A self-supervised learning system for pattern recognition by sensory integration.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics and Sensory Systems

Background:

• Artificial neural networks excel at nonlinear mapping for pattern recognition.
• Supervised learning methods like back-propagation require known desired outputs, which are often unavailable in unpredictable environments.
• A gap exists in unsupervised or self-supervised approaches for category detection with ambiguous sensory data.

Purpose of the Study:

• To present a novel self-supervised learning system for automatic category detection.
• To enable systems to learn object categories by integrating information from multiple sensors without predefined outputs.
• To develop a system capable of adaptive sensor priority control based on input pattern deformation.

Main Methods:

• Development of a self-supervised learning framework for category detection.
• Integration of multi-sensor information, assuming noisy and ambiguous input patterns.
• Implementation of a learning algorithm that automatically identifies object categories.
• Inclusion of a mechanism for dynamic sensor priority adjustment during recognition.

Main Results:

• The self-supervised system successfully learned categories across various tasks using artificial and actual sensory inputs.
• Demonstrated effective object recognition capabilities.
• Successfully applied to learning Japanese vowels from mouth shape data, yielding distinct outputs for each vowel.

Conclusions:

• The proposed self-supervised learning system effectively automates category detection by integrating multi-modal, noisy sensory data.
• The system exhibits adaptability in recognizing objects and managing sensor priorities based on input characteristics.
• Validated efficacy through simulations and a real-world application (vowel recognition).