Recurrent Autoassociative Networks and Sequential Processing
Ivelin Stoianov, James Boswell · 1999
A novel connectionist architecture that develops static representations of structured sequences is presented. The model is based on SRNs trained on an autoassociation task in a way that guarantees the development of unique static representations. The model can be applied in modeling Natural Language, cognition, etc. Introduction Connectionist Natural Language Processing has gone a long way since the development of the NETtalk model [13] and the Simple Recurrent Networks (SRN) [5], but one problem is still remaining: how to develop static representations of sequential linguistic objects. Static representations of words, sentences, and so on are necessary in order to apply direct, "holistic" operations over sequences [7], such as associations to other modalities, extracting items, and so on. In order to allow those operations, the developed representations should uniquely characterize the original objects. A first attempt to build such representations was suggested by J. Pollack [11]. H...