Google

If a model is showing high avoidable bias, several practical fixes can help: increase model capacity, revise the input features using error analysis, reduce or remove regularization, or adjust the architecture so it better matches the task. Simply adding more training examples is usually not the right response for this problem, because extra data often improves variance more than bias.

Methods to Reduce High Avoidable Bias

Insights from error review can point to new input features that address a specific pattern of mistakes. Adding those features may lower both bias and variance. If the added features make variance rise in practice, regularization usually can offset that increase.

Using Error Review to Guide Feature Design

When a model is underfitting, lowering or removing regularization methods such as L2 penalty, L1 penalty, or dropout can reduce avoidable bias. The trade-off is that less regularization usually raises variance, so the model may fit the training data better while becoming more sensitive to noise.

Using Less Regularization to Ease Underfitting

Which change is most likely to reduce avoidable bias when a model is underfitting?

More training data and avoidable bias

A way to lower avoidable bias is to reduce or eliminate ____.

Match each action with its main effect or purpose.

Order the response to weak training-set performance.

Question:
In a short response, explain how increasing model capacity, adding more informative inputs, lowering regularization, and choosing a better-fitting architecture can reduce high avoidable bias. Then contrast these actions with collecting additional training examples.

Sample answer:
A larger model can reduce bias by giving the learner enough capacity to fit the training data more closely. Adding or redesigning inputs can remove repeated error patterns when error analysis shows the model is missing important information. Lowering L1, L2, or dropout regularization can also reduce avoidable bias, although it may increase variance. Choosing an architecture that matches the problem better can improve performance by changing the bias-variance tradeoff. By contrast, collecting more training data usually helps with variance and does not usually fix high bias on its own.

Key points:
- More model capacity can lower bias by improving fit to the training set.
- Input changes should be driven by observed error patterns.
- Reducing regularization lowers bias but can raise variance.
- Architecture choice can influence both bias and variance.
- More data is mainly a variance remedy, not a direct solution to high bias.

Rubric:
A strong response explains each of the four tactics, states how each affects bias and variance, and makes clear that more training data is generally not the main fix for high bias.

How can four bias-reduction tactics help when a model underfits?

Case context:
A small product team is evaluating a classifier that still makes many mistakes on the data it was trained on. A review of the errors shows that many of the failures occur on examples with a rare but repeated pattern: records from older devices with noisy measurements. The team is deciding between enlarging the model, adding a new input feature, lowering regularization, changing the network design, or collecting more labeled examples.

Question:
What is the most likely diagnosis, which interventions are appropriate, and which proposed intervention is not the right way to address the diagnosed issue?

Sample answer:
The model is showing high avoidable bias. Reasonable responses include using a larger model so it can fit the training set better, adding a feature that captures the older-device/noisy-measurement pattern identified in error analysis, reducing regularization if necessary while accepting that variance may rise, or revising the architecture to better match the task. Collecting more training data should not be treated as the main fix for this problem, because more data mainly helps when variance is the issue and usually does little to reduce bias.

Key points:
- Weak performance on the training set points to high avoidable bias.
- Increasing model capacity is a valid remedy.
- Error analysis can justify feature changes targeted at the recurring failure pattern.
- Lowering regularization may improve fit but can raise variance.
- Changing the architecture can help reduce bias and may also affect variance.
- More labeled data is not the primary solution for high bias.

Rubric:
The response should diagnose high avoidable bias, name the four supported intervention categories, connect feature changes to the recurring error pattern, note the variance tradeoff of reducing regularization, and explain why more data is not the main fix for bias.


Fixing poor training-set performance in a model

Question:
Answer in one to three sentences: Why might lowering or removing penalties such as L1, L2, or dropout help when a model has high bias, and what drawback should you expect?

Sample answer:
If the model is too constrained, reducing regularization can let it fit the training data more closely and lower bias. The main drawback is that the model may become more sensitive to the training set and overfit more easily, which increases variance.

Key points:
- Regularization can be lowered or removed.
- Bias decreases.
- Variance increases.

Rubric:
The answer must state both effects: lower bias and higher variance.

Why can weakening regularization help a model that is underfitting?

Which choice best applies error analysis to reduce avoidable bias?

Changing a model’s architecture can affect both bias and variance.

Learn Before

Related