Learn Before
Handling variance after enlarging a fraud-detection model
Case context: You are building a fraud-detection neural network for mobile payment transactions. The current model shows high avoidable bias on the training data, so you expand the network by adding extra hidden layers. After retraining, the training error improves, but the model now performs worse on fresh data and shows high variance.
Question: What is the best next step to address the new variance problem without undoing the increase in model size?
Sample answer: Add regularization. Regularization is the standard way to reduce the variance introduced by a larger model.
Key points:
- Increasing the network size reduced the original high bias.
- The enlarged model now overfits, which means variance has increased.
- Regularization is the appropriate fix for the new variance problem.
Rubric: Full credit for identifying regularization as the next step and connecting it to the variance created by the larger model.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
Why would a team choose a larger neural network during development?
True or False: A larger model can reduce underfitting, but it may also become more sensitive to noise in the training data.
Fill in the blank: When a larger model begins to overfit, _____ is commonly used to reduce the resulting variance.
Match model adjustment terms with their descriptions.
Order the actions for handling high avoidable bias.
Which change most directly increases a neural network's capacity?
True or False: Increasing the width or depth of a neural network can reduce underfitting.
Fill in the blank: Increasing the model’s capacity can reduce _____ error caused by an overly simple model.
Match each concept with what usually happens when model capacity increases.
Steps for Increasing Model Capacity Safely
Explain why enlarging a model can help, and what new problem it may create.
Handling variance after enlarging a fraud-detection model
Why can a larger model lower avoidable bias?