Sensor Data Classification With Dirichlet Uncertainty Calibration

The patent combines a multi-class classifier with a generative model into a unified system. The classifier processes sensor data to identify classes, while the generative model simultaneously evaluates the likelihood of the data under each class. This merging allows the system to provide both out-of-distribution detection (via the classifier) and calibrated in-domain uncertainty (via the generative model's likelihood estimates), resolving the contradiction between these two capabilities.

The generative model acts as an intermediary between the sensor data and the final classification output. It receives the sensor data and class labels from the classifier, then computes calibrated uncertainty measures by evaluating how well the data fits each class distribution. This intermediary component enables the system to provide reliable uncertainty calibration without compromising out-of-distribution detection capability.

The system implements feedback by using the generative model to evaluate the confidence of each classification. When the generative model detects low likelihood or high uncertainty (indicating potential misclassification on edge cases or out-of-distribution data), it provides feedback that triggers further processing or alternative classification strategies. This feedback mechanism maintains high classification speed for confident predictions while improving reliability for uncertain cases.

The system applies partial action by selectively invoking additional processing only when needed. For clear, in-distribution cases, the system quickly outputs classification results without additional computation. For edge cases or out-of-distribution data detected by the generative model, the system performs additional validation or alternative classification. This approach maintains high overall productivity while improving reliability for problematic cases.