RE: [agi] Embedding AI agents in simulated worlds
hi, One is that I wonder whether it's worth building into Novamente a pre- set predisposition to distinguish between 'me' and 'not' me. My guess is that this will emerge pretty simply and naturally. Some external observations will correlate closely with internal sensations (these are the external manifestations of me) and others will not. And, if external observation Y correlates closely with internal sensation X, and X correlates with internal sensation Z (that may not correlate with any external observation), then Y will come to correlate with Z [through one of a couple cognitive mechanisms]. A cluster of tightly-linked nodes should then emerge, corresponding to me-ish things (internal or external) So, if the cognitive mechanisms work as expected (which may require some parameter-tuning as it's a new application context, and the cognitive mechanisms are now tuned for datamining applications), no pre-set predisposition to distinguish me/not-me will be needed I think this is a relatively simple unsupervised learning problem, given a reasonable volume of data obtained from spontaneously acting and observing in a (real or simulated) world I would start by setting up in 'itch' to discover whether data flowing through Novamente is sourced from 'outside me' or from 'within me'. The second 'itch' would be to label data generated outside of 'me' as being to do with 'me' or to do with 'other'. I think a baby does the latter by noticing close correlations between internal feelings/intentions and seeing things happen 'outside' - I send messages to my arms/legs and I see objects move in a closely correlated way (turns out that as often as not I see my arms and legs move, etc.). OK, so what I'm saying is that Novamente will do this about the way you hypothesize a baby does it... Another itch that I imagine you already have built into Novamente is to try to find closely correlated streams of data - this would tend to speed the process of creating 'objects' or standard actions. The search for correlations is intrinsic to Novamente cognition, yeah It might be worth running a few experiments to see if it significantly speeds up learning for a Novababy to have the 'itches' about 'me' / 'not me' built in at the start. Well, if it sped things up, it would only be a short-term effect, because after a certain point, this distinction is just gonna be DONE and LEARNED, and then utilizing it will be basically independent of whether it was acquired thru generic cognition or through a specially-programmed itch ... Another thought is about the way you've split leaning into direct environmental learning, learning to be taught, and then learning symbolic communication. I think learning symbolic communication is inseparable from learnng to be taught. And direct environmental learning is inseparable from the precursors to speech. Let me clarify: These aspects of learning are not really separated inside Novamente, they are internally to be carried out by basically the same processes. However, I have separated them from the point of view of thinking about teaching and training Novamente, simply to jog my mind in a productive direction when designing teaching exercises... A long time ago I picked up at second hand a rather crude notion of the Piagetian stages of learning. What I absorbed was the notion that Piaget said that kids must first learn concretely before learning abstractly. This had a surface common sense ring to it, but I've now decided that I don't at all agree that concrete learning has to preceed abstract learning. I agree with Piaget, though I think his statement is only true in a statistical sense. Abstractions are generally built up hierarchically, so we have -- concrete observations -- level 1 patterns in observations -- level 2 patterns, in level 1 observations concrete observations -- level 3 patterns, ... -- etc. Learning level N patterns is only possible once a large body of level N-1 patterns has been mastered. Ascending the hierarchy takes time, because the more abstract the patterns get, the larger the space of possible patterns that must be searched in order to find the correct patterns. (Of course, a mind is not using a dumb search algorithm, the search metaphor is just being used for its evocative power.) So, I think that the process of streaming data into objects and actions and relationships and characteristics is already a proccess of abstract learning. Sure -- it's just on a lower level of the hierarchy of abstraction I've just described... The second reason for thinging that abstract thinking starts very early in human babies is that their primary carers are talking to them all the time - of using highly abstract notions like I love you, you georgeous little thing, how clever, oh, don't be messy etc. etc. The baby hears the words (jigsaw puzzle-like at first - ie. a fuzzy set of sounds in amongst other words they know) and then
Re: [agi] Embedding AI agents in simulated worlds
