Skip to content

Overfitting Escape Lab: a game on common machine learning failure

“Overfitting Escape Lab” is an interactive escape-room style web app. The player is trapped in a colorful, cartoonish AI research lab. Each room represents a common machine learning failure (overfitting, vanishing gradients, data leakage, bias, catastrophic forgetting). In every room the player sees a live training visualization (loss curves, accuracy, confusion matrix) and interactive controls (sliders for learning rate, regularization strength, dropout, etc.).
They must diagnose the problem and apply the correct fix to unlock the door.
Add a timer, humorous AI assistant characters that give deliberately bad advice, collectible debug tools, and short educational tooltips that appear when the player succeeds.

How to play

  1. Enter the lab and start the clock.
  2. Each room opens with a broken baseline. Read the curves before touching knobs.
  3. Tune the experiment, run it, then name the failure. Metrics and diagnosis have to be right.
  4. Staff AIs (Otto, Gabe, Lena, Barry, Fran, Piper) give confident, terrible advice. Their trap button applies it.
  5. A successful unlock drops a debug tool and a notebook page. Later rooms use those as hints.
  6. Clear the five research rooms, then the On-Call Vault — a production model with several bugs stacked.

The rooms, in order

RoomFailure
Memorization ClosetOverfitting
The Quiet ShaftVanishing gradients
The Filing CabinetData leakage
Half-Lit BenchBias
Chalk Dust ArchiveCatastrophic forgetting
On-Call VaultCompound production failure

Sliders, feature chips, and live train/val charts, confusion matrices, and cohort slices are all in the room. Pause freezes the clock. Best escape time is saved for the next tour.


Discover more from Science Safari

Subscribe to get the latest posts sent to your email.

Leave a Reply

error: Content is protected !!