> ML SERIES :: index

Non-graded · offline · explore at your own pace

Machine learning, from a straight line to neural networks

Fourteen interactive pages that build on each other. Start with what it means for a machine to learn from data, fit your first model with gradient descent, learn to spot and fix overfitting, then work up through neural networks and the ways attackers target ML systems.

Sinclair College - Dr. Mike Libassi

01 What is AI & ML How AI, machine learning, and deep learning nest, and the three ways a machine can learn. 02 Structure of ML systems Features, labels, train/test splits, and the pipeline from raw data to a working model. 03 Python ML toolkit NumPy, pandas, and scikit-learn, plus the fit-and-predict pattern every later page uses. 04 Linear regression Fit a line by hand, measure the error, then let gradient descent find the best fit. 05 Optimizing training Learning rates, feature scaling, mini-batches, and why momentum and Adam converge faster. 06 Underfitting & overfitting Bias versus variance, and reading training and validation error to find the sweet spot. 07 Regularization L1 and L2 penalties, early stopping, and dropout to keep a model from memorizing. 08 Classification & logistic regression The sigmoid turns a score into a probability and draws a decision boundary. 09 Evaluating models Confusion matrices, precision, recall, and ROC curves, and why accuracy can lie. 10 Unsupervised learning k-means clustering and anomaly detection when the data has no labels. 11 Multi-layer networks Neurons, activation functions, and backpropagation through hidden layers. 12 Convolutional networks Slide small filters across an image to find edges, shapes, and objects. 13 Recurrent networks & attention Models with memory for sequences, and the attention idea behind transformers. 14 ML security & ethics Adversarial examples, data poisoning, model theft, and bias in deployed models.
Foundations · 01–04 Core techniques · 05–10 Neural networks & security · 11–14

RESOURCES

REF Notation & glossary The math symbols, Greek letters, and key terms used across the series, with how to read each equation out loud. NEXT Further learning Set up Jupyter or Google Colab, then find the best free courses, books, and datasets to keep going.
A small neural network classifying an email Four input features at the top feed two hidden layers of five neurons each, connected by weighted lines. Two output neurons at the bottom give the probability that the email is spam (0.94) or not spam (0.06). links CAPS sender length input layer features hidden layer 1 learned patterns hidden layer 2 combined patterns 0.94 spam 0.06 not spam output probability
How a network reads an email. Each input is a feature of the message, and each line is a learned weight: thicker means stronger, cyan weights are positive and amber weights are negative. The hidden layers combine those signals until the output gives a probability. By page 11 you will build one of these yourself.