A method that drops less useful tokens to make AI work faster.
It is like clearing junk from a backpack before a hike. Keep the map. Lose the old banana peel.
It helps with long text and images. It saves work, but may toss out something important.
Token
Token Pruning removes Tokens that matter less during processing.
Attention
Attention scores can help Token Pruning choose important Tokens.
Transformer
Token Pruning gives the Transformer fewer Tokens to process.