UNIT 6 — Neural Networks (Topic-

by-topic, para→Q&A)
Topic A — What is a Neural Network? (one-line hook)

Neural network = a brain-inspired machine learning model made

of layers of nodes that learn from data.

MCQs (1-mark)

1. Q: A neural network is: (A) biological (B) brain-inspired ML

model.
A: (B).
2. Q: ANNs automatically extract ___ from data.
A: features.

Short (2–3m)
Q: Define neural network.
A: A neural network is a machine learning model that mimics
biological neurons; it processes inputs via interconnected nodes to
learn patterns and make predictions.

Long (5m)
Q: Explain significance and one real-world example.
A: Neural networks can learn complex, non-linear relationships
without hand-coded rules. They’re used in spam filters, image
tagging and search ranking (e.g., Google). Because they learn
features automatically, they handle messy, high-dimensional data
well. (Write 3 concise examples.)

Topic B — Parts of a Neural Network (Input / Hidden /

Output)
Memory hook: ―Ingredients → Chefs → Inspector‖ (input → hidden →
output).

MCQs

1. Q: Which layer holds features? A: Input layer.

2. Q: More than one hidden layer → ? A: Deep Neural Network.

Short (2–3m)
Q: Describe hidden layer role.
A: Hidden layers process inputs using neurons which apply weights
and activation functions; they transform raw features into higher-
level signals used for final prediction.

Long (5m)
Q: Explain all three layers and how data flows.
A: Input layer receives features. Hidden layers (one or many) apply
weighted sums + activation to produce intermediate
representations. Output layer maps these to final predictions/labels.
Data moves forward (forward propagation); errors are used later to
adjust weights (backpropagation). Add tiny diagram or sentence
showing w·x + b → activation.

Topic C — Components: Neuron, Weight, Bias, Activation,

Connections, Learning Rule

Hook: neuron = tiny calculator; weight = dial; bias = extra push;

activation = yes/no gate.

MCQs

1. Q: Weight shows ___ of connection. A: importance.

2. Q: Bias is added ___ activation. A: before.

Short (2–3m)
Q: Purpose of activation function?
A: Introduces non-linearity; decides if neuron ―fires‖ so network can
learn complex patterns beyond straight lines. Examples: Sigmoid,
ReLU, Tanh.

Long (5m)
Q: Explain backpropagation & learning rule (write steps).
A: 1) Forward pass: compute outputs. 2) Compute loss (predicted vs
actual). 3) Backpropagate error to compute gradients. 4) Update
weights using optimizer (e.g., gradient descent). This loop repeats
for epochs until loss reduces. Emphasize gradients and weight
updates.

Topic D — Working: Forward + Backprop (with tiny example)

Hook: ―Attempt → Mark → Fix → Repeat.‖

MCQs

1. Q: Forward propagation calculates ___. A:

activations/prediction.
2. Q: Backprop uses ___ to update weights. A: gradients.

Short (2–3m)
Q: Show formula for a neuron.
A: Sum = w1x1 + w2x2 + ... + bias; output = activation(sum). If
threshold-based, output =1 if sum≥threshold else 0.

Long (5m)
Q: Work a tiny numerical example (like in book: x1=2,x2=3,x3=1).
A: Write weighted sum, add bias, compare to threshold → decide
active/inactive; show how output passes to next layer. Then say: with
an error, backprop adjusts weights slightly to correct. (Use book’s
mini-example during memorization).

Topic E — Types of Neural Networks (Perceptron, FFNN, CNN,

RNN, GAN)
Hook: each NN is a specialist — perceptron = basic, CNN = vision, RNN
= memory, GAN = artist + critic.

MCQs

1. Q: Best for images? A: CNN.

2. Q: RNN is used for? A: sequential data (text, speech).

Short (2–3m)
Q: Perceptron vs FFNN.
A: Perceptron = single layer binary classifier. FFNN (MLP) = multi-
layer feedforward network, handles complex tasks via hidden layers.

Long (5m)
Q: Describe CNN and RNN with applications.
A: CNN uses convolutional filters to detect local patterns (edges →
textures → objects), ideal for image recognition, medical imaging.
RNN has loops to carry state across time steps, suitable for language
modeling, speech recognition, time-series forecasting. Mention GAN:
generator + discriminator competing to create realistic images.

Topic F — Impact & Ethics (short)

Short (2–3m)
Q: One societal impact and one risk.
A: Impact: automation & personalization (recommenders,
healthcare tools). Risk: bias, privacy issues, job shifts.

Unit-level ―Top exam Qs‖ (memorize these exact tiny

answers)

 Define NN. (1-liner)

 Parts of NN (i,ii,iii — para).
 Working of a neuron (formula + sentence).
 Explain backprop (steps).
 Types & one example each.
UNIT 8 — Data Storytelling
(Topic-by-topic, para→Q&A)
Topic A — What is a Story; why storytelling? (hook)

Stories stick — data + story = decisions made.

MCQs

1. Q: Data storytelling =? A: Communicating insights through

data + narrative.
2. Q: Three key elements? A: Data, Visuals, Narrative.

Short (2–3m)
Q: Why is data storytelling powerful?
A: It turns complex facts into memorable narratives, making
insights persuasive and actionable for non-technical audiences.

Long (5m)
Q: Steps to create a data story.
A: 1) Collect & clean data. 2) Choose visuals suited to data type. 3)
Find insight (trend/anomaly). 4) Build narrative arc (context →
evidence → insight → action). 5) Note ethics & limitations. Add short
example (survey→bar chart→recommendation).

Topic B — Elements: Data, Visuals, Narrative

Hook: think of a play — script (narrative), actors (data), stage set

(visuals).

