Predictive processing as a scalable computational principle for embodied multitask intelligence

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Robotics

Background:

• Human adaptability relies on integrating multimodal sensory information into predictive models.
• Current AI often requires dimensionality reduction or manual feature engineering for complex sensory data.

Purpose of the Study:

• To develop a scalable hierarchical multimodal recurrent neural network based on predictive processing and the free-energy principle.
• To enable direct integration of high-dimensional visuo-proprioceptive inputs without preprocessing.
• To establish a computational foundation bridging brain theory, AI, and embodied robotics.

Main Methods:

• A hierarchical multimodal recurrent neural network was designed, grounded in predictive processing.
• The network directly processed over 30,000-dimensional visuo-proprioceptive inputs from a humanoid robot.
• The model was trained end-to-end on teleoperation data for tasks like repositioning and wiping.

Main Results:

• The framework demonstrated self-organized hierarchical latent dynamics for task transitions and uncertainty inference.
• The model showed robustness to visual degradation through effective multimodal integration.
• Emergent properties included inference of occlusion and asymmetric interference in multitask learning.

Conclusions:

• The developed neural network successfully learns to predict complex sensory streams in real-time.
• The framework offers a generalizable computational approach for embodied AI and robotics.
• Future applications include closed-loop robot control driven by predictive proprioception.