How hard is it to teach a program to recognize everyday things? Attempt to build a rigid, rule-based classifier. Watch how quickly edge cases break the system, demonstrating the feedback loop of machine learning.
Help train our code by classifying the card below.
Information about why this broke the rule goes here.
Use this interactive game live with an audience to demonstrate why rigid heuristics (if/else rules) crash in production and how machine learning closes the gap.
Select Sandwich. Show Card 1 (Turkey BLT). Ask: "Who thinks this is a sandwich?" (Unanimous yes). Click YES. Explain: The machine compiles a rule: if it has bread slices and filling, it is a sandwich. Show Card 2 (PB&J). Click YES. Rule stays stable. Easy!
Show Card 3 (Open-Faced Roast Beef). Ask: "Is this a sandwich?". Watch the room split. If they say YES, show how the program allows single slices. If they say NO, show how it restricts it to exactly two slices. Explain: Humans already disagree on boundary limits.
Reveal Card 4 (Hot Dog). Ask: "Is a hot dog a sandwich?". If they say YES: Click YES. Point out that now, our rule has become so broad that it will also classify tacos, wraps, and oreos as sandwiches. If they say NO: Click NO. The code updates to exclude "connected buns". Point out: "Aha! But a Subway footlong bun is connected. Did we just break our code for Subway?"
Explain Pattern, Exception, Feedback Loop: Traditional programming forces us to write endless code patches for edge cases. Machine learning shifts the burden from writing rules to sourcing data, but we still encounter this exact feedback loop: we discover a failure, collect representative samples, retrain, and test for regressions.