A self-learning cognitive architecture exploiting causality from rewards.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cognitive Science

Background:

• Current deep reinforcement learning (RL) methods often require large labeled datasets.
• Understanding raw video content without explicit labels remains a significant challenge in artificial intelligence.
• Mimicking human vision and learning offers a promising direction for more efficient AI.

Purpose of the Study:

• To develop a novel cognitive architecture for understanding raw video content in terms of objects, without relying on labels.
• To enhance reinforcement learning (RL) by identifying and utilizing causally relevant objects.
• To improve the efficiency and transfer learning capabilities of AI systems in visual environments.

Main Methods:

• A cognitive architecture inspired by human vision, employing foveal vision and memory to decompose frames into objects.
• Internal canvas projection for world description and causal analysis of object-reward relationships for object relevance extraction.
• Implementation of learning by causality (Wiener-Granger causality) using object trajectories and reward time series.
• Development of a non-parametric directed information estimator based on Renyi's entropy.

Main Results:

• The proposed architecture successfully extracts most relevant objects from raw video frames.
• Demonstrated 'understanding' of the world in an object-oriented manner.
• Outperformed state-of-the-art deep reinforcement learning methods in training speed and transfer learning.

Conclusions:

• The novel cognitive architecture enables effective video content understanding and object extraction without labels.
• Causal learning significantly accelerates and improves reinforcement learning processes.
• This object-oriented approach represents a significant advancement in AI for visual learning and decision-making.