Dense Out-of-Distribution Detection by Robust Learning on Synthetic Negative Data

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Standard machine learning models struggle with inputs outside their training data distribution, leading to confident but incorrect predictions.
• Dense prediction tasks, like image analysis, are particularly vulnerable as anomalies can be partial.
• Existing out-of-distribution detection methods using real negative datasets may overestimate performance due to data overlap.

Purpose of the Study:

• To develop a robust method for dense out-of-distribution detection.
• To address the limitations of using real negative datasets for training and evaluation.
• To improve the reliability of machine learning models in real-world, unpredictable scenarios.

Main Methods:

• Generating synthetic negative data patches along the inlier manifold's border.
• Utilizing a jointly trained normalizing flow with a coverage-oriented learning objective.
• Employing a principled information-theoretic criterion for anomaly detection during training and inference.

Main Results:

• The proposed method achieves state-of-the-art performance on benchmarks for out-of-distribution detection.
• Demonstrated effectiveness in road-driving scenes and remote sensing imagery.
• Achieved superior results with minimal additional computational cost.

Conclusions:

• The novel approach of generating synthetic negative data significantly enhances out-of-distribution detection capabilities.
• The method provides a more reliable and principled way to identify anomalous inputs in dense prediction.
• This research offers a significant advancement for deploying machine learning in safety-critical applications.