Super AI=theoretical.

3. BFS=Queue/Complete/Optimal | DFS=Stack/Incomplete/Not Optimal | A*=g(n)+h(n)/Optimal

4. Greedy BFS uses ONLY h(n). A* uses g(n)+h(n). A* is optimal (admissible h). Greedy is NOT.

5. Alpha-Beta goal = Reduce nodes in Minimax. α=best MAX score. β=best MIN score.

6. Branching factor = states generated FROM a given state (NOT paths to goal).

7. Precision=TP/(TP+FP) | Recall=TP/(TP+FN) | f(n)=g(n)+h(n) for A*

8. Naive Bayes assumption = features INDEPENDENT given class (NOT correlated).

9. Cross-validation = REDUCE OVERFITTING. Overfitting = good train, bad test performance.

10. CNN=images/spatial features. RNN=sequences/memory h_t. Transformer=parallel/attention.

11. RNN problem = Vanishing gradient during long sequences.

12. Positional encoding = needed because Transformer reads ALL at once (loses word order).

13. BERT=Bidirectional/MLM+NSP/Encoder-only. GPT=Autoregressive/Next-token/Decoder-only.

14. GAN: Generator GENERATES samples. Discriminator judges real/fake. Don't confuse!

15. VAE primary purpose = generate new data via probabilistic latent space.

16. Zero-shot = predict UNSEEN tasks directly. Few-shot = learn from few examples.

17. Prompt = input to language model. All 3 patterns (instruction, CoT, few-shot) are valid.

18. NLP chatbot pipeline: Speech Recog → Text Preprocessing → Intent Recog → Response Gen.

19. ChatGPT ADA can analyze datasets + run Python; standard ChatGPT cannot.

20. De Morgan: ¬(P∨Q)=(¬P)∧(¬Q). NOT a logical connector: Aggregation.

🔢 Complete Formula & Definition Sheet

A* f(n)

f(n) = g(n) + h(n)

Admissible h

h(n) ≤ h*(n) [never overestimates]

Manhattan Dist

|x1-x2| + |y1-y2|

Euclidean Dist

√((x1-x2)² + (y1-y2)²)

Precision

TP / (TP + FP)

Recall

TP / (TP + FN)

Accuracy

(TP + TN) / (TP + TN + FP + FN)

F1 Score

2 × (P × R) / (P + R)

Bayes Theorem

P(A|B) = P(B|A)×P(A) / P(B)

Total Probability

P(B) = Σ P(B|Aᵢ)×P(Aᵢ)

Mean (μ)

μ = Σxᵢ / N

Variance

σ² = Σ(xᵢ - μ)² / N

Std Deviation

σ = √(Σ(xᵢ-μ)² / N)

BFS Time+Space

O(b^d) where b=branching factor, d=depth

DFS Time

O(b^m) where m=max depth

DFS Space

O(b×m)

Water Jug States

(cap1 + 1) × (cap2 + 1)

Perceptron

output = activation(Σwᵢxᵢ + b)

Weight Update

w = w - η·x (η = learning rate)

RNN Hidden State

h_t = tanh(W_x·X_t + W_h·h_{t-1} + b)

Self-Attention

Attention(Q,K,V) = softmax(QK^T/√d_k)·V

Positional Enc.

PE(pos,2i) = sin(pos/10000^(2i/d_model))

De Morgan 1

¬(P∨Q) = (¬P)∧(¬Q)

De Morgan 2

¬(P∧Q) = (¬P)∨(¬Q)

Implication

P→Q ≡ ¬P∨Q

Biconditional

P↔Q ≡ (P→Q)∧(Q→P)

K-Means Centroid

New centroid = mean of all points in cluster

RL Bellman Eq.

V(s) = R + γ × max_a V(s')

Discount γ≈1

Future rewards are important

Discount γ≈0

Only immediate rewards matter

🏆 Top 100 Most Important Points

⚖️ Master Comparison Tables

CNN vs RNN vs Transformer

Feature	CNN	RNN	Transformer
Best for	Images, Grid data	Sequences, Time series, Text	NLP, Any sequence
Processing	Parallel (spatial)	Sequential (one step at a time)	Fully Parallel
Memory	No temporal memory	Hidden state h_t	Self-attention (global context)
Key Layer	Conv + Pooling	Recurrent connection	Multi-Head Attention
Vanishing Grad	Less problematic	Major issue (long sequences)	Not an issue
Match (CA Q5)	Image processing	Sequence modeling	Parallel processing
Also handles	Object detection	Speech, Stock prices	Translation, Generation

BERT vs GPT

Feature	BERT	GPT
Direction	Bidirectional (L ↔ R)	Unidirectional (L → R)
Architecture	Encoder-only	Decoder-only
Training Task	MLM (mask 15%) + NSP	Autoregressive next-token prediction
Best Use	Understanding: QA, Classification, NER, Sentiment	Generation: Text, Code, Dialogue
Context	Sees past AND future simultaneously	Sees only past tokens
Examples	BERT-base, RoBERTa, DistilBERT	GPT-2, GPT-3, GPT-4, ChatGPT

Supervised vs Unsupervised vs Reinforcement Learning

Feature	Supervised	Unsupervised	Reinforcement
Training Data	Labeled (X, Y pairs)	Unlabeled (X only)	Rewards/Penalties
Goal	Predict labels/values	Find patterns/clusters	Maximize cumulative reward
Algorithms	SVM, KNN, Linear Reg, DT	K-Means, DBSCAN, PCA	Q-Learning, Policy Gradient
Example	Email spam filter, Image classifier	Customer segmentation	AlphaGo, Game agents
Feedback	Direct (label comparison)	No direct feedback	Delayed reward signal

Search Algorithms Master Reference

Algorithm	Type	Complete	Optimal	Time	Space
BFS	Uninformed	✅ Yes	✅ Equal cost	O(b^d)	O(b^d)
DFS	Uninformed	❌ No	❌ No	O(b^m)	O(bm)
IDS	Uninformed	✅ Yes	✅ Yes	O(b^d)	O(bd)
Greedy BFS	Informed	❌ No	❌ No	O(b^m)	O(b^m)
A*	Informed	✅ Yes	✅ (admissible h)	Exp.	Exp.
Hill Climbing	Local	❌ No	❌ No	—	O(b)
Sim. Annealing	Local	✅ Prob.	✅ Prob.	—	O(1)

Narrow vs General vs Super AI

Feature	Narrow AI (Weak)	General AI (Strong)	Super AI
Scope	One specific task	Any human intellectual task	Surpasses humans in ALL
Learning	Task-specific only	Cross-domain learning	Self-improvement
Examples	Siri, Alexa, ChatGPT, Netflix	Humanoid robot (theoretical)	Sci-fi AI (theoretical)
Status	✅ EXISTS NOW	❌ Theoretical	❌ Theoretical

🃏 Flashcards — Click Card to Flip

Loading...

1 / 40

❓ MCQ Bank (105+ Questions)

📝 Mock Test — 30 Questions / 30 Minutes

📋 Instructions

30 MCQs simulating actual INT428 exam pattern
1 mark each, 0.25 negative marking per wrong answer
Timer: 30 minutes. Covers all 6 units.
Submit at any time or let timer run out