Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. International Journal of Pattern Recognition and Articial Intelligence , 8(1), 391-414, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....is not guaranteed; 3) learning constants and other parameters must be arrived at heuristically. These latter two problems, premature convergence and parameter setting, have been approached using several optimization procedures, which can be divided into two groups: Incremental Decremental (see [8], by Alpaydim et al. for a review) or genetic algorithms, which are reviewed for instance in [9] Incremental algorithms, are based on adding hidden neurons to a network of minimum size until the required precision is reached. These methods start with few hidden neurons and increase their ....

Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. International Journal of Pattern Recognition and Articial Intelligence , 8(1), 391-414, 1994.

....since the learning parameters and network size depend on the problem to solve. Thus the problem of automatic MLP parameter setting and optimization remains an open question. There are two ways of approaching the optimization of BP parameters for certain problem: Incremental decremental (see [10], by Alpaydim et al. for an interesting review) or genetic algorithms (see either Yao [11,12] or Balakrishnan et al. 13] for an interesting review) Incremental algorithms, such as Cascade Correlation by Fahlman and Lebi ere [9] and the Tiling and Perceptron Cascade Algorithm presented by ....

Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. International Journal of Pattern Recognition and Articial Intelligence , 8(1), 391-414, 1994.

....rapidit. Les arbres d induction apparaissent donc comme de moins bons candidats pour modliser des processus automatiques. Deuximement l apprentissage avec les arbres d induction ncessite un seul passage de la base d exemples#; pour les rseaux base d exemples mmoriss quelques passages suffisent [Alpaydin 91] sauf quand on les combine avec des techniques base du gradient) A l inverse les PMC ncessitent de nombreuses prsentations des exemples pour apprendre. Ceci est gnralement considr comme un point faible des PMC, mais comme prcisment la cration d un automatisme est un phnomne lent et sensible la ....

....et al. 86] Le Cun 87] sont essentiellement monoblocs et homognes. Les dfauts de cette approche ont t bien identifis#: difficult pour choisir la taille du rseau [Fiesler 94] Jutten Chentouf 95] structure arbitraire pas ncessairement adapte la tche, mauvaises capacits de gnralisation [Alpaydin 91] Le Cun 89] difficults d apprentissage [Fahlman Lebiere 91] Haykins 94] dfauts qui ont amen les chercheurs proposer d autres approches#: l approche modulaire manuelle , qui consiste diviser le rseau en petits modules et entraner sparment chaque module [Hrycej 92] Kmmerer Kpper ....

[Article contains additional citation context not shown here]

Ethem Alpaydin. GAL: Networks that grow when they learn and shrink when they forget. Rapport technique, TR 91-032, ICSI, Berkeley, mai 1991.

....on the problem to solve, on the type of network (number of hidden units to use) and on the training test sets. This leaves the problem of automatic MLP parameter setting and optimization open. There are two ways of approaching the optimization problem parameters: Incremental decremental (see [7], by Alpaydm et al. for a good review) or genetic algorithms (see [8] by Yao for a good review) Incremental algorithms, such as Cascade Correlation by Fahlman and Lebiere [9] the Tiling and Perceptron Cascade by Parekh et al. 10] or the methods proposed by Zhang [11] or Rathbun et al. 12] ....

Ethem Alpaydm. GAL: Networks that grow when they learn and shrink when they forget. International Journal of Pattern Recognition and Artificial Intelligence , 8(1), 391-414, 1994.

....the technique used in this paper, g lvq, intends to optimize several codebook parameters at the same time: codevector labels, codebook size (or number of levels) and codevector initial values. Some other methods for optimizing lvq have been based on incremental approaches (for a review, see [2]) which are still local error gradient descent search algorithms on the space of networks or dictionnaries with different size. Other methods use genetic algorithms [12] to set the initial weights of a codebook with a maximum implicit weight; this maximum length limits the search space, which ....

Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. Technical Report TR-91-032, International Computer Science Institute, May 1991.

....optimum is not guaranteed; 3) learning constants must be guessed heuristically. These problems, premature convergence and parameters setting, have been approached using several optimization procedures. These procedures applied to the MLP can be divided in two groups: Incremental decremental (see [3], by Alpaydim et al. for a good review) or genetic algorithms (see [4] by Balakrishnan et al. for a good review) ffl Incremental algorithms, such as Cascade Correlation by Fahlman and Lebi ere [5] the Tiling and Perceptron Cascade by Parekh et al. 6] or the methods proposed by Zhang [7] or ....

Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. International Journal of Pattern Recognition and Artificial Intelligence , 8(1), 391-414, 1994.

....mentioned above, classification accuracy for supervised learning, or distortion for unsupervised learning, while leaving size fixed. Several size values are usually tested, and the best is chosen. Some other methods for optimizing LVQ have been based on incremental approaches (for a review, see [8]) which are still local error gradient descent search algorithms on the space of networks or dictionnaries with different size; or on genetic algorithms [9] but in this case they have got a maximum implicit size. An incremental methodology proposed by Perez and Vidal [10] seems to offer the ....

Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. Technical Report TR-91-032, International Computer Science Institute, May 1991.

....independently of the connectionist module (when an expert gives new rules) but rules can also be extracted from the connectionist module [3] 2. 2 The connectionist module The connectionist module (CM) is made up of a prototype based network which is designed to be incremental and supervised [2, 1]. It is a feed forward network as shown in fig. 3b. In this kind of network, each hidden unit corresponds to a prototype of input patterns (or situations) The basic model which we review here operates either in supervised learning mode or recognition mode. Learning consists in associating a ....

.... of this basic model, which associates a variable influence threshold for every prototype unit (i.e. each prototype unit has its own i , while in the basic algorithm of table 2 all units share the same ) This variant performs better [2] Contrary to other prototype based networks like GAL [1], no pruning is made in SYNHESYS, but it could easily be added. This kind of network was chosen because it is incremental, which is a very important property of learning systems [1] The second reason is that such a localist network (each unit corresponds to a precise prototype) will make easier ....

[Article contains additional citation context not shown here]

Ethem Alpaydin. Gal : Networks that grow when they learn and shrink when they forget. Technical Report TR-91-032, International Computer Science Institute, Berkeley, May 1991.

....performance, like classification accuracy for supervised learning, or distortion for unsupervised learning, while leaving size fixed. Several size values are usually tested, and the best is chosen. Some other methods for optimizing lvq have been based on incremental approaches (for a review, see [1]) which are still local errorgradient descent search algorithms on the space of networks or dictionnaries with different size. Other methods use genetic algorithms [12] to set the initial weights of a codebook with a maximum implicit weight; this maximum length limits the search space, which ....

Ethem Alpaydim. GAL: Networks that grow when they learn and shrink when they forget. Technical Report TR-91-032, International Computer Science Institute, May 1991.

No context found.

Ethem Alpaydin. "GAL : Networks that grow when they learn and shrink when they forget", TR 91-032, International Computer Sci. Institute, (May 1991).

No context found.

Ethem Alpaydin. "GAL : Networks that grow when they learn and shrink when they forget", TR 91-032, International Computer Sci. Institute, (May 1991).

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC