ART-EMAP: A neural network architecture for object recognition by evidence accumulation

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Existing neural networks face challenges in recognizing complex patterns, especially in spatio-temporal data and ambiguous 3D object views.
• Adaptive Resonance Theory (ART) and evidence integration methods offer potential for improved pattern recognition but require synthesis.

Purpose of the Study:

• To introduce a novel neural network architecture, ART-EMAP, for enhanced pattern class recognition.
• To apply ART-EMAP to spatio-temporal image understanding, prediction, and 3D object recognition from 2D views.
• To extend fuzzy ARTMAP capabilities through a four-stage incremental development.

Main Methods:

• Developed ART-EMAP, synthesizing Adaptive Resonance Theory (ART) and Evidence integration for Dynamic Predictive Mapping (EMAP).
• Implemented a four-stage extension of fuzzy ARTMAP:
• Stage 1: Distributed pattern representation.
• Stage 2: Confidence-based decision criterion for ambiguous object identification.
• Stage 3: Medium-term evidence accumulation.
• Stage 4: Unsupervised learning for performance fine-tuning.

Main Results:

• ART-EMAP demonstrates effective pattern recognition across supervised and unsupervised learning paradigms.
• Simulations using benchmark data (both noisy and noise-free) illustrate the efficacy of each ART-EMAP stage.
• The architecture successfully handles spatio-temporal image understanding and 3D object recognition from ambiguous 2D views.

Conclusions:

• ART-EMAP provides a robust framework for advanced pattern recognition tasks.
• The incremental, multi-stage design allows for adaptability and improved handling of complex visual data.
• This architecture offers a promising approach for applications requiring sophisticated image and object recognition.