A conscious AI system based on recurrent neural networks applying dynamic information equilibrium

EasyChair Preprint 676

Abstract

A basic structure and behavior of a human-like AI system with conscious like functions is proposed. The system is constructed completely with artificial neural networks (ANN), and an optimal-design approach is applied. The proposed system using recurrent neural networks (RNN) which execute learning under dynamic equilibrium is a redesign of ANN in the previous system. The redesign using the RNNs allows the proposed brain-like autonomous adaptive system to be more plausible as a macroscopic model of the brain. By hypothesizing that the “conscious sensation” that constitutes the basis for phenomenal consciousness, is the same as “state of system level learning”, we can clearly explain consciousness from an information system perspective. This hypothesis can also comprehensively explain recurrent processing theory (RPT) and the global neuronal workspace theory (GNWT) of consciousness. The proposed structure and behavior are simple but scalable by design, and can be expanded to reproduce more complex features of the brain, leading to the realization of an AI system with functions equivalent to human-like consciousness.

Keyphrases: Recurrent Neural Network, dynamic equilibrium, model of consciousness

@booklet{EasyChair:676,
  author    = {Yasuo Kinouchi and Kenneth James Mackin and Pitoyo Hartono},
  title     = {A conscious AI system based on recurrent neural networks applying dynamic information equilibrium},
  doi       = {10.29007/2hjj},
  howpublished = {EasyChair Preprint 676},
  year      = {EasyChair, 2018}}