Read on Twitter
5 differences between a machine learning system and the software you are building today:
1. The Team
Usually, a machine learning system needs the involvement of many different disciplines:
- Data Scientists
- Data Engineers
- Machine Learning Engineers
Plus, the same roles that are needed by a conventional software system.
Usually, a machine learning system needs the involvement of many different disciplines:
- Data Scientists
- Data Engineers
- Machine Learning Engineers
Plus, the same roles that are needed by a conventional software system.
2. The Development Process
Machine learning is a very experimental process. Creating a model requires a lot of exploration, usually not needed in software development.
Machine learning is a very experimental process. Creating a model requires a lot of exploration, usually not needed in software development.
3. The Testing Process
Testing a machine learning system is much more involved than testing a regular piece of software.
Here are three steps unique to machine learning:
Data validation
Testing model updates
Model validation
Testing a machine learning system is much more involved than testing a regular piece of software.
Here are three steps unique to machine learning:
Data validation Testing model updates Model validation
4. The Deployment Process
In a machine learning system, you are dealing with an entire pipeline, from data collection and model training all the way to automatic model monitoring.
This pipeline is much more complex than a regular CI/CD cycle in software development.
In a machine learning system, you are dealing with an entire pipeline, from data collection and model training all the way to automatic model monitoring.
This pipeline is much more complex than a regular CI/CD cycle in software development.
5. Monitoring
Data is constantly changing, and it directly impacts the performance of machine learning systems in production.
Systems decay even without going through any modifications.
This requires constant monitoring to detect and correct drift.
Data is constantly changing, and it directly impacts the performance of machine learning systems in production.
Systems decay even without going through any modifications.
This requires constant monitoring to detect and correct drift.
