An ML team is reviewing remediations for a model that has overfit. As it reconsiders how training is run and how data is handled, which are effective measures for reducing overfitting? (Choose TWO.)

1 / 1
Select all that apply
CorrectA, B

Explanation

Question Overview

Choosing TWO measures that reduce overfitting.

Requirements to satisfy
  • 1「reducing overfitting」Methods that curb overfitting
  • 2「which are effective measures」Strengthening regularization and increasing training data apply
Per-option explanation
ACorrect

Strengthen regularization, which penalizes model complexity.

Correct. Regularization is a technique that penalizes model complexity (the magnitude of the weights). The larger the weights, the more the model forcibly fits the fine differences in the training data, causing overfitting. Adding a penalty makes training keep the weights as small as possible to reduce that penalty, so the model is less swayed by the details of the training data. As a result, generalization performance on unseen data improves.

BCorrect

Increase the training data, including with data augmentation.

Correct. Increasing the training data makes the model less likely to overfit the details of specific samples and improves generalization performance. Augmenting the data with data augmentation is also effective.

CIncorrect

Switch to a more complex, larger model.

Making the model more complex increases its expressive power, but it advances overfitting to the training data even further.

This is the opposite direction from reducing overfitting, so it is incorrect.

DIncorrect

Deliberately reduce the training data.

Reducing the training data makes the model more prone to overfitting and worsens it.

It is the opposite of a reduction measure, so it is incorrect.

EIncorrect

Greatly increase the number of epochs and keep training.

The longer training continues, the more overfitting to the training data advances.

This is a situation where you should instead consider early stopping while watching validation accuracy, so it is incorrect.

Key Takeaway

Effective measures for reducing overfitting include 'strengthening regularization' (penalizing model complexity) and 'increasing the training data' (including data augmentation) (others include early stopping, dropout, and feature reduction). In contrast, 'keeping the model more complex' and 'reducing the data' worsen overfitting.

Related Links