Paper Summary: Mapping the Landscape of Human-Level Artificial General Intelligence

This is the paper review of the following paper.

Adams, Sam, Itmar Arel, Joscha Bach, Robert Coop, Rod Furlan, Ben Goertzel, J. Storrs Hall et al. “Mapping the landscape of human-level artificial general intelligence.” AI magazine 33, no. 1 (2012): 25-42.

The field of AGI is chaotic. For 70 years, there were many ups and downs. Still, we don’t have a consensus about how we can assess the human-level artificial general intelligence. This paper is an effort of many experts in the AGI field to derive a method to assess the AGI.

They start from Piaget and Vigotsky’s theory about developmental psychology. Piaget’s model starts with sensori-motor stage where an infant learns to coordinate perceptual and motor skills such as reaching, grasping, crawling, and walking. Vygotsky’s theory emphasize the role of others and social learning.

With this in mind, they present previous scenarios for assessing AGI. They are

  • General Video-game learning
  • Preschool learning
  • Reading comprehension
  • story or scene comprehension
  • school learning (highschool or college)
  • Wozniak Test

Scenario Milestones on the AGI Landascape

This paper is helpful for my research because it states the assessment scenario is an important challgenge in the AGI community.

One challenge is to find tasks and environments where all of these characteristic are active, and thus all of the requireements must be confronted. A second challenge is that the existence of an architecture that achieves a subset of these requirements does not guarantee that such an architecture can be extended to achieve other requirements while maintaining satisfaction of the original set of requirements.

Also it lists the requirements for the such scenarios that I can apply to my planned environment.

Required Characteristics for AGI Environments, Tasks, and Agents.

I criticize that they are all focusing on too challenging scenarios without plan to achieve it. The assessing scenarios should focus on the earlier stage when the agents begins to acquire language while developing sensor-motor skills. In this sense, general video-game learning and preschool learning are relevant. However, general video-game learning lacks how the skills from each game will be accumulated and transferred to other games. Preschool learning is similar to our planned environments, but the main difference is that pre-school learning assumes that the agent has already acquired the communication capability as can be shown in the tasks highligted.

A few example tasks for Preschool Environments. I highlighted the tasks where communication with human language is required.

However, we don’t know how we can teach such communication skill to achieve artificial agents yet.

Below is my mindmap for the related papers to artificial general intelligence.