IEEE International Conference on Neural Networks
DOI: 10.1109/icnn.1993.298572
|View full text |Cite
|
Sign up to set email alerts
|

Optimal Brain Surgeon and general network pruning

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
1
1
1

Citation Types

0
426
1
11

Publication Types

Select...
8
1

Relationship

0
9

Authors

Journals

citations
Cited by 504 publications
(438 citation statements)
references
References 2 publications
0
426
1
11
Order By: Relevance
“…To address these problems, constructive and destructive algorithms, with growing or shrinking networks, have been proposed. [3][4][5][6] A potential weakness of the constructive approaches is their sensitivity to noise, while the destructive approaches often include retraining steps that require prohibitive computational efforts. This paper illustrates how a novel pruning algorithm recently developed by the authors of the present paper 7) can be used to tackle the complex problem of predicting the silicon content in hot metal produced in a blast furnace.…”
Section: Introductionmentioning
confidence: 99%
“…To address these problems, constructive and destructive algorithms, with growing or shrinking networks, have been proposed. [3][4][5][6] A potential weakness of the constructive approaches is their sensitivity to noise, while the destructive approaches often include retraining steps that require prohibitive computational efforts. This paper illustrates how a novel pruning algorithm recently developed by the authors of the present paper 7) can be used to tackle the complex problem of predicting the silicon content in hot metal produced in a blast furnace.…”
Section: Introductionmentioning
confidence: 99%
“…Existem na literatura várias estratégias que objetivam soluções com elevada capacidade de generalização. (8,10,11) As RNAs apresentam limitações, por exemplo, a maneira pela qual o conhecimento é representado.…”
Section: Redes Neurais Artificiais E Sistemas Híbridosunclassified
“…(20) Foram desenvolvidas RNAs com uma camada intermediária, utilizando o algoritmo de treinamento por retropropagação do erro (7) e objetivando soluções com alta capacidade de generalização foram aplicadas as técnicas de Early Stopping, (8) Regularização Bayesiana, (9) algoritmo de Pruning Optimal Brain Surgeon (10) e o método Ensemble Modelling. (11) Os sistemas híbridos Neuro-fuzzy desenvolvidos foram baseados na arquitetura ANFIS (Adaptive Neuro-Fuzzy Inference System) proposta por Jang, (13) que mantêm as características básicas dos sistemas fuzzy mas têm incorporadas as propriedades de adaptação das RNAs.…”
Section: Introductionunclassified
“…Neurônios que tiverem todas as conexões cortadas serão eliminados e, portanto, ao final dos "cortes", sobrarão somente os neurônios realmente necessários à modelagem. A técnica de apodização 13,14 (pruning) reduz a complexidade da rede neural, melhorando sua capacidade de previsão, pois evita modelos sobre parametrizados (muitos neurônios) em que a possibilidade de sobreajuste (overfitting) é grande.…”
Section: Redes Neurais Com Apodizaçãounclassified