Hi Ben, I've just read your paper (Goertzel Pennachin) at: http://www.goertzel.org/dynapsyc/2003/NovamenteSimulations.htm I'm not expert in any of this - but I'm 10 years and three years into raising two kids so that gives me some experience that might or might not be useful I thought what you said made good sense. I've got two suggestions for modifications to your approach. One is that I wonder whether it's worth building into Novamente a pre- set predisposition to distinguish between 'me' and 'not' me. I would start by setting up in 'itch' to discover whether data flowing through Novamente is sourced from 'outside me' or from 'within me'. The second 'itch' would be to label data generated outside of 'me' as being to do with 'me' or to do with 'other'. I think a baby does the latter by noticing close correlations between internal feelings/intentions and seeing things happen 'outside' - I send messages to my arms/legs and I see objects move in a closely correlated way (turns out that as often as not I see my arms and legs move, etc.). Another itch that I imagine you already have built into Novamente is to try to find closely correlated streams of data - this would tend to speed the process of creating 'objects' or standard actions. It might be worth running a few experiments to see if it significantly speeds up learning for a Novababy to have the 'itches' about 'me' / 'not me' built in at the start. Another thought is about the way you've split leaning into direct environmental learning, learning to be taught, and then learning symbolic communication. I think learning symbolic communication is inseparable from learnng to be taught. And direct environmental learning is inseparable from the precursors to speech. I'll explain what I mean. A long time ago I picked up at second hand a rather crude notion of the Piagetian stages of learning. What I absorbed was the notion that Piaget said that kids must first learn concretely before learning abstractly. This had a surface common sense ring to it, but I've now decided that I don't at all agree that concrete learning has to preceed abstract learning. I have two reasons for thinking this. When babies are in the ealiest stages of development I think they face the hardest possible learning tasks - they are getting a staggering stream of sensory input data - most of which is meaningless. Out of this they have to sift signals that are meaningful - so from the minute their brains are able to process input data (while in utero) they engage in abstract learning - take the stream of raw data and abstract from it...so a tight coupling of environmental experience and abstract thinking is required form the first moment of mental capability. Babies are undoubtedly pre-programmed to be alert to certain patterns of data. This might be useful to get the baby responding in ways that helps its immediate survival. But it might be that the pre-programing sets up a process of awareness crystalisation - certain streams of data can be treated as meaningful - and then out of the soup of other non-meaningful data additional correlations to the currently meaningful data can be developed - a bit like the way we do jigsaw puzzles. So, I think that the process of streaming data into objects and actions and relationships and characteristics is already a proccess of abstract learning. The second reason for thinging that abstract thinking starts very early in human babies is that their primary carers are talking to them all the time - of using highly abstract notions like I love you, you georgeous little thing, how clever, oh, don't be messy etc. etc. The baby hears the words (jigsaw puzzle-like at first - ie. a fuzzy set of sounds in amongst other words they know) and then over time they associate behaviours, feelings, settings, other known words, etc. that invest these abstract terms with more and more meaning. But the abstract symbol comes first and the meaning later. The words are like pegs to hang meaning on. Given these ways of seeing things, it's not hard to say that 'learning to learn from a teacher' is already a process of symbolic learning. If a robot is circling another object and hoping the NovaBaby will realise that it wants the NovaBaby to go and get the object (or whatever) then it is teaching symbolic communication. But it's just doing it in a way that a mute person would teach it or the way that it would have to teach it to a deaf child. This form of teaching is no less abstract that the use of verbal symbols and it is no easier to learn (might even be harder as the action might not correlate so uniquely to the symbolic meaning that the teacher is trying to convey). My first child started speaking at 8 months and he was clearly understanding words long before that - so my guess is that symbolic reasoning starts very, very early - and that language take-off is more to do with
Re: [agi] Embedding AI agents in simulated worlds
Ben.. For the vision processing stuff, here are some resources that you may be familiar with: Intel's open source vision library offering most standard vision functions in C: http://www.intel.com/research/mrl/research/opencv/ Sourceforge open vision project: http://sourceforge.net/projects/opencvlibrary/ --Kevin - Original Message - From: Ben Goertzel [EMAIL PROTECTED] To: [EMAIL PROTECTED] Sent: Monday, August 18, 2003 12:48 PM Subject: [agi] Embedding AI agents in simulated worlds Hi all, Finally, I'm resuming a thread I started on this list a few weeks ago, before I left for the IJCAI conference... At the following URL, you will find some thoughts of mine Cassio's on hooking Novamente up to a simulated (virtual) world and teaching it therein: http://www.goertzel.org/dynapsyc/2003/NovamenteSimulations.htm Probably the most interesting parts are at the end, where we touch on artificial developmental psychology. This is not work we're doing right now, but it's stuff we're gradually moving towards as we refine and test our collection of cognitive algorithms on various sorts of data. Comments desired! -- Ben G --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED] --- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]