Day 23: Common Failure Modes

Goal

Recognize typical failures.

Learn

• Systems can fail from low light, hidden hands, fast motion, camera angle, rare signs, weak gloss labels, signer variation, missing facial expression, or dataset imbalance.
• Failures can also come from overfitting, duplicate clips, inconsistent annotation, or training on examples that are too narrow for real use.
• A good pipeline tracks failures so the team can improve capture, extraction, labels, model design, or review rules.

Example

• If the model fails when a signer wears a dark shirt, the background and clothing contrast may be hurting pose extraction.
• If it works for one signer but fails for others, the split may not have tested signer variation well.
• If generated motion is smooth but wrong, the production model may be optimizing motion similarity without enough language review.

Practice

1. List five failure modes.
2. For each one, write one prevention step and one review step.

Checkpoint

Pipeline note