Computer network

02 December 2024 22:38

In machine learning, learning paradigms define how algorithms learn from data to make predictions
or decisions, with the three main paradigms being supervised, unsupervised, and reinforcement
learning.

Here's a breakdown of each paradigm:

1. Supervised Learning:
• Definition:
In supervised learning, algorithms learn from a labeled dataset, meaning the input data has
corresponding known outputs or labels.
• Goal:
The goal is to train a model that can accurately predict the output for new, unseen inputs
based on the patterns learned from the labeled data.
• Examples:
○ Classification: Predicting whether an email is spam or not spam.
○ Regression: Predicting the price of a house based on its features.
• How it works:
The algorithm learns a mapping function from the input to the output, allowing it to make
predictions on new data.
•
2. Unsupervised Learning:
• Definition:
Unsupervised learning deals with unlabeled data, where the goal is to discover patterns,
structures, or insights without explicit guidance.
• Goal:
The goal is to find hidden structures, clusters, or relationships within the data.
• Examples:
○ Clustering: Grouping customers based on their purchasing behavior.
○ Dimensionality Reduction: Reducing the number of features in a dataset while
preserving important information.
• How it works:
The algorithm identifies patterns or structures in the data without being told what to look for.
•
3. Reinforcement Learning:
• Definition:
Reinforcement learning involves training an agent to make decisions in an environment to
maximize a reward.
• Goal:
The goal is to find the optimal strategy or policy for the agent to take actions that lead to the
highest cumulative reward.
• Examples:
○ Game Playing: Training an AI to play chess or Go.
○ Robotics: Controlling a robot to navigate an environment.
• How it works:
The agent interacts with the environment, receives feedback (rewards or penalties), and learns
to adjust its actions to achieve the desired outcome.
•
Other Machine Learning Paradigms:
• Transfer Learning: Leveraging knowledge learned from one task to improve performance on a
related task.
•
• Zero-Shot Learning: Recognizing and classifying objects or tasks without having seen any examples

Quick Notes Page 1

• Zero-Shot Learning: Recognizing and classifying objects or tasks without having seen any examples
during training.
•
• Semi-Supervised Learning: Using a combination of labeled and unlabeled data for training.
•
• Multi-Task Learning: Training a model to perform multiple tasks simultaneously.
•
• Meta-Learning: Learning how to learn, enabling models to adapt quickly to new tasks.

In machine learning, a hypothesis space is the set of all possible functions or models a learning
algorithm can consider. A finite hypothesis space contains a limited, countable number of
hypotheses, while an infinite hypothesis space contains an unlimited, uncountable number.

Finite Hypothesis Space:

• Definition:
A finite hypothesis space is a set of hypotheses where the number of possible functions is
limited and can be enumerated.
• Examples:
○ A set of decision trees with a fixed maximum depth.
○ A set of boolean functions with a fixed number of input variables.
○ A set of rules with a specific structure.
• Advantages:
○ Easier to search for the best hypothesis.
○ Can be learned efficiently.
• Disadvantages:
○ May not be able to represent complex relationships in the data.
○ Can be overly simplistic.
○
Infinite Hypothesis Space:
• Definition:
An infinite hypothesis space is a set of hypotheses where the number of possible functions is
unlimited and cannot be enumerated.
• Examples:
○ All possible linear functions.
○ All possible neural networks.
○ All possible polynomials.
• Advantages:
○ Can represent complex relationships in the data.
○ Can be more flexible.
• Disadvantages:
○ Harder to search for the best hypothesis.
○ Can be computationally expensive to learn.
○ Can overfit to the training data

PAC Learning (Probably Approximately Correct Learning) is a theoretical framework in machine

learning that helps us understand how well a learning algorithm can learn a function from a set of
examples. It was introduced by Leslie Valiant in 1984.
Let’s break it down in simple terms:

What does PAC mean?

• Probably: We allow a small probability that the algorithm might not perform well. This is
usually denoted by δ (delta).
• Approximately Correct: We allow the algorithm to make some errors. This is controlled by ε
(epsilon).
In essence, the algorithm is "probably (1 - δ) approximately (within error ε) correct."

Quick Notes Page 2

In essence, the algorithm is "probably (1 - δ) approximately (within error ε) correct."

✅ Key Concepts
1. Hypothesis Space (H): Set of all functions (hypotheses) that the algorithm can choose from.
2. Target Concept (c): The actual function we are trying to learn (unknown to the learner).
3. Learning Algorithm (A): Takes training examples and outputs a hypothesis h∈Hh \in Hh∈H.
4. Sample Complexity: How many examples are needed to learn a good hypothesis with high
probability.

PAC Learnable Definition

A concept class CCC is PAC-learnable if:
• There exists a learning algorithm AAA,
• That for any ϵ>0\epsilon > 0ϵ>0 and δ>0\delta > 0δ>0,
• With high probability (1−δ)(1 - \delta)(1−δ),
• Outputs a hypothesis hhh such that error ≤ϵ\leq \epsilon≤ϵ,
• In time and sample size polynomial in 1ϵ\frac{1}{\epsilon}ϵ1, 1δ\frac{1}{\delta}δ1, and the size
of the input.

