Wayfarer: Real-time Data Drift and Out of Distribution (OOD) Detection Package

Wayfarer is R-DEX’s flagship dataset drift and out-of-distribution (OOD) detection package that utilizes an ensemble of techniques to identify dataset drift and OOD samples. By minimizing the labeling workload for analysts, Wayfarer streamlines the process of keeping machine learning models aligned with their production data, ensuring models remain accurate over time. Compared to current solutions for data T&E and in particular data drift detection, R-DEX’s methods are more lightweight (less computationally expensive, hence more scalable) while maintaining comparable or better performance. Furthermore, R-DEX’s methods are easier to integrate into future pipelines due to their non-intrusive nature. Indeed, these techniques are portable and require minimal re-training, making them faster and much easier to use than existing methods. This results in more accurate real-time model predictions for operations. The primary goals of this SBIR Phase II are to 1) build a TRL 6 Phase II prototype 2) perform a comprehensive performance analysis of the proposed solution and 3) demonstrate the completed, integrated solution in a series of evaluations with the customer. This project will result in a fully autonomous, adaptive system that can operate exclusively at the edge, minimizing dependence on centralized retraining and analyst intervention.