Google

Machine learning strategy is the discipline of deciding which changes are most likely to improve a model, based on evidence from errors, data splits, and project constraints. It helps you prioritize the next experiment instead of guessing blindly.

Machine Learning Strategy

A good choice of model-improvement direction can move a project forward quickly, but a poor choice can consume months of effort. Learning to notice the clues in a problem helps you focus on changes that are likely to matter.

Choose the Right Improvement Path

Many machine learning tasks contain practical clues about which approaches are worth trying and which are likely to waste time. Paying attention to those clues can prevent long, unproductive rounds of trial and error.

Machine Learning Tasks Often Suggest Where to Focus Effort

In a machine learning project, strategy is used to map the overall objective to the concrete choices the team must make about data, modeling, and workflow.

Linking Project Goals to Engineering Decisions

Even modest changes in how a machine learning team ranks its tasks can noticeably change how much progress it makes.

Small Priority Shifts Can Change Team Throughput

The guide begins with strategies that apply to a wide range of machine learning methods, including both classic algorithms and neural networks, and then moves toward ideas that are especially useful for modern deep learning systems.

Broad Machine Learning Guidance First, Deep Learning Guidance Later

Machine Learning Yearning @ DeepLearning.AI

For a new machine-learning task, the best approach usually cannot be identified reliably before experiments are run. This uncertainty causes teams, including experienced practitioners, to test and compare multiple candidate approaches before selecting one that performs well enough.

Uncertainty About the Best Machine Learning Strategy Drives Experimentation

Possible improvements for an underperforming image classifier fall into two broad groups. Data-side options include collecting more labeled examples and making the training set more varied. Model-side options include training for more iterations, changing network capacity, adjusting regularization, or redesigning the architecture. Use evidence from the model's errors to decide which option to test rather than applying every change indiscriminately.

Learn Before

Related