General expressions

Question 1

What is a GAN?

Answer 1

A generative model with two neural networks — a Generator (creates fake data) and a Discriminator (detects fake vs real) — trained in competition to produce realistic synthetic data (e.g., images).

Question 2

Key AWS-relevant idea about GANs?

Answer 2

Used for synthetic data generation, image creation, and data augmentation in ML workflows.

Question 3

What is a VAE?

Answer 3

A generative model that encodes input data into a probability distribution (latent space) and then decodes samples from that distribution to generate new data.

Question 4

GAN vs VAE difference?

Answer 4

VAEs optimize probabilistic reconstruction (more stable training), while GANs use adversarial training (often sharper outputs).

Question 5

What is a Diffusion Model?

Answer 5

A generative model that gradually adds noise to data and then learns to reverse the noise process to generate high-quality synthetic outputs.

Question 6

Why diffusion model is important?

Answer 6

Powers modern image generators (e.g., Stable Diffusion-type models). Known for stability and high-quality outputs.

Question 7

What is RAG?

Answer 7

A technique that combines retrieval systems (e.g., vector databases) with LLMs to fetch relevant documents before generating a response.

Question 8

Why use RAG?

Answer 8

Improves accuracy, reduces hallucinations, and enables models to use up-to-date or private enterprise data.

Question 9

What is BERT?

Answer 9

BERT (Bidirectional Encoder Representations from Transformers) is a transformer-based language model that understands context by reading text bidirectionally (left and right).

Question 10

What is BERT mainly used for?

Answer 10

Text classification, sentiment analysis, question answering, and named entity recognition.

Question 11

What is ROUGE?

Answer 11

A metric used to evaluate text summarization by measuring overlap between generated text and reference summaries (recall-focused).

Question 12

What does ROUGE measure?

Answer 12

N-gram overlap, longest common subsequence, and recall similarity.

Question 13

What is BLEU?

Answer 13

A metric used to evaluate machine translation by measuring precision-based n-gram overlap between generated and reference text.

Question 14

BLEU vs ROUGE difference?

Answer 14

BLEU focuses on precision (common in translation). ROUGE focuses on recall (common in summarization).

Question 15

What is SVD?

Answer 15

A matrix factorization technique that decomposes a matrix into three matrices: A = U Σ V^T

Question 16

Why is SVD important in ML?

Answer 16

Used for dimensionality reduction, noise reduction, and recommendation systems.

Question 17

Where is SVD commonly applied in NLP?

Answer 17

In Latent Semantic Analysis (LSA).

Question 18

What is Word2Vec?

Answer 18

A neural network model that learns dense vector representations (embeddings) of words based on context.

Question 19

Two Word2Vec architectures?

Answer 19

• CBOW (predict word from context) * Skip-gram (predict context from word)

Question 20

Word2Vec Key idea?

Answer 20

Words appearing in similar contexts have similar vector representations.

Question 21

What is PCA?

Answer 21

A dimensionality reduction technique that transforms data into new orthogonal axes (principal components) that maximize variance.

Question 22

Core idea formula:

Answer 22

Maximize Variance = λ

Question 23

Why use PCA?

Answer 23

Reduce features while retaining most important information.

Question 24

What is LSA?

Answer 24

An NLP technique that uses SVD on a term-document matrix to uncover hidden (latent) relationships between words and documents.

Question 25

Why is LSA useful?

Answer 25

Improves semantic search and topic modeling by reducing noise and dimensionality.

Question 26

What is an SVM?

Answer 26

A supervised learning algorithm that finds the optimal hyperplane separating classes with maximum margin.

Question 27

Key idea?

Answer 27

Maximize the margin between support vectors and decision boundary.

Question 28

What allows SVMs to handle non-linear data?

Answer 28

The kernel trick (e.g., RBF kernel).

Question 29

What is Logistic Regression?

Answer 29

A classification algorithm that predicts probabilities using the sigmoid function.

