|
|
|
|
|
|
|
|
|
Переобучение
| Глоссарий терминов по нейронным сетям |
Ситуация, когда на обучающей последовательности ошибки сети были очень малы, но на новых данных становятся большими.
|
Overfitting, английский
- A term used in neural networks, recursive partitioning, and other automated modeling areas. if the training data is considered to consist both signal and noise (i.e., noisy data), a modeling technique has begun to overfit when it begins capturing the "noise" instead of the "signal." this usually occurs when a model is allowed to increase its number of "parameters`` such as regression coefficients, splits in recursive partitioning schemes, or hidden units in neural networks. the effect of overfitting is to reduce the applicability of the model to other data sets, (i.e., to limit its generalizability). in the extreme, the model can only "predict" its input. the most straight-forward way to eliminate this effect is to require a very large number of observations per parameter in the model.
- A problem in data mining when random variations in data are misclassified as important patterns. overfitting often occurs when the data set is too small to represent the real world.
|
|
Образец (эталон), русский
«правильный» результат на выходе нейронной сети. совокупность всех эталонов вместе с входным набором данных (фактов) образует обучающее множество данных.
Выходной слой, русский
Слой нейронов (узлов), выходные сигналы которых образуют выходной сигнал нейронной сети.
|
|
|
|
|
|
|
