| 摘要 |
The invention concerns a method for learning data classification in two distinct, classes (c11, c12) separated by a region divider (S), by a binary neurone comprising a descriptive parameter of the region divider with its input data being weighted by a weight (wi) and comprising the following steps: a) defining a cost function C: (I); b) initialising the weights (Wi), the radii (ri), the parameters ( sigma , T+, T-), the learning rate ( epsilon ) and the annealed velocities ( delta T+, delta T-); minimising the cost function by successive iterations; obtaining the values of the connection weights and the neurone radii. The invention is useful for classification and form recognition by a neural network.
|
