The Contour to Classification Game

Irene Lee, Safinah Arshad AliAAAI 2021 · DOI

Abstract

Teaching the internal mechanics of Convolutional Neural Networks (CNNs) to K-12 students is challenging due to the abstract nature of feature extraction and backpropagation. Researchers at MIT developed the "Contour to Classification Game," a browser-based multiplayer simulation where students act as nodes in a neural network. By shifting from word-based analogies to a visual system of drawing and compositing image contours, the game helps students build mental models of how AI transforms raw pixels into classifications through iterative learning.

Introduction

While students often understand that AI "learns from data," the specific mechanics—how a network extracts features or updates weights based on error—often remain a "black box." Previous educational tools used word-based analogies (e.g., describing images with text), which created a conceptual gap between human language and the mathematical reality of computer vision. Furthermore, physical classroom games were often too slow to demonstrate the temporal nature of gradient descent.

Methodology

The "Contour to Classification Game" evolves these concepts into a digital multiplayer experience involving 6 to 11 players across four layers:

  1. Input Layer: Players draw four distinct 30-pixel contours on a transparent overlay of an animal image.
  2. Hidden Layer 1: Nodes receive 16 total contours and must select the most salient segments to pass forward.
  3. Hidden Layer 2: A single node composites these segments into a partial outline.
  4. Output Layer: A final player selects an animal label based on the composite.
Figure 2: The game's hierarchical structure illustrating the feed-forward and back-propagation phases using image contours.

When a prediction is incorrect, the "Back Propagation Phase" highlights the active paths in red (incorrect) or green (correct). This allows players to re-strategize their drawing or selection process in subsequent rounds, mimicking the iterative process of gradient descent.

Figure 5: The game interface for a hidden layer node, where players must select and composite drawings from the previous layer.

Key Insights and Results

A pilot study with 38 students (ages 10–16) highlighted several key findings:

  • Visual Representation: Forcing students to work with "puzzle pieces" (contours) more accurately simulates convolutional feature identification than word-based metaphors.
  • Learning Gains: While students showed significant improvement in identifying network layers, gains in understanding complex training mechanics were more modest, suggesting a need for further iterative design.
  • Engagement: Qualitative feedback was highly positive, with students finding the "unveiling" and collaborative aspects engaging and "hilarious."

Conclusion and Future Work

The transition from word-based to contour-based interaction reduces cognitive load and aligns more closely with CNN architecture. Future work includes field-testing the contour-based version and adapting the "human-in-the-loop" model to other architectures like Recurrent Neural Networks (RNNs).

Summary updated

BibTeX
@inproceedings{lee2021contour,
  author = {Irene Lee and Safinah Arshad Ali},
  title = {{The Contour to Classification Game}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17835},
}
Figure