1. Discovering and understanding arbitrary relationships between behavior of entities and their environment over time. Holland, Holyoak, Nisbett and Thagard wrote Induction in which they postulated that inducing synchronic and diachronic associations were the essence of learning. They modeled these kinds of temporal relations with Classifier systems. Numenta performs spatial and temporal pooling (same thing) in their Grok product. The Schema Mechanism by Gary Drescher synthesized new arbitrary relationships observed in its environment over time as synthetic items (or meta schemata). There are many examples of systems that form arbitrary relationships between entities and environments, and there are many examples of systems that track the behavior of entities. I don't see anything special about what you are attempting.
2. This is circular reasoning. Things that are “interesting” are a product of observing all of reality and identifying things out of the ordinary, where “ordinary” is defined as commonly observed patterns over time. The spatiotemporal context is not just your database, it is your entire computational data model. Even if you use reality for hypothesis testing, a good hypothesis is hard to develop in a vacuum. Also, you can’t just “look things up” in reality; the data has to streamed into computers and indexed so that a computer can find and analyze it. First of all, you are NEVER, EVER, going to observe ALL of reality. That is imposible. So your first assumption is way off base. Human beings have limitations on their sensory inputs. They do not see the entire lightwave spectrum, nor hear the entire range of soundwaves nor feel temperatures from zero degrees kelvin to 1,000 degrees kelvin. So observing ALL of reality is a naive notion. You take slivers of reality, and process them. You take visual and audio frames, and so forth. Once you take a sliver, you can create and recognize patterns within your samples of reality. These patterns are your entities. There may be regular or irregular entities within your samples. But you will still store some of these patterns and ignore others. That is our nature as human beings. Even savants with eidetic memories do not store all modalities of memory. Their visual memory may be acute but their sonic memory may be deficient. So your requirement for computers to do what people do not do is also unrealistic if you are shooting for AGI. Finally you DO just look things up in reality. For example you take an image of a room and see a cat sitting on a bed. You take a later image of the same room and see the cat is no longer on the bed. That is "looking things up", reality checking, to ascertain your state, or current situation. 3. It turns out that subsets of the physical world do not even approximate a closed system. Models built from closed system assumptions suffer from chronic “black swan” events in practice unless they are limited to very narrow use cases. You mitigate this by fusing as many data sources as possible into a single spatiotemporal context. (Also required for data quality reasons.) I don't think you've fully baked what your learning algorithm(s) is(are) for an AGI. I suggest you do some more research before making sweeping statements that only reflect your lack of knowledge, or lack of an approach. If some things can't be done, then don't do them. Do what is feasible and computable and make strides there. Show a proof of concept, for God's sake. Let's see a prototype of what you CAN do. ~PM From: [email protected] Subject: Re: [agi] Re: Could Brain Emulation be NP-Hard? Date: Thu, 25 Jun 2015 14:51:20 -0700 To: [email protected] On Jun 25, 2015, at 12:49 PM, Piaget Modeler <[email protected]> wrote:What kinds of reasoning about the physical world are you aiming at? Discovering and understanding arbitrary relationships between the behavior of entities and their environment over time. You can escape the frame problem by not trying to model reality in its entirety. Use reality as a database look things up when needed, only track what you're interested in. This is circular reasoning. Things that are “interesting” are a product of observing all of reality and identifying things out of the ordinary, where “ordinary” is defined as commonly observed patterns over time. The spatiotemporal context is not just your database, it is your entire computational data model. Even if you use reality for hypothesis testing, a good hypothesis is hard to develop in a vacuum. Also, you can’t just “look things up” in reality; the data has to streamed into computers and indexed so that a computer can find and analyze it. Use open and close world assumptions advantageously. It turns out that subsets of the physical world do not even approximate a closed system. Models built from closed system assumptions suffer from chronic “black swan” events in practice unless they are limited to very narrow use cases. You mitigate this by fusing as many data sources as possible into a single spatiotemporal context. (Also required for data quality reasons.) AGI | Archives | Modify Your Subscription ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657 Powered by Listbox: http://www.listbox.com