Error Types
• True error (Generalization error): How well the hypothesis performs on unseen data.
• Empirical error (Training error): How well it performs on the training data.
PAC learning wants to minimize the true error.

Example
Imagine you’re trying to classify emails as "spam" or "not spam."
• Target function (c): The real rule for spam.
• Hypothesis space (H): All the rules your algorithm can choose from.
• PAC learning goal: Find a rule that is almost correct (say, 95% accuracy), with high confidence
(say, 99%).
So if ε = 0.05 and δ = 0.01, your algorithm should find a hypothesis that:
• Has at most 5% error
• With at least 99% probability

Ensemble Learning is a machine learning technique where multiple models (called base learners)
are combined to solve a problem and improve performance.
Think of it like this:
“Instead of relying on one person’s opinion (single model), you ask a group of people (ensemble
of models) and take a majority decision — it’s usually more accurate!”

Why use Ensemble Learning?

• Single models can be weak or make errors.
• Combining multiple models helps to:
○ Reduce variance (less overfitting)
○ Reduce bias (learn better)
○ Improve accuracy and robustness

Main Types of Ensemble Methods

1. Bagging (Bootstrap Aggregating)
• Train several models in parallel on different random subsets of the data (with replacement).
• Final output is usually the majority vote (for classification) or average (for regression).
• Reduces variance.
Famous example: Random Forest

Quick Notes Page 3

 Famous example: Random Forest
• Uses many decision trees trained on different parts of the data.

2. Boosting
• Models are trained sequentially.
• Each new model focuses more on the mistakes made by the previous ones.
• Final output is a weighted combination of all models.
• Reduces bias.
Popular boosting algorithms:
• AdaBoost
• Gradient Boosting
• XGBoost, LightGBM, CatBoost

3. Stacking (Stacked Generalization)

• Combines different types of models (e.g., decision tree, SVM, logistic regression).
• Their outputs are used as input features for a meta-model that learns how to best combine
them.
• More powerful but also more complex.

Simple Example:
Let’s say we have three models trying to classify if a fruit is an apple or not:
• Model A: says apple
• Model B: says not apple
• Model C: says apple
Bagging or Voting Ensemble would go with the majority vote → apple.
Boosting would say: “Hey Model B is weak, let’s train a new model that fixes its mistakes.”
Stacking would take outputs of A, B, and C and train a new model (like logistic regression) to decide
the final answer.

Bayesian learning in machine learning uses Bayes' theorem to incorporate prior knowledge and data
to estimate the probability of hypotheses, providing a framework for probabilistic predictions and
uncertainty quantification.

Here's a more detailed explanation:

Key Concepts:
• Bayes' Theorem:
The foundation of Bayesian learning, it allows updating beliefs (prior probabilities) in light of
new evidence (data) to arrive at a posterior probability.

• Formula: P(A|B) = (P(B|A) * P(A)) / P(B)

•
○ P(A|B): Posterior probability of event A given event B
○ P(B|A): Likelihood of event B given event A
○ P(A): Prior probability of event A
○ P(B): Evidence or probability of event B
• Prior Knowledge:
Bayesian methods allow incorporating existing knowledge or beliefs about a problem before
observing data.
•
• Posterior Probability:
The updated probability of a hypothesis after considering the data.
•
• Probabilistic Predictions:

Quick Notes Page 4

• Probabilistic Predictions:
Bayesian models provide not just a single prediction, but a probability distribution over possible
outcomes, allowing for uncertainty quantification.
•
• Bayesian Networks:
A type of probabilistic graphical model that represents a joint probability distribution of a set of
variables, using directed acyclic graphs to capture dependencies between them.
•
• Naive Bayes:
A simple probabilistic classifier based on Bayes' theorem, assuming that the presence of a particular
feature is independent of the presence of other features.
•
Advantages of Bayesian Learning:
• Uncertainty Quantification:
Bayesian models can estimate the uncertainty in predictions, which is crucial in applications
where understanding the reliability of predictions is important.
•
• Incorporating Prior Knowledge:
Bayesian methods can leverage prior knowledge or beliefs to improve the performance of machine
learning models, especially when data is limited.
•
• Handling Missing Data:
Bayesian methods can effectively handle missing data by incorporating prior information about the
missing values.
•
• Regularization:
Bayesian models can naturally incorporate regularization, which helps prevent overfitting and
improves generalization performance.
•
• Online Updating:
Bayesian models can be updated incrementally with new data, allowing for continuous learning and
adaptation.
•
Applications:
• Classification:
Naive Bayes is a popular algorithm for text classification and other classification tasks.
•
• Regression:
Bayesian regression models can provide probabilistic predictions for regression problems.
•
• Machine Learning Model Selection:
Bayesian methods can be used to select the best machine learning model for a given task.
•
• Uncertainty Estimation:
Bayesian models are used in various applications where uncertainty quantification is crucial, such as
medical diagnosis, financial modeling, and risk assessment.
•
• Deep Learning:
Bayesian methods can be used to create Bayesian deep learning models that can provide uncertainty
estimates for predictions.

Quick Notes Page 5