Day 19: Training

Goal

Understand training in plain English.

Learn

• During training, the model sees many examples and adjusts itself to reduce mistakes.
• For recognition, the input might be video or pose and the target might be gloss. For production, the input might be text or gloss and the target might be pose motion.
• Training needs repeatable data splits, clear targets, and logs. Otherwise it is hard to know whether the model improved or simply memorized a small set.

Example

• Simple training loop: load a batch, run the model, get a prediction, compare prediction to target, compute loss, update model weights, repeat.
• For a gloss classifier, the model might predict THANK-YOU when the target is PLEASE. The loss tells training that prediction was wrong.

Practice

1. Draw a training loop: input -> model -> prediction -> compare -> update.
2. Label where the pose file, gloss label, and model output appear.

Checkpoint

Quality note