From Continuous Observations to Symbolic Concepts: A Discrimination-Based Strategy for Grounded Concept Learning

Area of Science:

• Artificial Intelligence
• Robotics
• Cognitive Science

Background:

• Autonomous agents process continuous sensori-motor data but struggle to distill symbolic concepts for reasoning.
• Current deep learning methods for bridging continuous and symbolic domains require extensive data and lack transparency, generality, and adaptivity.

Purpose of the Study:

• To introduce a novel methodology for grounded concept learning in autonomous agents.
• To enable agents to construct human-interpretable conceptual systems from raw observations.
• To develop a module for mapping sensory input to symbolic concepts for higher-level reasoning.

Main Methods:

• A tutor-learner scenario is employed for concept formation.
• Concepts are built through discriminative combinations of prototypical values on interpretable feature channels.
• The approach was evaluated on the CLEVR dataset using simulated and computer vision-extracted features.

Main Results:

• The proposed method facilitates incremental learning with few data points.
• Learned concepts demonstrate generality, applying to unseen objects.
• Concepts can be combined compositionally, enabling complex reasoning.

Conclusions:

• The novel methodology offers a transparent, adaptive, and data-efficient approach to grounded concept learning.
• This method is suitable for robotic agents, bridging sensory perception and symbolic reasoning.
• The approach supports the development of more capable and generalizable artificial intelligence systems.