Train

Step 4 of the machine learning workflow

NoteTrain · at a glance

Workflow step	Step 4 of 5 · Train
What you’ll learn	How a model actually learns from data, and what can go wrong during training.
Standards assessed	ESB
Key terms	training set · loss · ensemble · generalization
Materials	readings · slides · homework · lab
Self-quiz	standards & practice →

Training is where a model learns. You feed prepared data to an algorithm, and it adjusts itself to fit the patterns in that data. This chapter covers what “fitting” really means and how ensembles combine many models into a stronger one.

Overview

Training takes the algorithm you chose in Step 3 and the data you prepared in Step 2, and produces a fitted model. The central tension is generalization: a model that fits its training data perfectly may still fail on new data.

(Flesh this out later.)

Key ideas

NoteKey idea: fitting the training data is not the goal

The goal is performance on new data. A model that memorizes the training set has learned nothing useful.

• Training vs. generalization — why fitting the data isn’t the same as learning.
• Loss — how a model measures how wrong it currently is.
• Ensembles — combining many weak models into one strong model.

(Expand each as you flesh out.)

Readings

Before class, read the assigned material on ensembles and training.

• Required: add reading + link

Slides

In class we’ll work through these decks:

• Slide deck links go here (ESB1, ESB2, ESB3)

Homework

• Assignment: link
• What it assesses: training a model and diagnosing under/overfitting.

Lab

TipTry it

Build an ensemble in the lab and compare it to a single model.

• Lab: link
• Objectives: add objectives

Self-quiz

Use the standards as a checklist, then try the practice questions.

Standards

ESB — Explain how ensembles improve on single models, and train one on a dataset.

You should be able to:

• Describe why combining models can outperform any single model.
• Distinguish training performance from generalization performance.

(Add any additional training-related standards.)

Practice questions

Click a question to reveal the answer.

NoteQ1. Why can an ensemble of “weak” models outperform a single strong model?

Different models make different errors. Averaging or voting across them cancels out individual mistakes, so the combined prediction is more reliable than any one model.

NoteQ2. Your model’s training accuracy keeps rising but validation accuracy starts falling. What does this signal?

Overfitting: the model is memorizing the training data at the expense of generalizing. It’s time to stop training, simplify, or regularize.

NoteQ3. (Add your own practice question here.)

(Answer goes here.)