Developmental word grounding through a growing neural network with a humanoid robot

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Current AI models often rely on pre-existing datasets for learning.
• Robots typically lack the ability for continuous, lifelong learning.
• Integrating multiple sensory inputs like speech and vision remains a challenge.

Purpose of the Study:

• To develop an unsupervised, online learning approach for robots to acquire word meanings.
• To enable incremental, lifelong learning capabilities in humanoid robots.
• To enhance the robot's ability to build a mental model from real-time data.

Main Methods:

• An unsupervised approach integrating speech and visual information without pre-prepared data.
• Online learning from continuous audio-visual input for the Incremental Knowledge Robot 1 (IKR1).
• A noise-robust, self-organized growing neural network for representing data topology.
• An active-learning mechanism termed "desire for knowledge" for targeted information acquisition.

Main Results:

• The developed approach significantly increases the efficiency of the robot's learning process.
• IKR1 successfully learns word meanings through integrated audio-visual input.
• The system constructs a mental model for the robot based on learned concepts and real-time scenes.

Conclusions:

• Unsupervised, online, and incremental learning is feasible for humanoid robots.
• The "desire for knowledge" mechanism enhances targeted learning efficiency.
• The approach provides a foundation for robots to develop an "inner world" and connect concepts to experiences.