Hello,
If I am training my Neural Network with training data comprised of 908 batches, as the rule of thumb, would 20 epochs for training be sufficient?
If 20 epochs are not sufficient, does this mean that I will have to re-train my neural network before performing any analysis with it?
Thanks,
It depends on the system to model (i.e. the data), but generally, the number of epochs exceeds 100. In addition, it is better to specify simultaneously another set of epochs for cross-validation to avoid the over fitting problem.
Instead of specifying the number of epochs, the error threshold between the target and the actual output is specified, say 0.001. The epochs can be set to a large number e.g. 1000. As soon as the error reduces below the threshold the iterations will automatically stop, irrespective of the number of epochs specified.
There is no rule. it is determined empirically, but it must be greater than 100
The number of epochs is not that significant. What's more important is the training and validation error. Our goal is to minimize the error for obtaining higher accuracy. As long as the error is decreasing you can increase the number of epochs and when you notice that the validation error starts increasing, it signifies overfitting. You can get optimal number of epochs by plotting network accuracy vs epochs.
Compute Epochs vs Absolute Error plot
Keep increasing the epochs value (number of iterations/passes) until your plot becomes horizontal. (Absolute error becomes constant considereing your tolerance limit)
Number of epochs for training a neural network model is either user defined or sometimes determined empirically. Besides this also, sometimes neural network models are trained until some predefined error metric is less than a threshold value.
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training