Would this be considered underfitting?

Training an LSTM (with 410 datasets) to simulate the response of a system.

Network settings are as follows:

layer = [
    sequenceInputLayer(3,"Name","Sequential Input Layer")
    lstmLayer(240,"Name","LSTM Layer")
    fullyConnectedLayer(50,"Name","Fully Connected Layer")
    dropoutLayer(0.5)
    fullyConnectedLayer(1,"Name","Fully Connected Layer2")
    regressionLayer("Name","Regression Output Layer")];

When training, the following learning curve is shown. The training and validation RMSE never converge and remain offset.

Does this indicate underfitting? If not what am I looking at, and is it acceptable?

Thank you in advance!

Answer 1

[Anshika Chaurasia]

Hi Lucas,

It is my understanding that you want to know whether your model is underfit or if it is not, then why training and validation loss are not converging.

“Underfitting occurs when the model is not able to obtain a sufficiently low error value on the training set.” – Deep Learning, by Ian Goodfellow

On seeing the graph, training and validation loss curves have low values. So, we can say model is not underfit.

In graph, validation loss is less than training loss because of the following reasons:

• Validation dataset is easier to learn as compared to training dataset. So, check whether validation dataset follows same distribution as training dataset.
• Regularization: Dropout is applied during training only. It helps in achieving better generalization on unseen datasets.

The reason for both validation and training never converge and remain offset could be that the model is not learning after certain epochs. You could try to experiment with hyperparameters like learning rate, no. of layers, dropout layer probability etc.

Answer 2

[Greg Heath]

A model is UNDERFIT

if and only if

No. of independent training equations < No. of unknowns

Hope this helps

Thank you for formally accepting my answer.

Greg

[Torsten K]

Dear Greg,

how to calculate the number of training equations Ntrneq = prod(size(ttrn)) = Ntrn*O if I have 1 Output and 105 timeseries with 600 timesteps each? The targets are organized as cell-array of dimension 1x600, where each cell contains a 1x105 double-array with the target value (so T{1,1}(1,1) contains the 1st timestep/1st timeseries Target, T{1,2}(1,1) contains the 2nd timestep/1st timeseries Target and so on).

I guess the mentioned equation is only for a single time-series, which means 1 sample. So, how can I calculate Ntrneq in my case?

Best regards

Torsten