StackingNet: Collective Inference Across Independent AI Foundation Models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Ensemble Methods

Background:

• Foundation models in AI excel in specific tasks but operate in isolation.
• Coordinating diverse, black-box AI models is crucial for trustworthy intelligent systems.
• Existing methods lack a framework for effective collaboration among independent AI models.

Purpose of the Study:

• To introduce StackingNet, a meta-ensemble framework for coordinating independent foundation models.
• To demonstrate StackingNet's ability to aggregate predictions and improve overall AI performance.
• To validate StackingNet's effectiveness across various AI domains without internal model access.

Main Methods:

• Developed StackingNet, a meta-ensemble framework that aggregates predictions from independent foundation models at inference.
• Applied StackingNet to tasks including language comprehension, visual attribute estimation, and academic paper rating.
• Evaluated StackingNet's performance against individual models and traditional ensembles.

Main Results:

• StackingNet consistently improved accuracy and reduced errors compared to individual models.
• The framework effectively ranked model reliability and identified underperforming models.
• Performance gains were attributed to variance reduction and consensus alignment, increasing with model diversity.

Conclusions:

• StackingNet provides a practical method for coordinating specialized AI models, fostering cooperation.
• This approach enhances AI capabilities by leveraging model diversity as a resource.
• Future AI progress can emerge from principled cooperation among multiple specialized models, not just larger single models.