MCQs

1. Q: Visuals help to? A: Convey complex info quickly.

 2. Q: Freytag’s Pyramid stages include? A: Exposition → Rising
action → Climax → Falling action → Resolution.

Short (2–3m)
Q: Role of narrative in a data story.
A: Narrative gives context and flow—it links charts to decisions and
explains ―why this matters‖, not just ―what happened.‖

Long (5m)
Q: Give examples of visuals & when to use them.
A: Line = trends over time; Bar = compare categories; Histogram =
distribution; Scatter = relationship; Map = geodata; Wordcloud =
text frequency. One-line justification each.

Topic C — Freytag’s Pyramid applied to data story

Short (2–3m)
Q: Briefly explain Freytag’s Pyramid for a data presentation.
A: Start with exposition (problem + data), build with rising action
(charts showing trends), reach climax (key insight), falling action
(implications) and finish with resolution
(recommendation/action).

Long (5m)
Q: Write a short data story using Freytag’s Pyramid (city traffic
case).
A: Intro: collect accident data; Rising: show hotspots via map &
heatmap; Climax: one intersection has 60% crashes; Falling: analyze
causes (visibility/time); Conclusion: propose speed cameras +
redesign. (Use as 5-mark answer.)

Topic D — Ethics in Data Storytelling

MCQ
Q: Key ethical rule? A: Accuracy and transparency.
Short (2–3m)
Q: List 3 ethical concerns.
A: Accuracy (don’t manipulate charts), transparency (cite sources &
limits), privacy (anonymize personal data).

Long (5m)
Q: How to ethically gather/share stories of people?
A: Get consent, anonymize identifiers, validate facts, avoid cherry-
picking, document limitations, and disclose methods/sources. Give a
tiny example (survey + consent form).

Unit ―Top exam Qs‖ (memorize short templates)

 Define data storytelling + 2 reasons it’s powerful.

 Name & explain 3 elements (data, visuals, narrative).
 Steps to make a data story (5 points).
 Freytag’s Pyramid (5 stages + tiny example).
 Ethics list and measures.

UNIT 7 — Generative AI (Topic-by-

topic, para→Q&A)
Topic A — What is Generative AI? (one-liner)

Generative AI creates new content (text, images, audio, video) by

learning patterns from datasets.

MCQs

1. Q: Primary objective? A: generate new data similar to training

samples.
2. Q: Examples? A: ChatGPT, DALL-E, Gemini.
Short (2–3m)
Q: Define Generative AI.
A: Generative AI uses deep learning models to learn data
distribution and produce novel, similar outputs (e.g., images,
stories, music).

Long (5m)
Q: Explain working in a paragraph (GANs & VAEs).
A: GANs pair a generator (makes fakes) with a discriminator
(judges real vs fake); through adversarial training, generator
improves. VAEs encode data into latent space and decode back,
learning smooth representations for sampling and anomaly
detection. Mention strengths: GANs → realism; VAEs → interpretable
latent space.

Topic B — Generative vs Discriminative models

MCQ
Q: Main difference? A: Generative models learn data distribution &
can create new samples; discriminative models learn boundaries to
classify.

Short (2–3m)
Q: Give 2 examples each.
A: Generative: GANs, VAEs, LLMs. Discriminative: Logistic Regression,
SVM, Random Forest.

Long (5m)
Q: When to use generative vs discriminative?
A: Use generative for content creation, data augmentation or
anomaly detection; discriminative for classification tasks where
accuracy and speed matter. Explain trade-offs: generative needs
more data/compute but gives richer outputs.

Topic C — Applications (image, text, audio, video)

MCQ
Q: Image generation tools? A: DALL-E, Stable Diffusion.

Short (2–3m)
Q: Text generation uses?
A: Content creation, drafting, code generation, Q&A (LLMs like
GPT/Gemini).

Long (5m)
Q: Give 4 applications and one risk for each.
A: Image gen (art/ads) — risk: copyright/deepfakes; Text gen
(chatbots/assistants) — risk: hallucinations; Audio gen
(music/voice) — risk: voice cloning; Video gen (ads/deepfakes) —
risk: misinformation. Offer mitigation: watermarking, source
citation, consent.

Topic D — LLMs & Transformers

MCQ
Q: LLMs are trained on? A: massive text datasets.

Short (2–3m)
Q: What is a transformer (1 line)?
A: Transformer is a neural architecture using self-attention to
model relationships across tokens, powering modern LLMs.

Long (5m)
Q: Describe an LLM with examples & limitations.
A: LLMs (e.g., GPT-4o, Gemini) are trained on trillions of words to
generate text, answer Qs, translate. Strengths: fluent text, versatility.
Limitations: high compute cost, can produce incorrect/confidently
wrong outputs, privacy & bias risks. Always verify outputs and cite
sources when used.

Topic E — Ethics & Social Implications

MCQ
Q: Deepfake tech threatens? A: authenticity & misinformation.

Short (2–3m)
Q: List main ethical concerns.
A: Deepfakes, bias/discrimination, plagiarism/copyright, lack of
transparency.

Long (5m)
Q: How should a company using generative AI (case: ―Creative
Horizons‖) act?
A: Use transparent labeling of AI content; check copyright & license
for training data; run bias audits and human review; add
watermarks; get consent for personal data; maintain provenance
logs. (This is the go-to answer in case-study Qs).

Unit ―Top exam Qs‖ (memorize)

 Define Generative AI + 2 examples.

 Explain GANs (generator + discriminator).
 What is a VAE (encoder/latent/decoder).
 LLM: features, examples, risks.
 Ethics measures (watermarking, consent, audits).