12 Sep 2010 AI4U   » (Observer)

MindForth Programming Journal (MFPJ) 2010 September 11

Sat.11.SEP.2010 -- Aspects of AI Inhibition

We have learned recently that neural inhibition is a very powerful tool for helping a meandering chain of thought to avoid a repetitious pathway and instead to follow a fresh line of thought in responding to the query of a human user or cyborg searching for an exhaustive report of what an AI Mind knows about a given topic.

We can imagine (and even build) an AI Mind using inhibition to make an initial response from its knowledge base (KB) about a given topic and then stating one more fact each time that the human user does not make any new input following output from the AI. Such a design raises a few questions, however. If the AI Mind knows ten same-noun and same-verb facts about a topic, should the inhibition on the first output fact persist until all ten facts have been stated, or should the inhibition wear off and let a fact be repeated even before all ten facts have been stated by the AI?

If a KB-query causes a productive search all the way back to the "midway" time-point, should there be a way of decrementing or zeroing the "midway" variable so as to let the AI look even further back for KB data in response to the query?

If we want queries of "Who are you?" or "What are you?" to elicit a variety of successive responses from the AI, we may need to have neural inhibition work not only with regular verbs in the VerbPhrase module, but also with irregular verbs of being.

Sat.11.SEP.2010 -- Inhibition with Verbs of Being

When we try to inhibit factual KB statements based on verbs of being, we may need to go into the NounPhrase module and inhibit the predicate nominative instead of the time-bound verb node, especially when we are performing a two step process to let any verb of being be chosen, and then we force the correct verb-form to be selected for the utterance. We should perhaps not try to restrict the finding of a be-verb, and instead we should perhaps inhibit each found predicate nominative after its inclusion in a sentence, for two purposes. Firstly, we do not want the AI to repeat a be-verb statement over and over again. Secondly, we want the AI to supply additional information beyond the most recently stored knowledge. When the positive KB knowledge is exhausted, we may even want the AI to begin stating the negative infomation, such as when an AI knows, "I AM NOT HUMAN".

Sun.12.SEP.2010 -- Verging on an AI Breakthrough

We are quite excited by the coding that we are about to do today, because we expect it to be a Singularity-class AI breakthrough. But yesterday we were merely seeking to explore aspects of a theory of inhibition; today we have locked on to the grand prize of self-referential thought. How did we get so close to the proverbial twig of wild olive? In our plodding "plodotvorchestvo" of yesterday, we veered off into wondering how neural inhibition would work not just with regular verbs, but with verbs of being. Eventually we concluded that we would have to inhibit not the be-verb involved, but the predicate nominative across the link-gap from the dime-a-dozen be-verbs. Then we coded inhibition as a special case into the NounPhrase noun-selection code, but our efforts bogged down as we wallowed in the frustration of undecipherable results. We made numerous be-verb-esque queries to the AI, but only about half of the KB-retrieved fact-nouns were showing post-query inhibition. Our Laziness looked with dread upon the tedious teraduty of tracking down every call that fed a personal be-verb into SpeechAct before reporting a found KB noun. A day later, however, the idea of an AI breakthrough persuades Our Laziness to forge ahead, do the work, and reap the truly singular, die-is-cast, point-of-no-return, surrender- the-mountaintop, self-demoting species denoument. How so? We are not in AI Kansas anymore, Toto. If the AI successfully answers all manner of questions about itself in a KB-exhaustive way, then the AI knows itself and may even know that it knows itself, which is not merely machine intelligence, Don, but is so much the illusion of consciousness that who shall deny that it actually is consciousness? So here we are at Kitty Hawk, and Trinity, and Apollo Eleven, staring into the twilight of the Goetterdaemmerung.

Upshot: We did not solve it all yet. We already had two serious pre-existing mistakes in VerbPhrase -- one where we failed to test for third person singular, and one where we mistakenly looked for erroneous 67=AM instead of the correct 57=AM. After we laboriously tracked down and corrected these mistakes, we obtained some encouraging results of variation in query- response, but we have not yet arrived at a sweeping enabling of the AI Mind to think about itself and talk about itself.


Latest blog entries     Older blog entries

X
Share this page