A Novel Independent Monitoring Model for Efficient Out-of-Distribution Detection in CNNs: Utilizing The Forward-Forward Algorithm
Junhee Hyeon, Chaejin Lim, Seonu Park, Abdullah Muhammad, Ki-Seong Lee, Dongil Han · 2023
Out-of-Distribution (OOD) detection is critical for identifying inputs incongruent with the trained classes of a model. Incorrect classifications, notably in sectors like healthcare, autonomous driving, and industrial inspection, can be catastrophic. Addressing this, our paper introduces a new method for OOD detection essential in these fields. Our approach employs a standalone monitoring model, not connected by weights to the image classification model, thus avoiding retraining needs. The Forward-Forward (FF) Algorithm helps in understanding the connection between the final feature maps of the classification model and its results, leading to a Positive or Negative classification. This strategy enhances OOD detection efficiency and integrates well with existing deep learning models.