Utilizing TOP2 Class for Hybrid Decision-Making to Enhance TOP1 Accuracy of Ensemble Models

Area of Science:

• Deep learning
• Computer vision
• Ensemble methods

Background:

• Traditional ensemble methods in deep learning visual tasks often rely on TOP1 class information, potentially overlooking valuable data.
• Existing techniques like majority voting and average probabilities have limitations in maximizing ensemble performance.

Purpose of the Study:

• To introduce a novel algorithm, TOP2 hybrid decision (TOP2 HD), designed to enhance the performance of deep learning ensemble models.
• To address the limitations of traditional ensemble methods by incorporating TOP2 class information for improved decision-making.

Main Methods:

• Developed the TOP2 hybrid decision (TOP2 HD) algorithm, which categorizes base models hierarchically based on their TOP1 class.
• Utilized the TOP2 class for ranking within the hierarchy to improve ensemble decision-making.
• Conducted extensive experiments using various models and datasets to evaluate TOP2 HD's performance.

Main Results:

• TOP2 HD significantly surpasses traditional ensemble methods, including majority voting, average probabilities, and stacking.
• The proposed algorithm demonstrates superior performance compared to many state-of-the-art ensemble strategies in the image domain.
• A functional relationship between ensemble test accuracy and the number of base models was identified.

Conclusions:

• TOP2 HD offers a superior approach to ensemble learning for visual tasks, enhancing TOP1 accuracy.
• The identified relationship allows for predicting the upper performance limit of ensemble models, aiding in deployment decisions.
• This research provides a valuable advancement in ensemble techniques for deep learning in computer vision.