Support Vector Machines (SVM)
Machine learning method mainly used for classification.
Tries to separate data into two classes using a decision boundary (hyperplane).
goal: find the boundary that maximizes the margin (distance) between the classes.
Only the data points closest to the boundary (support vectors) influence the final separating line.
For data that cannot be separated linearly, SVM uses the kernel trick to map data into a higher-dimensional space where separation becomes possible.
Common applications: image classification, fraud detection, medical diagnosis, and finance.
Comparison of Learning Types
Supervised Learning
Unsupervised Learning
Reinforcement Learning
data and labels
labeled training data
unlabeled training data
no initial training data
way of learning
using labeled dataset
using unlabeled dataset
interacting with environment
type of problem
regression and classification
clustering, dimensionality reduction
exploitation or exploration
learning objective
learn a mapping function to predict the correct output with a high level of accuracy
identify underlying patterns or similarities in the dataset
learn and make decisions in a sequential way based on the feedback from the environment
Artificial Neural Networks (ANN)
Inspired by the structure and functioning of the human brain.
Consist of interconnected nodes (neurons) that learn patterns from data.
Common Applications: speech recognition, language processing, diagnostics
Artificial Neuron Structure
Inputs are multiplied by weights.
Weighted inputs are summed.
Activation function transforms this result into an output.
Output is passed to other neurons.
Activation Functions (General Idea)
Introduce non-linearity.
Help the network learn complex patterns.
Different activation functions suit different tasks.
ANN Architecture
Input layer → Hidden layer(s) → Output layer.
Hidden layers extract and transform patterns.
Multiple hidden layers = deep learning.
Types of Neural Networks
Feedforward Networks: Information flows in one direction from input to output. Used for classification and regression.
Convolutional Neural Networks (CNNs): Good for images and visual pattern recognition.
Recurrent Neural Networks (RNNs): Designed for sequential or time-based data, remember previous information using loops.
Includes Long Short-Term Memory (LSTM) and grated Recurrent Units (GRU) cells for longer memory.
Self-Organizing Maps (SOM): Used for visualization and dimensionality reduction.
Backpropagation and Training
Training process adjusts weights to reduce prediction errors.
Forward pass: calculate output.
Compare output with target to compute error.
Backward pass: adjust weights using gradients to minimize the error.
Gradient descent repeatedly updates weights until errors are minimized.
Problems:
Vanishing gradients: early layers learn too slowly.
Exploding gradients: weights become unstable.
The model (agent) interacts with an environment.
Learns by trial and error to maximize cumulative rewards.
Uses a policy to decide optimal actions.
Markov Decision Process
states: situations or conditions, in which the agent finds itself
actions: choices that are available at each state
rewards: received by the agent
policy: guides agent’s decision-making
value: total reward achieved by following this policy
Genetic Algorithms
Inspired by natural selection.
Start with a population of possible solutions.
Selection, crossover, and mutation are used to evolve better solutions.
Useful for optimization problems and feature selection.
Fuzzy Logic
Deals with reasoning where information may be uncertain or vague.
Uses degrees of truth instead of strict true/false decisions.
Useful when data is imprecise.
Naïve Bayes Classification
Based on Bayes’ theorem.
Assumes independence between features (simplifying assumption).
Fast, simple, widely used for text and document classification.
Calculates the probability that a new item belongs to each class and picks the most likely one.
Benefits of ANN
Benefits:
Can model complex nonlinear relationships.
Learn and adapt over time.
Can process many inputs in parallel.
Robust even if some parts are damaged.
Zuletzt geändertvor einem Monat