Search PubMed⌕ Search

PubMed · 8823945

Learning and generalization in cascade network architectures.

Abstract

Incrementally constructed cascade architectures are a promising alternative to networks of predefined size. This paper compares the direct cascade architecture (DCA) proposed in Littmann and Ritter (1992) to the cascade-correlation approach of Fahlman and Lebiere (1990) and to related approaches and discusses the properties on the basis of various benchmark results. One important virtue of DCA is that it allows the cascading of entire subnetworks, even if these admit no error-backpropagation. Exploiting this flexibility and using LLM networks as cascaded elements, we show that the performance of the resulting network cascades can be greatly enhanced compared to the performance of a single network. Our results for the Mackey-Glass time series prediction task indicate that such deeply cascaded network architectures achieve good generalization even on small data sets, when shallow, broad architectures of comparable size suffer from overfitting. We conclude that the DCA approach offers a powerful and flexible alternative to existing schemes such as, e.g., the mixtures of experts approach, for the construction of modular systems from a wide range of subnetwork types.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

E Littmann, H Ritter. 1996-10-01. Learning and generalization in cascade network architectures.. https://doi.org/10.1162/neco.1996.8.7.1521

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related citations

A cross-linguistic speech error investigation of functional complexity.

This work is a systematic, cross-linguistic examination of speech errors in English, Hindi, Japanese, Spanish and Turkish. It first describes a methodology for the generation of parallel corpora of error data, then uses these data to examine three general hypotheses about the relationship between language structure and the speech production system. All of the following hypotheses were supported by the data. Languages are equally complex. No overall differences were found in the numbers of errors made by speakers of the five languages in the study. Languages are processed in similar ways. English-based generalizations about language production were tested to see to what extent they would hold true across languages. It was found that, to a large degree, languages follow similar patterns. However, the relative numbers of phonological anticipations and perseverations in other languages did not follow the English pattern. Languages differ in that speech errors tend to cluster around loci of complexity within each language. Languages such as Turkish and Spanish, which have more inflectional morphology, exhibit more errors involving inflected forms, while languages such as Japanese, with rich systems of closed-class forms, tend to have more errors involving closed-class items.

Generalization, Psychological↗