What is the conclusion from this Accuracy / Loss plot for Train and Validation

Question

What is the conclusion from this Accuracy / Loss plot for Train and Validation ?

• It seems, that the best results for Validation are after few (5) epochs.
• Also I'm not comfortable how the Loss and Accuracy looks for validation.

Answer 1

welcome to datascience.stackexchange.com

Based on the diagram you provided, it appears that your model is overfitting. The spikes in both the loss and accuracy curves, along with the training accuracy reaching 100% are strong indicators of this.

You can refer to the following kdnuggets article for solutions to overfitting:

And also check for data leakage.

Answer 2

Briefly speaking, you would have 3 issues.

a) Overfitting

b) Validation set or hold-out set is far smaller or contains unbalanced classes. Implementing early stopping, say, stop at the first or second peak.

c) Use other methods or change your Neural Network or Regression/Classification model types.