Mozer and Das (1993)

propose what could be considered as an asynchronous DTRNN-like architecture which effectively acts as a pushdown automata implementing a shift-reduce parsing strategy.^6.3The implementation is based on the following modules:

• A stack which may contain input symbols (terminals) shifted onto it or nonterminal symbols (variables).
• A set of demon units that can read the top two stack symbols. Each demon unit reacts to a particular pair of symbols, pops both of them from the stack, and pushes a particular nonterminal. When no demon unit reacts, a new input symbol is pushed onto the stack.

Grammars induced are of the form $X\rightarrow YZ$ where $Y$ and $Z$ are either nonterminals or terminals. Both the stack (as in Giles et al. (1990)) and the demons are continuous to allow for partial demon activity and variable-thickness symbols in the stack. This in turn allows for the use of a gradient-based method to train the network; the network is trained to obtain a stack containing simply the start symbol after reading a grammatical string, and any other symbol after reading a nongrammatical string. Mozer and Das (1993) successfully train the network to learn four simple grammars from relatively small learning sets.