Section Assignment 2

Train a neural network, then question it

Graded: Pass / No-Pass Due: at the end of your Section 2 meeting Submit: one file (or pasted text) on Canvas

Open the lab

Why this assignment

In Session 3 you saw how machines learn from examples, why models overfit, and three cases where prediction went into policy. Today you try both sides yourself. First you train a small neural network. Then you check a model the way an auditor would, asking what "predicts well" was measured against.

What to do

Part	Time	What you do
1 · Bills	~12 min	Train a neural network to predict which bills get out of committee. Find the smallest network that reaches 88% test accuracy, then watch a big network overfit a small, messy sample. Record at least 3 runs and answer 1A–1C.
2 · Benefit claims	~12 min	Train a model that flags 10% of benefit claims for fraud checks, with and without postcode and citizenship as inputs. Judge it first by the agency's records, then by the truth. Record at least 3 runs and answer 2A–2C.
3 · Reflection	~4 min	In 3–5 sentences: Is a model that predicts well good enough to use in policymaking? Use one thing you saw today.

All data on the lab page are simulated.

To pass

The file has your name, at least 3 recorded runs in Part 1 and at least 3 in Part 2.
All seven questions (1A–1C, 2A–2C, Part 3) have a real attempt. Short is fine; blank is not.
The answers are in your own words. You may work in pairs and discuss, but each person writes and submits their own answers.

Wrong answers don't fail. This is graded on a real attempt, not on getting the "right" numbers.

How to submit

At the bottom of the lab page, press Download my submission.
On Canvas, open Section 2 lab and upload the file. Or press Copy and paste the text into the text box.
Submit before you leave section.

If you can't finish in section

Submit what you have before you leave. You may then resubmit a complete version by 11:59 pm the same day. If you missed section, contact your TA.

Privacy

Your name and answers are stored only in your browser until you download them. The page doesn't send anything anywhere.