|Network architecture and size
||Network size and architecture are defined automatically by the adaptive training algorithm (Cascade Correlation)
||Implements classic Multi-Layer Perceptron with full control over the network architecture (you can customize number of layers and neurons)
||Cascade Correlation networks use the Quick-Propagation training algorithm to train the network
||You can select one of the 5 most proven training algorithms including Conjugated Gradients and Quasi Newton.
||Validation dataset is defined automatically as a part of the Training dataset. You can specify its size in %.
||You can select one of 2 options:
- Define a Validation subset as a part of the Training subset (in %)
- Specify custom validation dataset table
||Column types are defined manually. List of categories and numeric range is automatically detected from the values.
||You can define configuration for each data column including column types, list of supported categories and numeric ranges. If configuration is not specified, it will be detected automatically.
||You can set parameters to stop training by number of iterations, desired error level or error improvement value.
||You can set a number of iterations for the current training. Additionally, you can handle the IterationComplete event to perform any custom check for early stopping and immediately interrupt training if needed.
|Generalization loss control
||You can define the rate of generalization loss and the system will automatically stop training when it is reached.
||There is no built-in feature for generalization loss control, but you can define generalization loss logic in the IterationComplete event handler.
|Real-time information about training progress
||You can get the following values on each iteration: MSE, AE, CCR, IterationNumber. All errors are calculated for the whole dataset.
||NeuroFusion Pro calculates only internal network MSE value. All other errors that require data transformation are not calculated by default to avoid slowing down the training. You can easily calculate required errors on any amount of data records using the Query method of the neural model.
|Additional training of existing networks
||There is no option to train the existing neural network one more time. New training means new network.
||NeuroFusion Pro allows you to continue training the network if needed. For example, you can combine different training algorithms (start with Quick Propagation, then stop and continue with slower but more powerful Conjugated Gradients). Or you can do additional training of the network when the new data becomes available.