Machine learning has achieved superhuman performance in specific domains, but efficient generalization across domains remains elusive. Humans achieve this through declarative memory formation, closely linked to consciousness. This paper proposes that predictive processing (PP) systems, through online learning by hierarchical binding of unpredicted inferences, can flexibly generalize by forming working memories. The contents of these working memories, unified yet differentiated, are maintained by selective attention and align with observations of consciousness research. The paper describes how perceptual value prediction could have evolved for reinforcement learning of complex action policies, suggesting 'conscious experience' as a perceptual representation of the system's functioning. The proposal unifies feature binding, recurrent processing, and predictive processing, and offers a functional description facilitating experimental testing. Ethical implications of numerical experiments are also considered.

V A Aksyuk, Vladimir Aksyuk