Sequence generation with connectionist state machines

摘要

Summary form only given, as follows. Backpropagation networks withstate memories were trained to generate sequences of discrete events. Inone study, sequential networks were trained to produce `verbal'descriptions of objects in a microworld. In a second set of studiesnetworks were trained to manipulate a blocks world. One version requiredthe network to generate a sequence of actions for manipulating theblocks in response to instructions. A second version trained networks togenerate actions to move blocks from an initial configuration to a goalstate. In a final set of studies, networks generated strings offeatures. These networks were shown to take advantage of the structureof the output sequences and to apply output rules when generatingsequences