M.

Tech Class Notes Explained(Module 1)

Subject: Advanced Machine Learning & Deep Learning

Topic: Introduction to Machine Learning

1. Introduction to Machine Learning

 ML is a subfield of AI where algorithms learn from data to make

predictions or decisions without being explicitly programmed for every
task.

 Learning types:

o Supervised Learning: Learn from labeled data.

o Unsupervised Learning: Discover structure from unlabeled

data.

o Reinforcement Learning (RL): Learn from feedback (rewards).

o Semi-supervised: Combines small labeled + large unlabeled

data.

o Self-supervised: Learns representations by creating labels from

input data itself.

2. Training

 The process of adjusting a model's internal parameters to minimize the

prediction error on training data.

 Datasets:

o Training set: Used to fit the model.

o Validation set: Used to fine-tune model hyperparameters.

o Test set: Used for final performance evaluation.

3. Rote Learning

 A primitive form of learning where the system memorizes inputs and

their outputs.
  It doesn’t generalize, i.e., it can’t predict for unseen inputs.

4. Learning Concepts

 A concept is a function from input space to output labels (e.g.,

whether an email is spam).

 A concept class is a set of all possible functions (hypotheses) that can

define this mapping.

 Learning = Finding the best approximation of the true concept from the
concept class.

5. General-to-Specific Ordering

 Hypotheses are structured in a hierarchy:

o General hypothesis: Matches more examples (may include

incorrect ones).

o Specific hypothesis: Matches fewer, more accurate examples.

 This ordering helps efficiently search for the correct hypothesis in

algorithms like Candidate Elimination.

6. Version Spaces

 A version space is the subset of hypotheses that are still consistent

with all observed training examples.

 Defined by:

o S: Specific boundary → Most specific consistent hypotheses.

o G: General boundary → Most general consistent hypotheses.

 As training data increases, the space narrows.

7. Candidate Elimination Algorithm

 It updates S and G iteratively:

o Positive examples → make S more general.

 o Negative examples → make G more specific.

 Learning continues until S and G converge (the exact target concept is

found) or they diverge (inconsistent data).

8. Inductive Bias

 Every ML algorithm has assumptions that help it generalize beyond the

training data.

 Without bias, learning is impossible (many hypotheses can explain the

data equally well).

 Examples:

o k-NN: Nearby points are likely to share labels.

o Decision Trees: Favors simpler trees (Occam's Razor).

9. Decision Tree Induction

 A tree structure where each internal node splits data by an attribute

and each leaf represents a class label.

 Built using:

o Information gain (entropy reduction)

o Gini index (impurity measure)

 Overfitting avoided by pruning unnecessary branches.

10. Overfitting

 A model that fits the training data too closely, including noise, loses its
ability to generalize.

 Solutions:

o Cross-validation: Robust performance estimate.

o Regularization: Penalizes complex models.

o Early stopping: Stops training before overfitting.

 o Data augmentation: Increase data variety.

11. Nearest Neighbor Algorithm (k-NN)

 Stores all training data. For a new input:

o Calculate distance to all points.

o Pick the top-k closest and vote on label.

 Pros: Simple, no training needed.

 Cons: High computation during testing; sensitive to irrelevant features.

12. Learning Neural Networks

 A neural network is a collection of layers of interconnected nodes

(neurons).

 Backpropagation: Main training algorithm; calculates gradients and

updates weights.

 Activation functions:

o Sigmoid: Smooth but saturates.

o Tanh: Centered at 0.

o ReLU: Fast, sparse activation.

 Challenges: Deep nets suffer from vanishing/exploding gradients.

13. Supervised Learning

 Labeled data = inputs + correct outputs.

 Algorithms:

o Linear regression: Predicts continuous values.

o Logistic regression: Binary classification.

o SVM: Maximize margin between classes.

o Neural nets: Flexible, handles nonlinearity.

  Evaluation:

o Accuracy, Precision, Recall, F1, ROC-AUC.

14. Unsupervised Learning

 Only input data; no labels.

 Goals:

o Find hidden structure (clusters, patterns).

 Methods:

o Clustering: Group similar data.

o Dimensionality Reduction: Simplify features.

 Applications: Market segmentation, anomaly detection, compression.

15. Reinforcement Learning (RL)

 Learning by interacting with an environment.

 The agent takes actions → receives rewards → learns optimal strategy

(policy).

 Key terms:

o Policy (π): Maps states to actions.

o Value function (V): Expected future rewards.

o Q-value (Q): Expected reward for state-action pair.

 Algorithms:

o Q-learning: Learn value of actions directly.

o Actor-Critic: Learn both value and policy.

 Applications: Game AI (AlphaGo), robotics, self-driving.

📊 Summary Table Explained

Concept Why It Matters

Rote vs. Concept

Shows importance of generalization
Learning

General-Specific
Organizes hypothesis search efficiently
Ordering

Version Space Narrows down consistent hypotheses

Inductive Bias Drives meaningful generalization

Real-world models must generalize, not

Overfitting
memorize

Supervised vs.
Determines learning strategy
Unsupervised

Reinforcement
Learns from feedback, not labels
Learning

📌 Suggested Diagrams

1. Version Space: Visual of narrowing between S and G.

2. Decision Tree: Split on attributes and class labels at leaves.

3. k-NN: Graph showing test point and neighbors.

4. Neural Net: Layered diagram with activations.

5. RL Agent: Feedback loop with environment and reward arrows.