The role of self supervision in artificial intelligence (AI) approaches may be the most critical aspect of emerging next generation research.

Supervision often implies biologically unrealistic interference using some form of external state machine, monitoring and conditioning by supplying desired outcomes and thereby directing and constraining ongoing learning.

While supervision has likely been necessary for the vast majority of software based implementations so far, it is a long ways from the far more desireable situation of biologically plausible self supervision.

From a preliminary perspective we might infer that biological senses as neural nets are unsupervised, such as the hominid retina.

Apparently though a retina may have some internal compression or enhancement effects, they are for the most part simply reacting blindly and continuously to input energy, translating it to an embedded format for transport between and through downstream networks.

Self supervision is implied in the phrase "self metaperspective" along with an inferred overview of factors of potential importance to the supervision function.

One might envision a large feedback loop of special components that replaces the external supervision state engine.

Preliminary considerations involve what such a self supervision system might be tasked with and how it might complete such.

The implied meta mechanics of the supervision system will be able to derive some form of overall view and possibly, optimistically, to decide on that basis what to do next.

There must be some mechanism for creating the meta view and discerning or comparing alternatives based on that view, in such fashion as to shift emphasis to some potential future meta view of preferred form.

Similarly running projective (future time line) comparitive elements leading to a set of preferred views might then serve as a trial and error framework for actions leading to enhancing the probabiity of a prefered meta view.

This might in lieu of directed purpose of the traditional form in hominids, be termed a type of virtual purpose, based on entirely mechanical probabilistic behaviour.

It seems overall that one might like to develop a distributed learning system that pushes the envelope.

To some extent, mimicry of "human-like" or "sentient", "naturally evolved" reasoning capabilities, are implied:

• "reason" in a variety of ways, using substantial amounts of appropriately represented knowledge
• learn from experiences so that performance gradually improves as knowledge and experience accumulates
• explain "itself" (dynamic self context metaperspective)
• "accept" direction (identification of authoritative influence sources - integration of influence into self context awareness model )
• "aware" of "own" behaviour (self context awareness)
• reflect on "own" capabilities (integration of self context potential model)
• respond in robust manner to surprises (very rapid learning)