“Machine Learning Yearning” by Andrew Ng, published in 2018, is a practical guide for building and improving machine learning systems. The book is designed for engineers and technical team leaders who want to get better results from their machine learning projects. Ng’s insights are drawn from years of experience in both academia and industry, and the book is structured to provide actionable advice on every aspect of machine learning projects.

Major Point: Many machine learning projects fail not because of poor algorithms but due to a lack of clear objectives and properly defined metrics.

Concrete Example: Ng discusses a scenario where a team spends months tuning a model’s hyperparameters, only to realize that the model’s performance metric didn’t align with the business’s actual needs.

Action: Ensure that you define clear and measurable objectives for your project from the outset. Determine key performance indicators (KPIs) that are closely aligned with business goals.

Major Point: Using a single number evaluation metric helps in making objective comparisons between different models and versions.

Concrete Example: In a voice recognition system, rather than using a combination of metrics, Ng suggests using word error rate (WER) as the single metric to evaluate performance.

Action: Choose one metric that will determine the success of your project. Ensure it’s the most relevant one to capture the overall performance of your model effectively.

Major Point: Analyzing the types of errors your model makes can provide invaluable insights into where improvements are needed.

Concrete Example: Ng describes a scenario where an image recognition system fails to correctly identify images with poor lighting. Error analysis reveals that many misclassifications are due to lighting conditions, guiding the team to augment the training data with more varied lighting scenarios.

Action: Regularly perform error analysis by categorizing the types of errors your model makes. Use this information to guide data collection and preprocessing steps.

Major Point: Breaking down errors into orthogonal, i.e., non-overlapping, categories can help in isolating causes and improving performance incrementally.

Concrete Example: In a spam email classifier, Ng suggests decomposing error into false positives and false negatives. Addressing these separately can provide focused solutions.

Action: Decompose your model’s error into orthogonal metrics. This will help in systematically isolating and addressing different error categories.

Major Point: Creating a reliable validation set is crucial for model evaluation.

Concrete Example: Ng notes that a common mistake is using a validation set that doesn’t represent real-world data distributions, leading to overly optimistic performance estimations.

Action: Ensure that your validation set is representative of the actual data your model will encounter in production. Randomly split data while maintaining real-world proportions.

Major Point: High bias implies underfitting while high variance indicates overfitting. Understanding and diagnosing these can aid in troubleshooting model issues.

Concrete Example: Ng explains a case where a speech recognition model performed poorly. By analyzing bias and variance, the team realized the problem was due to underfitting, prompting them to increase model complexity.

Action: Regularly perform bias and variance analysis on your model’s errors. Use this analysis to decide whether to adjust model complexity, collect more data, or enhance feature engineering.

Major Point: In many ML applications, more data can often lead to better performance than more complex algorithms.

Concrete Example: Ng presents a scenario involving a cyclical prediction task where adding additional training examples led to significant improvements in model accuracy.

Action: Focus on augmenting your training dataset. Use techniques like data augmentation for digital data or actively seek more real-world data through user interactions and data partnerships.

Major Point: Quantify how error scales with data size to make informed decisions about collecting more data.

Concrete Example: In a machine translation system, Ng describes how plotting error rates against the amount of training data showed diminishing returns after a specific threshold, signaling when to switch focus to model or feature improvements.

Action: Create learning curves by plotting your model’s performance against different amounts of training data. Use this to determine when to gather more data or refine the model.

Major Point: Pre-trained models can serve as excellent starting points for many applications, reducing training time and required data.

Concrete Example: Ng discusses how using pre-trained convolutional neural networks (CNNs) for an image recognition project and fine-tuning them for specific tasks led to faster and more accurate results than training from scratch.

Action: Search for relevant pre-trained models in your domain and start with transfer learning. Fine-tune these models on your specific dataset to improve performance quickly.

Major Point: Sometimes your data distribution differs from the pre-trained model’s original training data. Domain adaptation techniques can help bridge this gap.

Concrete Example: In a sentiment analysis project, Ng explains how adapting a pre-trained language model to a specific product review dataset resulted in more accurate predictions.

Action: Use domain adaptation techniques to adjust pre-trained models to better fit your data. This might involve retraining certain layers or using domain-specific fine-tuning methods.

Major Point: Machine learning development is inherently iterative. Regularly refining models based on new data and error analysis is essential.

Concrete Example: Ng details the iterative loop of deploying a model, gathering new data from user interactions, analyzing errors, and retraining the model to handle new data better.

Action: Establish an iterative process for continuous improvement where models are regularly updated based on new data and performance analyses.

Major Point: Systematic hyperparameter tuning can significantly impact model performance.

Concrete Example: Ng shares how grid search and random search techniques help in finding optimal hyperparameters more efficiently than manual tuning.

Action: Use grid search, random search, or specialized frameworks like Bayesian optimization to systematically tune hyperparameters and improve model performance.

Major Point: Deployment involves more than just having a good model. Issues like model responsiveness, scalability, and inference time are crucial.

Concrete Example: Ng illustrates a case where a high-performing model couldn’t be used due to high latency, necessitating optimizations for faster inference.

Action: Consider practical constraints such as hardware limitations, latency requirements, and scalability from the beginning. Use tools and techniques to optimize model inference time and resource usage.

Major Point: Post-deployment, monitoring model performance and regularly retraining with new data ensures continued relevance.

Concrete Example: In an online ad recommendation system, Ng underscores the importance of tracking model performance metrics in real-time to detect drift and retrain the model as user preferences evolve.

Action: Implement continuous monitoring systems to track model performance in real-time. Set up automated retraining pipelines to update models with new data regularly.

“Machine Learning Yearning” by Andrew Ng is a treasure trove of practical advice for anyone involved in machine learning projects. From setting clear objectives and choosing the right evaluation metrics to harnessing more data and deploying models effectively, the book emphasizes a systematic and iterative approach to machine learning. By following the actionable steps outlined, technical teams can significantly improve their model’s performance and usability in real-world applications.