One-to-Many Neural Networks

For the purposes of the card one unit means a full set of features seen at one time.

• Start with the engram used to synthesize an output and run the network as in a one-to-one model.
• The network usually has an extra "stop" symbol added as an output which indicates the AI is done.
• There is some mechanism for the network to store history and basically send state back to its own input. Sometimes with recurrent networks (which carry a little bit of state) or having a block of outputs that are just copied back to the input.
• Run the network and find the output its predicts, copy necessary feedback state, and continue until stop symbol or failsafe.

Insert engram, turn crank, (copy feedback state to next cycle), get outputs.

GRUs and LSTMs are common here because they are able to carry their own state and attempt to learn when to open and close themselves for new state.