Dynamic and Evolving Neural Network as a Basis for AGI

EasyChair Preprint 7922, version 1

Abstract

Artificial general intelligence (AGI) should be founded on a suitable framework. Existing rule-based design is problematic, since it has to be manually updated if new and unaccounted for data is encountered. Current Deep Learning (DL) is also insufficient to become AGI, but it has the potential to be extended into one. Therefore an appropriate AGI has to be defined, followed by its appropriate DL implementation. We introduce an AGI, in the form of cognitive architecture, which is based on Global Workspace Theory (GWT). It consists of a supervisor, a working memory, specialized memory units, and processing units. Additional discussion about the uniqueness of the visual and the auditory sensory channels is conducted. Next, we introduce our DL module, which is dynamic, flexible, and evolving. It can be also considered as a Network Architecture Search (NAS) method. It is a spatial-temporal model, with a hierarchy of both features and tasks, tasks such as objects or events.

Keyphrases: deep learning, dynamic, evolving, general intelligence

@booklet{EasyChair:7922,
  author    = {Shimon Komarovsky},
  title     = {Dynamic and Evolving Neural Network as a Basis for AGI},
  howpublished = {EasyChair Preprint 7922},
  year      = {EasyChair, 2022}}