An Integration of Deep Learning and Neuroscience for Machine Consciousness

Consciousness can be understood as a computational tool that evolved to integrate information across the brain's modular organization. Specialized modules process information unconsciously, while subjective conscious experience arises from the global availability of data through a nonmodular global workspace in the parieto-frontal brain. Large neurons with long axons enable long-distance connectivity, stabilizing and transmitting selected information to all other modules. This paper discusses necessary elements for designing conscious artificial-intelligence devices, arguing that implementing these features would likely result in a machine considered conscious. It reviews theories and specific problems that must be solved, and debates implications for neuroscience and machine learning research.

Study at a glance

Abstract

Conscious processing is a useful aspect of brain function that can be used as a model to design artificial-intelligence devices. There are still certain computational features that our conscious brains possess, and which machines currently fail to perform those. This paper discusses the necessary elements needed to make the device conscious and suggests if those implemented, the resulting machine would likely to be considered conscious. Consciousness mainly presented as a computational tool that evolved to connect the modular organization of the brain. Specialized modules of the brain process information unconsciously and what we subjectively experience as consciousness is the global availability of data, which is made possible by a nonmodular global workspace. During conscious perception, the global neuronal work space at parieto-frontal part of the brain selectively amplifies relevant pieces of information. Supported by large neurons with long axons, which makes the long-distance connectivity possible, the selected portions of information stabilized and transmitted to all other brain modules. The brain areas that have structuring ability seem to match to a specific computational problem. The global workspace maintains this information in an active state for as long as it is needed. In this paper, a broad range of theories and specific problems have been discussed, which need to be solved to make the machine conscious. Later particular implications of these hypotheses for research approach in neuroscience and machine learning are debated.