Requirements of AGI
What features or capabilities must an AGI have?
- transfer learning : Being able to use knowledge from one task in other tasks.
- online and incremental learning: The AI has to be able to learn at any time in an incremental fashion. Batch learning is a no go for AGI or most components of it.
- integrate different types of learning: Unsupervised, Supervised, Semisupervised and Reinforcement learning.
- open world: At any point of time new tasks and/or new knowledge can be submitted to the system.
- operating under AIKR [On the Working Definition of Intelligence]: The amount of memory of a physical machine is limited. For this reason, all (sub)tasks have to share the memory and not important tasks must be forgotten with some mechanism. Each task has to share the finite computational power of a computing system with all other tasks. Some mechanism has to exist to decide how much computation is done on all tasks.
- attention: Deciding which tasks should get how many resources(memory and time). Is related to AIKR because a system under AIKR must have a attention system.
- conceptual learning : Ability to abstract concepts from the raw perception. The AGI must not make decisions directly based on the pixel values.
- incomplete and contradicting evidence: An AGI has to have some way to revise already existing knowledge with new knowledge. Is related to conceptual learning.
- self: Any agent acting in the world or in its own mind must have a sense of a self. The actions of the agent are bound to a self symbol which represents the acting of the agent in the world. This is useful for learning actions from other agents by imitation. This is required for other self properties, too.
- self-introspection/self-awareness: Finding patterns in the recorded resource usage history and finding patterns in the recorded actions done by the agent. The agent can use these patterns to avoid making the same mistake again or to improve it’s performance and resource utilization.
Any AGI system must implement all of these points, not just one or two as it is the case with all contemporary AI systems of today.
What feature or abilities can an AGI have?
- self-reflection/self-modification : Self-reflection allows the agent to probe and reason about its own code (in some representation). Self-modification is the action of changing this representation (which include configuration values).
- ability to write new modules: See self-reflection.