Machine Learning for Everyone - Lecture by Kylie Ying

Introduction

• Kylie Ying: Physicist, engineer, worked at MIT, CERN, Free Code Camp.
• Focus: Machine Learning (ML) for beginners.
• Topics: Supervised and Unsupervised learning models, basic logic/math, programming on Google CoLab.

Resources

• UCI Machine Learning Repository: Source of datasets.
• Example dataset: Magic Gamma Telescope dataset.
• Tools: Google CoLab, NumPy, pandas, matplotlib.

Dataset Overview

Magic Gamma Telescope Dataset

• Information collected by a gamma telescope camera/detector.
• Attributes: Length, width, size, asymmetry, etc.
• Goal: Predict if particles are gamma particles or hadrons.
• Steps: Import data, preprocess data, label columns, transform labels into numerical values.

Supervised Learning

Overview

• Definition: Subfield of computer science focusing on algorithms that allow computers to learn from data without explicit programming.
• AI vs. ML vs. Data Science: AI simulates human tasks, ML focuses on data-driven predictions, Data Science finds patterns and insights in data.
• Supervised Learning: Uses labeled inputs to train models.
• Tasks: Classification (binary, multi-class) and Regression.

Key Concepts

• Features: Input attributes used to predict labels.
• Classification: Predicting discrete classes (binary or multi-class).
• Regression: Predicting continuous values.
• Training, Validation, Testing datasets: Split data for training models and evaluating performance.
• Models: K-Nearest Neighbors (KNN), Naive Bayes, Logistic Regression, Support Vector Machines (SVM), Neural Networks.

Models and Examples

K-Nearest Neighbors (KNN)

• Concept: Classify data points based on the majority label of nearest neighbors.
• Distance metric: Euclidean distance.
• Implementation: KNeighborsClassifier from sklearn.
• Evaluation: Accuracy, Precision, Recall, F1 score.

Naive Bayes

• Based on Bayes' theorem and conditional probability.
• Assumption: Independence of features.
• Implementation: GaussianNB from sklearn.
• Evaluation: Similar metrics as KNN.

Logistic Regression

• Concept: Models probability of class membership using a logistic function.
• Implementation: LogisticRegression from sklearn.
• Evaluation: Similar metrics, plus visualization of decision boundary.

Support Vector Machines (SVM)

• Concept: Finds hyperplane that best separates classes with maximum margin.
• Implementation: SVC from sklearn.
• Evaluation: Often performs well, especially with clean data.

Neural Networks

• Structure: Input layer, hidden layers (neurons), output layer.
• Training: Backpropagation, adjusting weights to minimize loss.
• Implementation: TensorFlow for defining and training models.
• Evaluation: Compare performance with simpler models.

Regression

Concept

• Predicting continuous values using input features.
• Linear Regression: Fit a line to minimize prediction error (residuals).
• Assumptions: Linearity, Independence, Normality, Homoscedasticity.

Evaluation Metrics

• Mean Absolute Error (MAE)
• Mean Squared Error (MSE)
• Root Mean Squared Error (RMSE)
• R-squared (coefficient of determination)

Example: Bike Sharing Dataset

• Steps: Import data, preprocess, split into training/validation/testing sets.
• Models: Simple and multiple linear regression, neural networks for regression.
• Visualization: Scatter plots, regression lines.

Unsupervised Learning

Overview

• Definition: Learning from unlabeled data to find patterns or structure.
• Types: Clustering (e.g., K-Means) and Dimensionality Reduction (e.g., PCA).

Models and Examples

K-Means Clustering

• Concept: Partition data into K clusters based on feature similarity.
• Steps: Initialize centroids, assign points to nearest centroid, recompute centroids, iterate until stable.
• Evaluation: Visualize clustering results.

Principal Component Analysis (PCA)

• Concept: Reduce dimensionality by projecting data onto principal components with the largest variance.
• Steps: Compute principal components, transform data to lower dimensions.
• Evaluation: Visualize data in reduced dimensions.

Example: Seeds Dataset

• Data: Geometric attributes of wheat kernels.
• Goals: Apply K-Means clustering and PCA to identify different wheat varieties.
• Visualization: Compare original classes with clustered results.

Conclusion

• Summary of supervised (classification and regression) and unsupervised learning models.
• Encouragement to explore further, experiment with models and datasets.
• Invite feedback and collaborative learning.