Object-centric proto-symbolic behavioural reasoning from pixels.

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Robotics

Background:

• Autonomous agents require bridging low-level sensory input and high-level reasoning.
• Unsupervised learning methods are needed to avoid costly data annotations.
• Object-centric representations offer a promising interface between perception, action, and reasoning.

Purpose of the Study:

• To present a novel, brain-inspired deep learning architecture for autonomous agents.
• To enable agents to learn from pixels using object-centric representations for interpretation, control, and reasoning.
• To demonstrate the utility of this approach in synthetic environments requiring logical reasoning and continuous control.

Main Methods:

• Developed a novel deep learning architecture inspired by the brain.
• Employed object-centric representations learned from pixel data.
• Utilized synthetic environments (2D and 3D dSprites) for task evaluation.

Main Results:

• The agent learned emergent conditional behavioral reasoning and logical composition.
• The agent successfully controlled its environment based on deduced logical rules.
• The agent demonstrated online adaptation to environmental changes and robustness to world model violations.

Conclusions:

• Object-centric representations serve as a key inductive bias for unsupervised learning in autonomous agents.
• The proposed architecture facilitates behavioral reasoning by manipulating grounded object representations.
• Future work can extend this approach to more complex, real-world scenarios.