AI & Machine Learning Interview Questions

Fundamental Concepts

1. What is the difference between Artificial Intelligence, Machine Learning, and Deep Learning?

Answer:

• Artificial Intelligence (AI): Broad field of creating intelligent machines that can simulate human thinking and behavior.

• Machine Learning (ML): Subset of AI that uses algorithms to learn patterns from data without being explicitly programmed.

• Deep Learning (DL): Subset of ML that uses neural networks with many layers to learn complex patterns from large amounts of data.

Analogy: AI is the goal of creating intelligent machines, ML is the approach to achieve it, and DL is a powerful technique within ML.

2. What are the main types of Machine Learning? Explain each.

Answer:

• Supervised Learning: Model learns from labeled data (input-output pairs). Used for prediction and classification.

  • Examples: Spam detection, house price prediction

• Unsupervised Learning: Model finds patterns in unlabeled data. Used for clustering and association.

  • Examples: Customer segmentation, anomaly detection

• Reinforcement Learning: Model learns through trial and error using rewards and punishments.

  • Examples: Game playing AI, robotics

3. What is the difference between Classification and Regression?

Answer:

• Classification: Predicts discrete/categorical outcomes (classes/labels).

  • Examples: Spam/Not spam, Dog/Cat, Positive/Negative sentiment

• Regression: Predicts continuous numerical values.

  • Examples: House prices, Temperature, Stock prices

4. What is Overfitting and Underfitting? How do you prevent them?

Answer:

• Overfitting: Model learns training data too well, including noise, and performs poorly on new data. (High variance)

• Underfitting: Model fails to capture underlying patterns in data. (High bias)

Prevention:

• For Overfitting:

  • Cross-validation

  • Regularization (L1/L2)

  • Early stopping

  • Reduce model complexity

  • Get more training data

• For Underfitting:

  • Increase model complexity

  • Add more features

  • Reduce regularization

  • Train for longer

5. What is Cross-Validation and why is it important?

Answer: Cross-validation is a technique to evaluate model performance by splitting data into multiple folds.

Common approach - k-Fold Cross Validation:

1. Split data into k equal parts

2. Use k-1 parts for training, 1 part for testing

3. Repeat k times, using each fold as test set once

4. Average the results

Importance:

• More reliable performance estimate

• Better use of limited data

• Reduces overfitting risk

Technical Concepts

6. What is the Bias-Variance Tradeoff?

Answer:

• Bias: Error from overly simplistic assumptions. High bias causes underfitting.

• Variance: Error from sensitivity to small fluctuations in training data. High variance causes overfitting.

Tradeoff: As model complexity increases:

• Bias decreases

• Variance increases
  The goal is to find the right balance.

7. Explain the Confusion Matrix and its metrics.

Answer: A table showing model performance on classification problems.

text

              Predicted: No   Predicted: Yes
Actual: No     True Negative    False Positive
Actual: Yes    False Negative   True Positive

Key Metrics:

• Accuracy: (TP + TN) / Total

• Precision: TP / (TP + FP) - How accurate positive predictions are

• Recall/Sensitivity: TP / (TP + FN) - How well we find actual positives

• F1-Score: Harmonic mean of precision and recall

8. What is the difference between L1 and L2 Regularization?

Answer:

• L1 Regularization (Lasso): Adds absolute value of coefficients to loss function. Can shrink coefficients to exactly zero (feature selection).

• L2 Regularization (Ridge): Adds squared value of coefficients to loss function. Shrinks coefficients but rarely to zero.

When to use:

• L1: When you suspect many irrelevant features

• L2: When all features might be relevant

9. What is Gradient Descent and how does it work?

Answer: Optimization algorithm used to minimize the loss function by iteratively moving towards the minimum.

Process:

1. Initialize parameters randomly

2. Calculate gradient (derivative) of loss function

3. Update parameters: θ = θ - learning_rate * gradient

4. Repeat until convergence

Types:

• Batch GD: Uses entire dataset for each update

• Stochastic GD: Uses one sample per update

• Mini-batch GD: Uses small batches of data

10. What is the difference between Bagging and Boosting?

Answer:

• Bagging (Bootstrap Aggregating):

  • Trains multiple models independently on random subsets of data

  • Reduces variance

  • Example: Random Forest

  • Models run in parallel

• Boosting:

  • Trains models sequentially, each correcting the previous one

  • Reduces bias

  • Examples: AdaBoost, Gradient Boosting, XGBoost

  • Models run sequentially

Practical Implementation

11. How do you handle missing values in a dataset?

Answer:

• Numerical data:

  • Mean/Median imputation

  • Regression imputation

  • K-NN imputation

• Categorical data:

  • Mode imputation

  • Create "Missing" category

• Other approaches:

  • Delete rows/columns (if small percentage missing)

  • Use algorithms that handle missing values (like XGBoost)

12. What is Feature Engineering and why is it important?

Answer: Process of creating new features or transforming existing ones to improve model performance.

