Stanley In my view, one cannot consider abstraction in isolation, but rather as a concept within a circular, abstraction/deabstraction value chain. Think, pure objects.
To further our general understanding of what I'm trying to communicate, perhaps we would be willing to consider that such a value chain is not forward, reverse, or re-engineering. However, it lends itself to engineering in the sense of employing an SDLC and applying method to science. I think for computational models, this is essential. The question is, how should such a component of abstraction/deabstraction be successfully engineered as a component of an AGI service? Considering what you shared about your architecture, I would understand you are saying that you would develop the blueprint by virtue of moving from an abstraction of information, to a codable (deabstraction) of said information "text". If so, that is in compliance with the classical SDLC where analysis and design moved from conceptual- to logical- to physical, architectural layers. Such a choice-based exercise is fairly straight forward. But it seems you were alluding to another aspect of abstraction as well, one that is less "simple", where information is "floated" as an outcome of the component in action, then "discovered" by the system (obviously by a generic, a scanning/locating function) to be used within the informational context of the overall system? In short then, would emerge another systems value chain; Making (in the sense of abstracting/deabstracting), Finding (locating the floated value), Framing (abstracting the floated value to systemic relative reality), and Sharing (Presenting the compressed value intra and extra systemically in feedback loops). Suppose this value chain I stated was slotted in architecturally as a vector function between your version of "abstractor" and "promoter"? Would it be possible to consider that a generic, "X Value Chain" would become the thread that all AGI functionality would flow from in an architectural sense? My suggestion would be that, within an AGI system, there exists is no more up or down progression within the system, but rather an architecture resembling an "instantaneous" informational starburst. I know, this would probably play havoc with linear and component-based programming techniques, as it should. I think the simplest reason for this is that AGI functionality probably is not a teachable construct. Would we need a different programming approach? Probably. Would this approach be derived from an AGI-services architecture? Probably. Which leaves the question; what would the SDLC for an AGI-services architecture look like? I'm meeting up with one of the ghurus in Systems- and Information engineering in South Africa next week to hopefully try and discuss this question. Mr. Viljoen pioneered IE methodology in the 1980's, which IBM Global acquired in the early 1990's for their future use. I hope to be able to share my views and be enlightened by his mind. Rob ________________________________ From: Stanley Nilsen <[email protected]> Sent: Monday, 29 October 2018 6:55 AM To: [email protected] Subject: Re: [agi] Abstraction is not simple On 10/28/18 6:14 PM, Jim Bromer via AGI wrote: ... thinking of how abstraction might be used to produce recognition. First of all, a useful abstraction might rely on an algorithm not only to get it out of a data (or a 'text') but the data or some characteristic of the data might need to be put through a transformation by the algorithm. Although this is not radically different it is a new way of looking at 'abstraction.' Secondly, different kinds of transformative abstractions of data-text might be needed to build a configuration of abstractions that could then be used in recognition and subsequent analyses. I think that is a new and interesting way of looking at the concept of abstraction. Jim Bromer Hi Jim, I believe that abstraction is important for the goal of making an intelligent entity, one that will recognize and make good choices with the facts. My idea of abstraction is fairly simple, probably incomplete. With that in mind, a few thoughts and examples: I'm thinking abstraction has to do with taking away detail - loss of information for a reason. For example, the IQ score is an abstraction. It doesn't tell you much about the person with a specific IQ. But it serves a purpose (perhaps) of allowing one to make a comparison of one person to another (dubious.) Encoding is more like what we do with text. The text requires that there be a device that decodes the symbols, and after decode, ends up with richer concepts. This happens because there is agreement about what the encoder and decoder do with the symbols. Abstraction is very different than decoding. It uses some formula to arrive at an end result, but the end result doesn't contain nearly the information that was originally involved in performing the abstraction. This is the beauty of the abstraction, and the beast in the abstraction. Beauty in that we have something simple to work with, but a beast if we try to use the abstraction for the wrong purpose. For an example of the beast of IQ, consider… We agree that a person's IQ reflects something about their ability to be aware of the meaning behind a few example questions or problems. This is good, but it isn't really a predictor of a person's suitability for any given task. When it comes to people, IQ doesn't tell you if a person has sustainable interest in a specific domain. Another example: Compare an apple and a peach. We want to totally abstract both objects and say which one is preferable. First problem is what does preferable mean? Lets say we can take one or the other but can only choose one. And lets say that we have an agent who acts as an intermediary. The operation involves the agent producing a number for one object and a number for the other object (fruit objects.) Using the abstracting process, the agent acquires the object with the higher number. Lets say we train the agent to know that we prefer peaches to apples, but then again, many times a store bought peach is far inferior to a ripe picked peach. So we start to give the abstractor rules… 1) if a peach, value starts as 1 2) if apple, value starts as .9 3) if fruit is spoiled, value is reduced to .1 percent of original. 4) if peach is tree ripened, value increased by 30 percent 5) if peach is store bought value decreased by 16 percent 6) if apple is tree ripened, value is reduced by 20 percent 7) if apple is golden delicious increase value by 5 percent … and so on, we accumulate a bunch of rules for the abstractor to use in the evaluation. A couple of things we can say about this process. First, the abstractor works best with lots of information - ONLY if the abstractor has rules for how to “use” the information. If the abstractor is sophisticated, it will take many things into consideration. Second, the considerations will be based on who the abstractor is working for – sort of a “knowing” the end consumer. A chef may look at fruit differently than a worm farmer or a pig feeding operation. Finally, it might be said that we have a good abstracting agent if it was trained with appropriate rules and ends up guiding our acquisition as we hope it would. And, yes I agree with you Jim, abstraction could aid in recognition. If one looks at recognition as being a process of making a choice between multiple possibilities. The abstraction phase could compare the known facts against a set of facts that are typical of each possible item. Rules would modify the abstraction number for each item and a highest ranking “winner” emerges. Notice that in the real world, the items to be recognized could be numerous. The abstractor would have lots of work to do to come up with an abstraction number for every object. This is where parallel processing saves the day. We distributed the facts to thousands of processors and each one has a dozen or so rules to consider and then “promotes” it's match number. The highest promoted number wins, we have our match. The “recognition” is made according to comparison of magnitude of numbers. The numbers are basically totally abstract, (with very little resemblance to anything,) but, that is what is needed to arrive at the choice. In my architecture for an intelligent machine, I separate the intelligence of doing abstraction from the moment by moment operation of promoting - promoters are simple components duplicated by the thousands and each one can only consider a few rules. The abstractor is the genius and does the programming of promoters. Stan Artificial General Intelligence List<https://agi.topicbox.com/latest> / AGI / see discussions<https://agi.topicbox.com/groups/agi> + participants<https://agi.topicbox.com/groups/agi/members> + delivery options<https://agi.topicbox.com/groups/agi/subscription> Permalink<https://agi.topicbox.com/groups/agi/T586df509299da774-M3dad61520150f51f36bf6fbf> ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T586df509299da774-M39cef6a1740fe30722c8e312 Delivery options: https://agi.topicbox.com/groups/agi/subscription
