Introduction to Machine Learning

• Machine learning involves computers learning without explicit programming

Defining Machine Learning

• Machine learning consists of computational methods leveraging experience to enhance performance and make accurate predictions
• The success of a learning algorithm relies on the data, linking machine learning to data analysis and statistics

AI, ML, NN, DL Overlap

• Artificial Intelligence (AI) encompasses techniques enabling machines to perform tasks like humans
• Machine Learning (ML) refers to algorithms allowing computers to learn from examples without explicit programming
• Artificial Neural Networks (ANN) are brain-inspired machine learning models
• Deep Learning (DL) represents a subset of ML utilizing deep artificial neural networks for modeling and data representation hierarchy

Situations Requiring Machine Learning

• Tasks too complex to program are suitable for machine learning
• Analysis of extensive and intricate datasets benefits from machine learning
• Machine learning tools adapt to input data, unlike rigid programmed tools

Common Machine Learning Applications

• Healthcare: Predicting diseases and medical imaging
• Transportation: Self-driving cars, traffic prediction, and ride-sharing optimization
• Finance: Trading and fraud detection
• Retail and e-commerce: Customer segmentation and chatbots
• Social Media: Content recommendation and feeds
• Education: Personalized learning and predicting failure
• Environmental protection: Deforestation and wildlife monitoring, and energy optimization

Problems Tackled by Machine Learning

• Tackles Text/document classification
• Tackles Natural language processing
• Tackles Speech processing
• Tackles Computer vision applications
• Tackles Fraud detection, playing games, remaining useful life
• Machine learning finds practical application in an expanding array of areas

Key Figures and Developments in AI History

• Arthur Samuel coined the term "machine learning" in 1959, his checkers playing program demonstrated that machine can improve autonomously.
• 1943: Development of the artifical neuron
• 1950: The Turing Test
• 2012: AlexNet
• 2016: IBM Deep Blue
• 2020: AlphaFold
• 2022: ChatGPT

Reasons For Recent Growth of Machine Learning

• Machine learning's recent progress are attributed to Data Volume, Data Storage and Computing powers

Core Machine Learning Terminologies

• Example: A data item used for learning or evaluation
• Features: Input variables in a machine learning model
• Feature vector: The set of features associated with an example
• Labels: Values or categories assigned to examples
• Hypothesis set: Functions mapping features to labels
• Training: Process to determine ideal parameters
• Trained model: The best hypothesis found during training
• Loss function (cost function): Measures the difference between predicted and true labels
• Hyperparameter: Free parameters specified as learning algorithm inputs vs parameters learned from data

Training/Validation/Test Samples

• Training sample: A set of examples used to train the learning algorithm
• Validation Sample: Examples used to choose the appropriate values while learning, such as alpha
• Test sample: Separate set of examples used to evaluate the learning algorithm's performance

Types of Machine Learning

Supervised Learning:

• The model learns from a labeled dataset with known target or outcome variables, each paired with the correct output
• It divides into classification and regression

Supervised Learning: Classification

• Aims to assign a category or categorical label to each item and sorts data points into predefined classes

Supervised Learning: Regression

• Used to predict a continuous numerical output based on input features, modeling relationships between variables

Unsupervised Learning

• Involves training models on unlabeled data to find patterns, structures, and relationships without explicit supervision
• Divides into clustering, association, and dimensionality reduction

Unsupervised Learning: Clustering

• Explores raw, unlabeled data, grouping it by similarities or differences, discovering natural groups in uncategorized data

Unsupervised Learning: Dimensionality Reduction

• Involves unsupervised learning techniques to reduces the number of features, or dimensions, in a dataset
• Applications include feature engineering, image compression, data visualization, etc.

Semi-Supervised Learning

• Blends supervised and unsupervised learning, using both labeled and unlabeled data
• Involves training a model with labeled data, predicting labels for unlabeled data, and retraining with the combined dataset, thereby lessening manual labeling

Reinforcement Learning

• Reinforcement Learning focuses on learning from consequences in a trial-and-error process

Key Machine Learning Challenges

• Machine learning depends on selecting an algorithm and training data; challenges arise from "bad algorithms" and "bad data"

Insufficient Training Data

• Most Machine Learning algorithms require a lot of data to work properly

Nonrepresentative Training Data

• Training data needs to be both large and representative of the population
• Non-representative data inaccurately reflects the underlying population or true distribution, potentially creating, sampling bias and imbalanced class distribution
• Address this using random sampling and resampling, such as oversampling and undersampling

Poor Quality Data

• Concerns inaccurate, incomplete, inconsistent, or irrelevant data can hurt machine learning
• It is worth the effort to clean your data when you find: incorrect values, measurement errors, typos, empty fields, inconsistent units, multiple identical records or outliers

Irrelevant Features

• Relates to garbage in garbage out
• Input variables lack meaningful information, reducing model predictive power
• Critical part of machine learning is the coming up with a good set of features to train on which is called feature engineering: It includes, feature selection, feature extraction and creating new features by gathering new data
• Correlation analysis & Dimensionality reduction can identify irrelevant features

Overfitting and Underfitting Training Data

• Overfitting results when the model learns data including both noise and desired factors
• Underfitting results when the modesl oversimplifies the data and can't perform both in training and testing