Vertex AI (60-70%, heavy emphasis)
- Vertex AI AutoML: You can train models without coding, but learn when to use AutoML effectively versus building custom models.
- Vertex AI Pipelines and Orchestration: These are essential for building real-world ML systems. Expect questions on how pipelines work, the different components, how to troubleshoot them, and streamlining and automating your ML workflows for efficient and -scalable machine learning solutions on Google Cloud.
- Vertex AI Experiments: Keep track of different models and experiments. Make sure you learn how to track, compare, and analyze different model versions and training runs.
- Vertex AI Metadata: Helps you organize all your ML stuff (models, datasets, etc.). Know how to use it to track your work and see how everything connects.
- Vertex AI Model Registry: Where you store your trained models. Know how to register them, deploy them, and manage different versions.
- Vertex AI Endpoints: How you make your models available for others to use. Understand how to deploy models for online prediction and the different deployment options available.
- Vertex AI Feature Store: A newer service that helps you organize, store, and serve machine learning features. It can be important for improving model accuracy and consistency.
- Vertex AI Training: Go beyond AutoML and explore custom training options, including pre-built containers, custom containers, and distributed training strategies.
- Vertex AI Prediction: This component handles serving predictions from your deployed models. Understand different prediction methods (online, batch) and scaling options.
- Vertex AI Explainable AI: Helps you understand how your models make predictions. It’s increasingly important for model transparency and debugging.
- Vertex AI Model Monitoring: Vertex AI has specific tools for monitoring model performance, drift, and fairness.
ML fundamentals are essential (20-25%)
- Types of algorithms: Know the different types of ML algorithms (like linear models, clustering, regression, and classification) and when to use each one.
- Evaluation: Understand how to interpret metrics like recall, precision, and accuracy.
- Monitoring: Know how to keep an eye on your models after you deploy them and spot problems like model drift or bias.
- Hardware: Understand when to use TPUs, GPUs, or regular CPUs for your ML tasks.
- TensorFlow and TFRecords: Be comfortable with TensorFlow and know how to use TFRecords to feed data to your models efficiently.
- Data splitting and cross-validation: Understand how to split your data for training, validation, and testing. Know different cross-validation techniques.
- Bias and fairness in ML: Be aware of potential biases in data and models. Understand techniques for mitigating bias and ensuring fairness.
- MLOps principles: This includes concepts like continuous integration, continuous delivery, and model versioning.
- Security in ML: Understand security considerations for data, models, and infrastructure.
Other GCP services (10-15%) and final review
- Cloud Natural Language API: This one is for working with text. Understand what it can do, when you’d use it, and how much it costs.
- BigQuery ML: Allows you to build and deploy models directly within BigQuery. It can be useful for simpler models or when your data is already in BigQuery.
- Cloud Storage: Fundamentals for storing and accessing data for your ML workloads.
- Cloud Logging and Monitoring: Essential for monitoring your ML pipelines and deployed models.
- Cloud Functions: Used for lightweight serverless deployments or triggering actions in your ML workflows.
- Dataflow, Pub/Sub, Firebase: These are important for building data pipelines and connecting different parts of your ML system.
Professional Machine Learning Engineer Sample Questions
https://docs.google.com/forms/d/e/1FAIpQLSeYmkCANE81qSBqLW0g2X7RoskBX9yGYQu-m1TtsjMvHabGqg/viewform
key decision points:
- Unstructured data storage: Cloud Storage.
- Structured data storage/analysis: BigQuery.
- Generic problem, no custom data: Pre-trained API.
- Custom problem, have data, no code skills: AutoML.
- Data in BigQuery, SQL-proficient team: BigQuery ML.
- Need full control/custom framework: Custom Training on Vertex AI.
- Orchestrate an ML-centric workflow: Vertex AI Pipelines.
- Orchestrate a complex business workflow with ML and non-ML parts: Cloud Composer.
- Guarantee training-serving consistency: TFX tf.Transform or Vertex AI Feature Store.
- Automate hyperparameter tuning: Vertex AI Vizier.
- Monitor deployed models for drift/skew: Vertex AI Model Monitoring.
Responsible AI
- Responsible AI is a framework of principles and practices to ensure that AI systems are developed and used in a way that is fair, safe, accountable, and respects privacy.
- The exam will expect familiarity with its key pillars.
- Fairness and Bias: This involves being aware of and mitigating potential biases in data and models that could lead to unfair outcomes for certain user groups.
- Tools within the TensorFlow ecosystem like Fairness Indicators and the What-If Tool are designed to help analyze and uncover these biases.
- **Explainability (XAI): ** This addresses the "black box" problem by providing methods to understand why a model made a specific prediction. Vertex Explainable AI integrates with prediction services to provide feature attributions (using methods like SHAP), which highlight the features that had the most significant influence on a given prediction.
- Safety and Privacy: This encompasses protecting models from adversarial attacks and ensuring that user data is handled securely and ethically, in line with Google's AI Principles.
- For generative models, this includes the use of safety filters to block harmful content generation.
Parnter content to leverage: https://www.skills.google/paths/17/course_templates/1171 Machine Learning Crash Course Best practices for implementing machine learning on Google Cloud
- Identify the Core Constraint: Every scenario question has a primary driver. Is the main concern cost-effectiveness, scalability, speed of development, interpretability, or the existing skill set of the team? The "best" solution is always relative to this primary constraint.
- Rapid-Fire Review: Internalize these key decision points:
- Unstructured data storage: Cloud Storage.
- Structured data storage/analysis: BigQuery.
- Generic problem, no custom data: Pre-trained API.
- Custom problem, have data, no code skills: AutoML.
- Data in BigQuery, SQL-proficient team: BigQuery ML.
- Need full control/custom framework: Custom Training on Vertex AI.
- Orchestrate an ML-centric workflow: Vertex AI Pipelines.
- Orchestrate a complex business workflow with ML and non-ML parts: Cloud Composer.
- Guarantee training-serving consistency: TFX tf.Transform or Vertex AI Feature Store.
- Automate hyperparameter tuning: Vertex AI Vizier.
- Monitor deployed models for drift/skew: Vertex AI Model Monitoring.