ML Ops Course Overview

Course Introduction

• High-paying job offers possible.
• Unique focus on adding a "creamy layer" on projects.
• Experience in international job offers through remote work.
• Emphasis on ML Ops as a new and valuable skill.

Instructor Background

• Lead Data Scientist at Triplet.
• Experience as an ML Ops engineer at Xenaml.
• Background in NLP products and early experience with AI.

Course Objectives

• Understanding specific ML Ops terminologies and concepts.
• Learning about data ingestion, model deployment, and tools (e.g., ML Flow, Xenaml).
• Conducting an end-to-end project from data to deployment.

ML Ops Importance

• Exponential data growth necessitates AI and ML solutions.
• Machine Learning involves more than just model training (only 20% of the project).
• Focus on the engineering aspect of ML (80%).

The Way ML Teams Function

• Roles in a typical ML team:
  • Data Scientists: Feature development and model training.
  • Data Engineers: Productionizing data pipelines.
  • ML Engineers: Deploying models.
  • Legal Advisors: Ensuring data usage compliance.

ML in Production Process

• Data collection ➡️ Model training ➡️ Model deployment
• Continuous feedback loop: retraining models based on new data and performance decay.

Deployment and Monitoring

• Deployment makes local models accessible to users (e.g., spam detection systems).
• Monitoring model performance is critical, as models can decay over time.

Importance of ML Ops

• ML Ops is a methodology extending DevOps to ML components, ensuring reliability in production.
• Ideal for scaling operations and managing ML lifecycle efficiently.

Deployment Challenges

• Latency: Slow models can lead to high abandonment rates.
• Fairness: Ethical considerations, examples from AI disasters (e.g., Microsoft's Twitter bot).
• Explainability: Users need to trust AI predictions.

ML Model Characteristics

• Differentiating between model-centric (improving model parameters) and data-centric (improving data quality) approaches.
• Advised to focus on data-centric modeling for better results.

ML Ops Project Structure

1. Business Problem Identification
   • Assess the cost of wrong predictions (e.g., sales forecasting scenarios).
2. Data Gathering
   • Analyze historical data and market trends.

Machine Learning Workflow Phases

• Data Engineering: Ingesting, cleaning, and preparing the data.
• Model Engineering: Training, validating, and deploying models.
• Pipeline Management: Organizing workflows using pipelines (e.g., Xenaml).

Conclusion

• Understanding the holistic approach to ML Ops is critical for success in the industry.
• Next steps include practical implementations with tools like Xenaml and ML Flow.