Question 30

Sigmoid function:

Answer 30

f(x)=rac{1}{1+e^{-x}}

Question 31

Output range?

Answer 31

Between 0 and 1 (probability).

Question 32

Used for?

Answer 32

Binary classification problems.

Question 33

What is a Decision Tree?

Answer 33

A supervised learning model that splits data based on feature values to make predictions.

Question 34

How does it decide splits?

Answer 34

Using metrics like Gini impurity or Information Gain.

Question 35

Advantage?

Answer 35

Easy to interpret and explain.

Question 36

What is a vector in ML?

Answer 36

An ordered list of numbers representing data in multi-dimensional space.

Question 37

Why are vectors important?

Answer 37

ML models operate on numerical vector representations.

Question 38

Common operation?

Answer 38

Dot product (measures similarity).

Question 39

What are embeddings?

Answer 39

Dense vector representations of objects (text, images, users, etc.) that capture semantic meaning.

Question 40

Why are embeddings important in generative AI?

Answer 40

Enable similarity search, clustering, and Retrieval-Augmented Generation (RAG).

Question 41

Example use case?

Answer 41

Converting text into vectors for vector databases.

Question 42

Batch Size

Answer 42

The number of training samples processed before the model updates its weights.

Question 43

Small vs Large batch size?

Answer 43

* Small → noisier updates, better generalization * Large → faster training, more stable gradients

Question 44

Top-K Sampling

Answer 44

A text generation method that selects the next token from the **K most probable tokens**.

Question 45

Effect of smaller K?

Answer 45

More deterministic output. Larger K → more diversity.

Question 46

Top-P (Nucleus Sampling)

Answer 46

Selects the smallest set of tokens whose cumulative probability ≥ P, then samples from them.

Question 47

Why use Top-P instead of Top-K?

Answer 47

It dynamically adjusts the candidate pool based on probability distribution.

Question 48

Temperature

Answer 48

A parameter that controls randomness in token selection.

Question 49

Low vs High temperature?

Answer 49

* Low (e.g., 0.1) → deterministic, focused * High (e.g., 1.0+) → creative, diverse

Question 50

Confusion Matrix

Answer 50

A table used to evaluate classification performance by comparing predicted vs actual values.

Question 51

What are the 4 components?

Answer 51

* True Positive (TP) * True Negative (TN) * False Positive (FP) * False Negative (FN)

Question 52

Why important?

Answer 52

Helps compute accuracy, precision, recall, and F1-score.

Question 53

Root Mean Squared Error (RMSE)

Answer 53

A regression metric measuring the square root of average squared prediction errors.

Question 54

Formula:

Answer 54

RMSE = \sqrt{\frac{1}{n}\sum (y_i - \hat{y}_i)^2}

Question 55

Why square errors?

Answer 55

Penalizes large errors more heavily.

Question 56

Mean Absolute Error (MAE)

Answer 56

A regression metric measuring the average absolute difference between predictions and actual values.

Question 57

Formula:

Answer 57

MAE = \frac{1}{n}\sum |y_i - \hat{y}_i|

Question 58

RMSE vs MAE difference?

Answer 58

RMSE penalizes large errors more; MAE treats all errors equally.

Question 59

Correlation Matrix

Answer 59

A table showing correlation coefficients between multiple variables.

Question 60

Correlation coefficient range?

Answer 60

From -1 to +1. * +1 → perfect positive correlation * -1 → perfect negative correlation * 0 → no linear correlation

Question 61

Shapley Values

Answer 61

A model explainability method that assigns each feature a contribution value to a specific prediction.

Question 62

Why important?

Answer 62

Provides fair, mathematically grounded feature importance.

Question 63

Used in?

Answer 63

Model interpretability for complex models (e.g., tree ensembles, neural networks).

Question 64

Partial Dependence Plots (PDP)

Answer 64

A visualization showing the marginal effect of one feature on model predictions.

Question 65

Why useful?

Answer 65

Helps understand how a feature influences output while averaging out other features.