Why doesn't the Artificial Neural Network stop training at the point where it achieves the best performance indicated by the lowest MSE on the performance plot?

Why doesn't the Artificial Neural Network stop training at the point where it achieves the best performance indicated by the lowest Mean Squared Error (performance) on the performance plot?

Answer 1

[Ganesh]

Hi @Sunita,

The ANN cannot stop at the lowest point, because at the time of training there is no lowest point. Ideally, the lowest loss that a model can achieve is 0. However, during training, the model has no idea of whether the loss will increase or decrease further down the line.

A model training works in the following way:

1. The model starts off with initial weights.
2. The model makes a prediction and assesses the loss attained for a data point.
3. The model corrects the weight according to the loss attained.
4. The model makes a prediction using the corrected weight and assesses the loss for next data point.
5. The process continues until all data points are exhausted.
6. The performance plot shows the loss attained at each step.

The model thus, has no way of knowing it’s performance on data it has not seen.

It is to be noted that the lowest mean squared error does not mean that the model is performing the best at that point. It means that the model performs well for the data it has already seen. Allowing it to train over more data would ensure that the model makes accurate predictions for unseen data.

Hope this answer helps!

[Matt J]

It means that the model performs well for the data it has already seen. Allowing it to train over more data would ensure that the model makes accurate predictions for unseen data.

Even if it has already seen the data in earlier epochs, the loss shown on the training curve is still only an evaluation of performance with respect to the current minibatch. The training curve will never show you the loss of the model with respect to the complete data set - you just have to trust that the minibatch loss is comparable to the global loss.

[Sunita]

Thank you for your response. My question pertains to the following data obtained during model training:

1. The model stopped at epoch 22 with an MSE of 0.0242
2. The best performance was observed at epoch 16 with an MSE of 0.031
3. However, the final MSE between actual and predicted values is 0.0308

Shouldn't the third data align with the best performance (second) value ?

[Matt J]

The best performance was observed at epoch 16 with an MSE of 0.031

An MSE of 0.031 is worse than an MSE of 0.0242. MSE is something you want to minimize.

Shouldn't the third data align with the best performance (second) value ?

No, see here:

https://www.mathworks.com/matlabcentral/answers/2063317-why-doesn-t-the-artificial-neural-network-stop-training-at-the-point-where-it-achieves-the-best-perf#comment_3007012