Operating Systems For Conscious Learning

Conscious learning is a new kind of developmental systems that conduct task-nonspecific learning while their consciousness is not only learned but also plays a critical role in deciding what to learn, how to learn and how well the learning is doing. It requires a new kind of operating system (OS) running on robots. This new kind of OS builds on a mobile operating system (like Android or iOS), but it is for Autonomous Programming For General Purposes (AFPGP) to learn in real-time through the sensors and effectors of the robot/phone. The traditional machine-learning approach is for a human to program a robot for a narrow set of given tasks (e.g., deep learning and ChatGPT). By definition, this corresponds to weak AI. Weak AI is brittle because the machine does not learn consciousness. This paper is based on strong AI— APFGP. APFGP is supported by (1) task-nonspecificity, (2) modality-nonspecificity, (3) conscious learning, and (4) OSFCL (Operating Systems For Conscious Learning). Due to limited space here, this paper briefly overviews (1), (2), and (3), but focuses on (4). The novelty of OSFCL includes: treating agent resources and their changes uniformly as three categories— sensors, effectors, and computational resources. Multisensor and multi-motor integration in a DN (Developmental Network) is automatic, based on maximum likelihood across a lifetime. Therefore, the DN model is free from the Post-Selection misconduct that has flooded the AI field. The importance of the work is to relieve humans from the seemingly intractable programming and data annotation for an open-ended array of robot bodies, tasks, and task environments. An Emergent Universal Turing Machine (EUTM) inside the DN is the core of (1) to (4). As far as we know, this is the first paper that introduces OSFCL.