Hi Osman,
When training a neural network using MATLAB's ‘trainbr’ (Bayesian Regularization Backpropagation) function, several factors influence when training should stop and how to interpret the results. The training process can be stopped based on many factors like maximum epochs (2000 in this case), performance goals (MSE), when the gradient of the performance function is below a specified minimum gradient, validation failures, and a predefined time limit.
Stopping at the end of the specified epochs is acceptable, but it may indicate that the network hasn’t fully converged or reached an optimal solution. Ideally, training should stop when the network achieves a satisfactory performance level, not just because it reached the maximum epochs. If your network consistently stops at the maximum epochs without achieving desired performance, you can adjust the learning rate and can perform regularization to avoid overfitting.
The acceptable range for performance depends on the specific application and the scale of your data. For example, in regression tasks, the performance value should be as low as possible, ideally approaching zero. In classification tasks, performance is often measured by accuracy, where higher values are better. Based on domain specific requirements, level of error in accuracy and the tolerance level can be determined. Also, compare the performance value with benchmarks or previous models to determine if it is acceptable.
For more information, refer the following documentations:
- Speed up deep neural network training: https://www.mathworks.com/help/deeplearning/ug/speed-up-deep-neural-network-training.html
- How do I improve my neural network performance: https://www.mathworks.com/matlabcentral/answers/137-how-do-i-improve-my-neural-network-performance
- Assess Neural Network Classifier Performance: https://www.mathworks.com/help/stats/assess-neural-network-classifier-performance.html
Hope this helps!
