Embracing the Power of Known Class Bias in Open Set Recognition From a Reconstruction Perspective

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open set recognition (OSR) models face challenges with known class bias, where models trained on known classes incorrectly classify unknown classes.
• Existing OSR methods either eliminate known class bias or use reconstruction methods to circumvent it.

Purpose of the Study:

• To challenge conventional approaches to known class bias in OSR.
• To propose that known class bias can be beneficial for reconstruction-based OSR methods.
• To introduce a novel framework, Bias Enhanced Reconstruction Learning (BERL), to leverage this bias.

Main Methods:

• The BERL framework enhances known class bias at class, model, and sample levels.
• Class-level enhancement uses supervised contrastive learning to prevent overgeneralization.
• Model-level enhancement employs a diffusion model with class priors for guided reconstruction.
• Sample-level enhancement utilizes a self-adaptive diffusion model strategy based on information bottleneck theory.

Main Results:

• Experiments on various benchmarks demonstrate the effectiveness of the BERL framework.
• The proposed method shows performance superiority compared to existing OSR approaches.
• BERL successfully utilizes known class bias as a positive incentive for reconstruction.

Conclusions:

• Known class bias, traditionally seen as detrimental, can be advantageous for reconstruction-based OSR.
• The BERL framework offers a novel and effective approach to open set recognition by exploiting class bias.
• The findings suggest a paradigm shift in addressing known class bias in OSR models.