Interval-Based Least Squares for Uncertainty-Aware Learning in Human-Centric Multimedia Systems.

Area of Science:

• Multimedia Signal Processing
• Machine Learning
• Computer Vision

Background:

• Existing machine learning (ML) methods in multimedia signal processing often ignore uncertainty in ground-truth data.
• This leads to overemphasis on single-target values, which is problematic when human feedback is inherently variable.
• This limitation is particularly relevant in immersive multimedia systems with higher degrees of freedom and uncertainty.

Purpose of the Study:

• To propose an uncertainty-aware loss function that explicitly accounts for data uncertainty.
• To improve the optimization and validation processes in ML models by acknowledging ground-truth variability.
• To enhance prediction accuracy in scenarios with inherent uncertainty, such as human-centric systems.

Main Methods:

• Development of a novel uncertainty-aware loss function, denoted as [Formula: see text].
• Integration of the proposed loss function into ML model training and validation pipelines.
• Application and evaluation of the method on diverse multimedia tasks.

Main Results:

• The proposed uncertainty-aware loss function effectively reduces prediction errors.
• Demonstrated utility in blind estimation of perceptual quality for audiovisual signals, panoramic images, and distorted images.
• Experimental results validate the theoretical benefits of accounting for data uncertainty.

Conclusions:

• The novel loss function offers a more robust approach to ML in multimedia signal processing by handling ground-truth uncertainty.
• The method is particularly beneficial for modern applications like crowdsourcing and immersive systems where uncertainty is high.
• This work provides a valuable tool for developing more accurate and reliable ML models in the presence of data variability.