What To Try Next

Deciding What to Try Next in Machine Learning

Project Guidance

Introduction to Effective ML Development

• You’ve learned many powerful machine learning algorithms:
• Linear regression
• Logistic regression
• Neural networks (deep learning)
• Decision trees (coming next week)
• The key challenge now: How to use these tools effectively
• Some teams spend 6 months building systems that skilled teams could complete in weeks
• Efficiency depends largely on making good decisions about what to try next

Note

This week focuses on saving you time by providing a framework for making better decisions throughout your ML projects.

The Machine Learning Decision Problem

Common Scenario

Example scenario:

• You’ve implemented regularized linear regression for housing price prediction
• Your model makes unacceptably large errors in predictions
• What should you try next?

Possible Next Steps

• Get more training examples
• Try a smaller set of features
• Get additional features
• Add polynomial features (x₁², x₂², x₁x₂…)
• Decrease regularization parameter (λ)
• Increase regularization parameter (λ)

Decision Challenge

• Some approaches will be fruitful
• Others will waste time
• Without guidance, teams may invest months in unproductive directions
• Example: Spending months collecting more data when it won’t help

The Value of Diagnostics

Time-Saving Tools

• Diagnostics: Tests that provide insight into what is/isn’t working with your algorithm
• Help guide improvements to model performance
• Answer critical questions like:
• Is it worth spending weeks/months collecting more training data?
• Which features are most valuable to the model?
• Is the model overfitting or underfitting?

Tip

While diagnostics take time to implement, they can save you months of work by preventing you from pursuing unproductive approaches.

Looking Ahead

• This week will cover various diagnostics to guide algorithm improvement
• First step: Learning how to properly evaluate your algorithm’s performance
• Next video: How to evaluate machine learning algorithms

Caution

Making informed decisions about what to try next is often the difference between efficiently building effective ML systems and wasting months on approaches that yield minimal improvements.

Efficiently building machine learning systems requires making good decisions about what to try next. While there are many possible approaches to improving model performance, diagnostics provide critical insights that help you invest your time wisely. This week’s content will equip you with tools to evaluate your algorithms properly and make better development decisions.