Global versus local Heuristic Terminal Attractor
| Autor: | Francisco Javier Marín Martín, Francisco Sandoval Hernández, F. García |
|---|---|
| Rok vydání: | 1995 |
| Předmět: | |
| Zdroj: | Lecture Notes in Computer Science ISBN: 9783540594970 IWANN |
| DOI: | 10.1007/3-540-59497-3_216 |
| Popis: | The Heuristic Terminal Attractor (H.T.A) [2,3] is one of the most widely used algorithms for training feedforward neural networks. This algorithm ensures the completion of the learning process in finite time, as well as reaching the global minimum of the error function. The H.T.A. implementation introduces a local adaptive gain factor, in the way that it only affects the weights which belong to a given neuron. The weights actualization rule works with partial information from the weights vector, making the local algorithm slower in learning time. In this paper, we introduce the global gain strategy for H.T.A., which updates the matrix weights using all the weights of the network, instead of partial information of the matrix weights. This global strategy provides shorter learning times than the local one. A theoretical computational study is given to compare the viability of global versus local algorithm, for an only processor and in the case in which a processor is available for every neuron. The results are shown for the encode/decode problem and for the pattern recognition of the alphabetical capital letters. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|