Scaling Up: How a RAM Boost Improved Certus's ML Accuracy

As reported on previously, Certus has the largest training datasets for various use cases. 

However, these datasets became so large that the previous computer could no longer process them. To address this, Certus recently invested in a significantly upgraded ML computer, increasing RAM by 4x. 

RAM (Random Access Memory) is the computer's short-term memory, used to temporarily hold data that's actively being processed. In machine learning, this refers to the datasets being used to train a model. Therefore, the increase in RAM allows the system to work with these larger datasets. As a result, larger datasets expose machine learning models like Certus’s TEAM-ML models, DART and FALCON, to greater variability in real-world conditions, reducing overfitting and improving its ability to generalize accurately across diverse scenarios.

DART is Certus Innovations’ ML-based technology that allows for the identification, classification, and localization of gunshots, and FALCON is their technology that allows for the identification, classification, and localization of drones.

This upgrade allowed the DART classifier to increase its test accuracy from 96.5% to 98.8% despite adding four more calibers. Even more impressive, the FALCON classifier remained above 99% accuracy despite almost doubling the number of sUAS in the dataset.

While Certus Innovations is constantly innovating to develop their ML models further, this serves as a reminder that feeding a model more of the right data can sometimes drive some of the biggest performance gains. Building upon that, Certus recently received a massive sUAS dataset from Dr. Peter Zulch of AFRL Rome, which they are eager to combine with their own.