Day 5: What the Model Learns From

Goal

Understand that models learn from processed examples, not human meaning directly.

Learn

• A model learns patterns from numbers, labels, sequences, and comparisons. It does not automatically know Deaf culture, signer intent, or whether a phrase is natural ASL.
• If the input is pose data, the model sees coordinates over time. If the input is gloss, it sees label sequences. If the input is video, it sees pixels or visual features.
• Better data gives the model a better chance. Wrong labels, missing fingers, weak frame timing, and inconsistent metadata all become training problems.

Example

• A pose-only record might keep wrist and fingertip movement but lose skin tone, clothing, background, and some mouth detail.
• That can protect some privacy and simplify training, but it can also remove information that matters for ASL meaning.

Practice

1. Compare a real signing clip with a pose skeleton preview.
2. Make two columns: kept by pose data and lost by pose data.
3. Add one sentence about how the lost information could affect the model.

Checkpoint

Quality note