Framing ML Problems
Project: @gortium’s Journey To Become a Machine Learning Engineer
Repository: GCP_ML_Engineering_Cert_Framing_ML_Problems
Key terms
- machine learning: Learning algorithms
- supervised learning: Learning from data with labels
- unsupervised learning: Learning from data without labels
- Reinforcement Learning: Learning from try and error and a reward function
- ML Solution Readiness: Assessment of the readiness of a ML solution
- Data availability
- Data quality
- Data limitation
- Responsable AI practices
- ML Uses Cases: A task a ML can solve with a business context. Exemple:
- Segmentation
- Fraud detection
- Demand forecasting
- ML Problem Types:
- Classification
- Progression
- clustering
- Business Success Criteria: Specific measurable goal or outcomes that indicate whether a ML solution is successful in addressing the business problem. Exemple:
- Improving accuracy
- Reducing a cost
- Increasing efficiency
- Bias in ML:
- Fair model predictions
- Unbiased model predictions
- Accurate model predictions
- Business Impact Assessment: Impact of a specific ml solution and to communicate to stakeholders
- Data readiness: Data is useful for ML training
- Data availability
- Data quality
- Data limitation
- Responsible AI Practices: Design in a way that it is ethical and trust worthy to avoid negative consequences and legal consequences
Translating business challenges into ML use cases
Github for teaching MLOps
- Reproduccibility: Codespaces
- Access to GPU: Machine Learning Codespaces
- AI Coding Assistant: Copilot
- Continuous Integration & Deploy: Github Actions
MLOps?
- WHY?: You need to retrain and redeploy ML model continuously
- HOW?: Github codespaces, copilot, actions
- WHAT?: Github actions link to any platform
Transclude of Framing-ML-Problems-2024-03-17-22.23.16.excalidraw
MLOps rule of 25
- 25% DevOps
- 25% Data engineering
- 25% MLOps
- 25% Stakeholder communication