Symbol processing in RAAM neural networks.
The ability to construct and manipulate recursive symbol structures is regarded
as fundamentally important in the domain of cognitive modelling.
The aim of this thesis dissertation is to explore how well Pollack's Recursive
Auto-Associative Memory (RAAM) networks can represent and facilitate the
manipulation of highly-recursive structures.
Using mainly skewed and balanced binary trees, the representational power
of the RAAM architecture is examined for structures which are lexically simple
and syntactically complex. This is in contrast to much published work on RAAM
networks, in which the structures encoded are lexically complex but syntactically
A new RAAM tree-processing operation, which allows partial information
about a set of siblings to be used as a parent pointer, is described and tested.
Several empirical investigations are motivated and carried out, to determine
how effectively RAAM networks can encode highly-recursive structures. The
investigations demonstrate the sensitivity of the RAAM architecture with respect
to the initial conditions, training parameters and the training strategies used.
This work also introduces some new techniques which help to address the twin
problems of extended training times and obtaining successful RAAM encodings.
A completely new method for performing terminal detection is presented
as well as a technique for refining Pollack's (1990) terminal detection method.
In both cases, the rate at which successful RAAM encodings are obtained is
significantly better than using Pollack's method. In addition, the new implicit
terminal detection method might allow improved RAAM generalisation, although
this conjecture has not yet been tested.RAAM networks have been used as an important counter-example to influential
analyses of the shortcomings of connectionist cognitive models. The limited
success of the RAAM networks in this study brings into question connectionist
hopes for an effective RAAM-based cognitive model.