Giles et al. (1990)

use a second-order DTRNN of the kind used by Giles et al. (1992), but supplemented with an external analog (continous) stack, to learn successfully two simple context-free languages. A set of special output units, called action units, are used to control the stack. The stack is continuous, in that it may contain elements of variable thickness; this is so that gradient-based algorithms may be used.