Skylight believes in delivering world-class models to the maritime community. It follows several principles to ensure that machine learning models are able to perform well outside the confines of a controlled environment and into the real world to help drive impact.

Leverage state of the art deep learning

Train via supervised learning against expertly annotated large datasets 

Build new neural nets as needed if the state of the art doesn’t exist.  

Leverage large deep learning networks for feature extraction and lightweight classification networks that are easy and cheap to train.  

Test and evaluate in real world conditions

Test the model as early as possible in the research and development cycle and stress test extensively (global scale, large time scales) but not exhaustively (fail fast)

Listen to users 

Make it easy for them to provide feedback 

In app feedback

Leverage this feedback to improve models.

High engineering standards 

Build continuous integration/continuous deployment machine learning pipelines that enable high velocity iteration and improvement 

Extensive testing (unit, integration, ML specific)

Performance

Aim for operations grade reliability = high precision / expert-level human level accuracy 

Minimize latency as much as possible

Weight precision higher than recall (but aim for high performance in both)

Build open, transparent, and explainable AI 

Free, and permissive licensing (including commercial applications)