Toward Generalized Artificial Intelligence by Assessment Aggregation With Applications to Standard and Extreme...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Specialized Convolutional Neural Networks (CNNs) often lack generalizability.
• Developing Artificial Intelligence (AI) that can generalize across diverse tasks remains a significant challenge.

Purpose of the Study:

• To propose a novel plural learning framework for deriving Generalized Artificial Intelligence (GAI).
• To enhance AI capabilities by integrating specialized CNNs through a two-stage training process.

Main Methods:

• A two-stage training framework: specialization and generalization.
• Specialization stage: individual CNNs learn to predict independently.
• Generalization stage: an integration network learns from specialized CNN outputs (softmax probabilities).

Main Results:

• Demonstrated generalization through multimodel, multimodal, and distributed schemes.
• Multimodel: CNNs on the same data modality cooperate.
• Multimodal: CNNs specialize in different input types.
• Distributed: CNNs exchange assessments for joint decision-making.

Conclusions:

• The proposed framework significantly improves AI performance in both standard and extreme classification tasks.
• The integration network effectively learns from diverse assessment measures provided by specialized CNNs.
• This approach offers a robust method for building more generalized AI systems from specialized components.