Machine Learning 1 - Week 1 Lecture

Machine Learning 1 - Week 1 Lecture

• The lecture covers course introduction, supervised machine learning, and an overview of the course.
• The suggested textbooks are:
  • An Introduction to Statistical Learning by Gareth James, Daniela Witten, Trevor Hastie, Robert Tibshirani, and Jonathan Taylor
  • Introduction to Machine Learning with Python by Andreas C. Müller & Sarah Guido
• Course resources include a syllabus, course site, course outline, and a data camp.
• Supervised machine learning is a type of machine learning where the algorithm is trained on labelled data.
• The algorithm learns from labelled data to predict future outcomes.
• Examples of supervised machine learning problems include:
  • Predicting if an image is a tree.
  • Predicting customer churn.
  • Predicting if a unit will fail.
  • Predicting fraudulent transactions.
  • Classifying diseases based on X-rays.
• Machine learning differs from classical programming as rules are learned from data instead of specified in the code.
• Machine learning is crucial for building algorithms that adapt and improve through experience rather than relying on explicitly programmed rules
• Machine learning attempts to discover patterns in data to create models that can be used to predict outputs or behaviors for new data or inputs.

What is Supervised Machine Learning and Why Use It?

• Supervised machine learning is a process where an algorithm is trained on labelled data.
• The algorithm learns the relationship between input and output variables.
• The goal is to build a model that can accurately predict the output for new data points.

Learning From Data

• Machine learning algorithms discover patterns to build models, making iterative adjustments.
• The resulting model can be a decision boundary.
• It is different from classical programming where rules are already programmed and the data has direct answers.

Biologically Inspired

• Neural networks are biologically inspired algorithms.
• These networks have input, hidden, and output layers.
• They work by adjusting weights to improve their predictions based on training data.

Learning by Trial and Error

• The concept is similar to how children learn through trial and error.
• Learning by trial and error involves adjusting actions based on a reward system to improve outcomes over time.

Machine Learning

• Machine learning, as defined by Arthur Samuel (1959), is the ability of computers to learn without explicit programming.
• The field automates building models from data, adjusting parameters to improve prediction accuracy with time.
• Machine learning is defined as learning from experience as described by Herb Simon (1978), where learning programs improve based on experience.
• Tom Mitchell (1997) defines machine learning as learning a task T with performance measure P, improving with experience E to predict outcomes from inputs.

Task T, Performance Measure P, and Training Experience E

• Examples are:
  • task (T) - recognizing the numbers 0-9
  • performance measure (P) - how accurately the algorithm recognizes numbers it hasn't seen
  • experience E - a range of examples of the numbers 0 -9, which are used by an algorithm to accurately classify numbers between 0-9

Artificial Intelligence

• Alan Turing introduced the imitation game and the Turing Test, a method of determining if a machine can exhibit intelligent behavior.
• There exists a correlation between Artificial Intelligence and Machine Learning.
• Artificial Intelligence is broader and encompasses machine learning but it is not specific or limited to it.

The Turing Test

• The Turing Test is a measure of a machine's ability to exhibit intelligent behavior.
• A human evaluator engages in conversations with both a human and a machine without knowing which is which.
• If the evaluator cannot reliably distinguish the machine from the human, the machine is said to have passed the test.

Why Now?

• Advancements in computing power (e.g., GPUs) and data availability have made machine learning more practical.
• Increased accessibility of computing resources (cloud services like AWS, Azure, and Google Cloud) has allowed more people to train machine learning models easily.
• Progresses in algorithms, theories, and tools to handle more complex tasks.

Supervised Learning: Regression and Classification

• Supervised learning aims at predicting continuous or discrete values.
• Regression tasks involve predicting continuous values.
• Classification tasks involve predicting discrete values (categories or classes).

Categories of Machine Learning

• Supervised: algorithms get paired data to predict outcomes.
• Unsupervised: no labels, so algorithms find hidden patterns.
• Reinforcement: algorithms learn through rewards and penalties, changing actions to maximize rewards over time.

Unsupervised Learning: Clustering

• Finding hidden patterns in data without any initial labels or categories.

Reinforcement Learning

• Learning through interactions with an environment, adjusting actions to maximize rewards.
• The system has an agent and will attempt to adapt its actions in the environment for the highest possible reward in the long run.

Supervised Machine Learning: Regression and Classification Tasks

• Regression tasks are to predict continuous variables.
• Classification tasks are designed to predict discrete variables.

Supervised Machine Learning: Goal and Input/Output variables

• Determine a function from training data which gives the output for input data samples, capturing the relation between inputs and outputs.
• Input variables can be X1, X2, x p
• Output variable (y) can be a function of all inputs.
• Predicted output (ŷ) will be related to input data.

Supervised Machine Learning: Algorithm Selection and Assessment

• Train and test examples
• Goal: use the trained model on new unseen data points to make accurate predictions (generalization).

Evaluating a Regression Model

• Assessing the quality of continuous predictions involves using metrics to show quality of model predictions.
• Mean Squared Error (MSE) is one common metric.
• Mean Absolute Error (MAE) is another metric.

Evaluating a Regression Model - Overfitting

• A model is overfitting if it performs well on training data but poorly on new data.
• Simpler models are often better generalized models because they do no overfit to the training set.

The Bias-Variance Tradeoff

• Overfitting is associated with high variance and a lack of bias in comparison to underfitting.
• Underfitting is associated with lack of variance and high bias.
• Finding the appropriate or suitable model is vital and involves a balance of the two factors (bias and variance).

Evaluating a Classification Model

• Assess the quality of models predicting discrete variables.
• Consider accuracy, precision, recall, and F1-score to evaluate model performance.
• Metrics to accurately predict outcomes and assess model performance can be extracted using the confusion matrix.
• ROC curve can assess a classifiers capacity to differentiate between classes.

Classification Problems

• Classification problems involve categorizing data into discrete classes.
• Common classification use cases include:
  • deciding if a customer will default on a loan
  • determining if a transaction is fraudulent
  • identifying objects in images.

Regression Problems

• Regression problems involve predicting continuous values.

Evaluating Models - General Considerations

• Using test examples to evaluate models for a test example.
• Using test data alone, to give good indicators of how well a model will work in new and unseen data

Structure of Training and Prediction

• Split the data into training and testing sets.
• Train the model on the training set.
• Evaluate the model's performance on the testing set.
• Predict on the testing set.
• Examine the accuracy of the predictions.