Beyond novelty detection: incongruent events, when general and specific classifiers disagree

Area of Science:

• Machine Learning
• Artificial Intelligence
• Pattern Recognition

Background:

• Machine learning algorithms struggle with unexpected stimuli, leading to prediction conflicts.
• Identifying and processing incongruent events is crucial for robust AI systems.

Purpose of the Study:

• To develop a formal framework for representing and processing incongruent events in machine learning.
• To enable the detection of unexpected events arising from conflicting classifier predictions.

Main Methods:

• Defined a formal framework based on label hierarchies and partial order.
• Computed event probabilities using adjacent levels in the label hierarchy.
• Developed algorithms to detect incongruent events across diverse data types and applications.

Main Results:

• Successfully detected novel visual and audio objects.
• Identified new patterns of motion in video data.
• Demonstrated effectiveness in speech recognition (Out-Of-Vocabulary words) and multimodal scenarios.

Conclusions:

• The proposed framework effectively identifies and processes incongruent events.
• This approach enhances machine learning robustness when encountering unexpected stimuli.
• The method shows promise for real-world applications in object recognition, motion analysis, and speech processing.