User Tools

Site Tools


T-720-ATAI-2016 Main


Lecture Notes, F-12 11.03.2016

Seed-Based AGI

Seed A.k.a. bootstrap code – the initial knowledge given to a “baby learning machine”. A seed may consist of ontologies, states, models, internal drives, exemplary behaviors and programming skills.
Why it's important All systems whose knowledge, over time, is mostly self-acquired must get “bootstrapped” somehow – initial process of acquiring knowledge must be started somehow. This is the role of the seed.
What kinds of systems Seed-based programming is only possible for systems with a high potential for autonomy.
What kinds of worlds The importance of an appropriate seed increases with the level of ergodicity (non-reversibility/resettability) of the world. The seed must leverage processes already existing in the environment.
Do such systems exist? In AGI, only two: AERA (auto-catalytic endogenous reflective architecture) and NARS (non-axiomatic reasoning system). Both rely on non-axiomatic term logic for representing knowledge.
Historical note In the 50's Alan Turing wrote: “Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child’s?” and John von Neumann proposed ways to design self-replicating machines.

Seed-Based AI

Very different Bootstrapping a self-improving system with a seed is a radically different approach than any current method for creating “AI” systems.
What it requires A system that can self-improve must have the basic functions to:
- Identify relevant information.
- Automatically create sub-goals from a few top-level goals and drives.
- Ignore interfering data, information and processes.
- Incrementally improve its knowledge by correcting acquired knowledge.
- Integrate new knowledge with older knowledge.
- Evaluate when auto-acquired knowledge is good enough to build knew knowledge on.
- (for meta-learning systems) Evaluate when knowledge acquisition and other cognitive functions are ripe for improvement / replacement.
Recursive Self-Improvement (RSI) A system that get better with time, as well as getting better at getting better.
I.J. Good “Since the design of machines is one of these intellectual activities, an ultraintelligent machine could design even better machines; there would then unquestionably be an ‘intelligence explosion,’ and the intelligence of man would be left far behind.” (in Speculations Concerning the First Ultraintelligent Machine, Advances in Computers, vol. 6, 1965.

What's in a Seed?

Size Of course this depends on the World/task-environment, in particular its complexity and richness (a world may be complex but only contain very few atomic element types, it may be simple yet rich in atomic element types - the latter not necessarily making it (much) more complex.)
Content Again, this is context/world-dependent, because at least some notions in the seed must reference at least some measurable+manipulatable features of the domain, via the system's senctors.
Domain-dependence Seed is always domain-dependent in that to reference a manipulatable+observable feature of the bootstrapping domain it must be known a-priori to the seed-designer.


/var/www/ailab/WWW/wiki/data/pages/public/t-720-atai/atai-16/lecture_notes_f-18_11.03.2016.txt · Last modified: 2016/03/10 18:07 by thorisson2