Common techniques:

• Encoding: One-hot encoding, label encoding

• Scaling: Standardization, Normalization

• Creation: Polynomial features, interaction terms

• Transformation: Log transform, binning

Importance: Better features often beat better algorithms. Good features can make simple models perform exceptionally well.

13. What is One-Hot Encoding and when do you use it?

Answer: Technique to convert categorical variables into binary (0/1) columns.

Example: Colors: Red, Green, Blue becomes:

• Red: [1, 0, 0]

• Green: [0, 1, 0]

• Blue: [0, 0, 1]

When to use:

• For nominal categorical data (no order)

• When categories have no inherent ranking

• Before feeding to algorithms that expect numerical input

14. Explain Train-Test Split and why it's used.

Answer: Splitting dataset into two parts:

• Training set: Used to train the model (typically 70-80%)

• Testing set: Used to evaluate final model performance (typically 20-30%)

Why important:

• Evaluates how well model generalizes to unseen data

• Prevents overfitting

• Provides realistic performance estimate

python

from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

15. What is the purpose of Activation Functions in Neural Networks?

Answer: Activation functions introduce non-linearity into the network, allowing it to learn complex patterns.

Common activation functions:

• Sigmoid: Output between 0-1, used for binary classification

• Tanh: Output between -1 to 1, zero-centered

• ReLU: Most popular, solves vanishing gradient problem

• Softmax: Used in output layer for multi-class classification

Advanced Concepts

16. What is Transfer Learning and when would you use it?

Answer: Using a pre-trained model on a new related task.

Process:

1. Take model trained on large dataset (e.g., ImageNet)

2. Remove final layers

3. Add new layers for your specific task

4. Fine-tune on your smaller dataset

When to use:

• Limited labeled data

• Limited computational resources

• Similar domain to pre-trained model

17. What are Convolutional Neural Networks (CNNs) and where are they used?

Answer: Specialized neural networks for processing grid-like data (images, videos).

Key components:

• Convolutional layers: Extract features using filters

• Pooling layers: Reduce dimensionality

• Fully connected layers: Final classification

Applications:

• Image classification

• Object detection

• Medical imaging

• Self-driving cars

18. What is Natural Language Processing (NLP) and common techniques?

Answer: Field focused on interaction between computers and human language.

Common techniques:

• Tokenization: Splitting text into words/tokens

• TF-IDF: Term frequency-inverse document frequency

• Word Embeddings: Word2Vec, GloVe, FastText

• Transformers: BERT, GPT (state-of-the-art)

19. What is the difference between Supervised and Unsupervised Learning in practice?

Answer:

20. How would you approach a new Machine Learning problem?

Answer: Systematic approach:

1. Problem Understanding:

   • Define business objective

   • Identify success metrics

2. Data Collection & Exploration:

   • Gather relevant data

   • Exploratory Data Analysis (EDA)

3. Data Preprocessing:

   • Handle missing values

   • Feature engineering

   • Data splitting

4. Model Selection & Training:

   • Start with simple baselines

   • Try multiple algorithms

   • Hyperparameter tuning

5. Evaluation & Deployment:

   • Evaluate on test set

   • Deploy model

   • Monitor performance

Bonus: Scenario-Based Questions

21. How would you explain a complex ML model to a non-technical stakeholder?

Answer:

• Use analogies and simple language

• Focus on business impact, not technical details

• Show concrete examples of predictions

• Explain in terms of risk, cost, or benefit

• Use visualizations when possible

22. What metrics would you use for an imbalanced classification problem?

Answer: Accuracy is misleading for imbalanced data. Better metrics:

• Precision-Recall curve

• F1-Score

• AUC-ROC curve

• Cohen's Kappa

• Confusion matrix analysis

These questions cover the fundamental knowledge expected from freshers in AI/ML interviews. Focus on understanding the concepts deeply rather than just memorizing answers!