Sequences: Learning the Summary

A fixed object can be compressed once; a sequence arrives one piece at a time and may stop at any length. This Part builds order into the computation and makes the carried summary learnable. One update rule is shared across time, while gates decide what the state should keep, forget, and expose. An encoder and decoder can then turn one variable-length sequence into another, but every source fact still crosses one fixed-size bridge. The encoder had a state at every position; Part IV begins when we stop throwing those states away and learn where to look.