How to Fix Validation Performance That Degrades After Several Epochs?
Short answer
AB Enable early stopping and increase dropout to curb the overfitting that degrades validation performance after the peak epoch.
Question
Which solutions will mitigate this problem? (Choose two.)
Key Takeaways
The test is recognizing an overfitting curve (validation metric peaks then degrades) and mapping it to regularization techniques rather than to capacity increases, which are the usual trap answers.
Validation performance that climbs and then falls after a certain number of epochs is the classic signature of overfitting in a neural network. This page explains why early stopping (A) and increased dropout (B) are the two mitigations that directly address it, and why deeper or wider layers make it worse.
Common mistake: Many candidates pick 'increase the number of layers' or 'increase the number of neurons' because those feel like ways to 'improve the model', but adding capacity accelerates overfitting and makes the post-peak degradation start even sooner.
Expert Analysis
Why the Answer Is Correct
The curve described — validation performance rising sharply and then declining while training continues — is the textbook overfitting signature: the network keeps fitting training noise, so generalization peaks and then deteriorates. Early stopping (A) is the most direct fix because it halts training at the epoch where validation performance is best, restoring the peak checkpoint instead of the degraded final one. Increasing dropout (B) attacks the same problem at the architectural level by randomly deactivating neurons during training, forcing the network to learn redundant, more generalizable representations and flattening the overfitting curve. Together they are the standard pairing: one is a training-loop control, the other a regularization hyperparameter, and both target exactly the failure mode the question describes.
Why the Other Options Are Wrong
Options C and D both increase model capacity — more layers and more neurons respectively. Extra capacity does not address overfitting; it typically worsens it, since a larger hypothesis space can memorize training data faster and push the validation degradation point even earlier in training. Option E, investigating model bias, is a solution for underfitting (high bias), where both training and validation performance are poor and the model fails to capture the underlying pattern. In this scenario the model clearly learns well initially and then diverges from validation data, which is a variance problem, not a bias problem, so reducing bias sources would not mitigate it.
Community Comment Notes
All three commenters converge on the same diagnosis. Saransundar states plainly that "The issue is overfitting" and pairs early stopping, which stops training when validation performance declines, with dropout, which randomly disables neurons. GiorgioGss quotes the stem's own phrasing about improving then degrading after a specific number of epochs and calls it a "Clear sign to stop it early and to drop," which is a useful mnemonic. Ranneh frames it functionally: early stopping halts training at the model's best point, while dropout removes neurons that contribute little to generalization — the same reasoning, independently reached.
Exam Strategy
When a stem describes a metric that improves and then gets worse over epochs, immediately label it overfitting and scan for regularization or training-control options (early stopping, dropout, weight decay, data augmentation). Any option that adds layers or neurons is a capacity increase and should be eliminated in that scenario, while 'bias' options belong to underfitting questions.
Frequently Asked Questions
Why does increasing layers or neurons not fix the degrading validation curve?
Why is reducing model bias not the right fix for this epoch curve?
Discussion 3 contributions
Compiled from public community threads and reader submissions, lightly edited for clarity. Handles and vote tallies are shown as recorded, and comments reflect the views of individual contributors.
From public community threads 3
-
Saransundar 👍 7The issue is overfitting. Soln:- A. Early stopping:- Stops training when validation performance declines B. Increase dropout:- reduces overfitting by randomly disabling neurons
-
GiorgioGss 👍 2"improves substantially at first and then degrades after a specific number of epochs." Clear sign to stop it early and to drop
-
Ranneh 👍 1Enable early stopping to automatically halt model training when it at its best to prevent the model from overfitting and also adopting the dropout to remove specific neurons within the network thats are without a great impact on the model
Reader comments & corrections
No reader corrections yet — spotted an error or have a note? Share it below.
Related Analysis
Practice All MLA-C01 Questions
Access 115 questions with complete answers and detailed explanations.
View Full MLA-C01 Practice Test →