Ben, On Mon, Dec 24, 2012 at 9:53 PM, Ben Goertzel <[email protected]> wrote:
> >> However, I still believe that function needs to be learned or discovered, >> rather than programmed. >> > > I think that's too crude of a dichotomy.... Every digital AGI is > program, and an analog AGI would be wired together by humans.... So there > will always be some "programming" for any engineered intelligence. I'm > unclear where you are trying to draw the line between "programming the > ability to learn X" versus "programming a bias regarding how to do X", > since every finite-resources learning algorithm has some intrinsic > biases.... But I suspect that if you did articulate where you are trying > to draw that line, I would want to draw it somewhere else ;) > I'm not so sure. Some things are "built in", like our 4-D world and our 2-D eyes into it. Within that world are objects, some of which are fixed and some of which are not. On the other side, do those objects have arms and legs? How many? Do they walk on horizontal surfaces? Here we are getting into details that are probably better discovered than programmed. I understand the arguments for programming - it is probably faster than learning, doesn't incorporate residual defects of learning, etc. However, the main argument against learning is that no one knows how to learn these things, which is probably at least in part residue from not knowing how to perform ML at real-world speeds. There is a list of capabilities that apparently come from a single source, that there is simply NO WAY to build AGIs without overcoming. A few items on this list include: 1. Fast, i.e. single incidence, temporal learning. 2. Ability to learn almost anything as fast as human and animal babies can learn. Note that babies can learn some complex behaviors on their very first day out, e.g. to emit one yell and quietly await feeding, and open their mouth when their nose becomes blocked. 3. The ability to perform model-based reasoning, where the models are learned. 4. I suspect that you could add other items to this list. Once the recipe for this "secret sauce" is known, then it is just an engineering decision how much to program, and how much to learn, which I think was the point you were trying to make. However, until this secret sauce is known, AGIs must forever remain elusive because they will be unable to form comprehensive models, learn at human speeds, etc. What I have been (erroneously?) hearing is something like "We don't have the secret sauce (to learn efficiently), but if we decide to program most of the functionality we won't even need the secret sauce. I see this as foolish for two reasons: 1. Such systems could only ever become AI toys whose capabilities are FAR short of human. 2. There is an immense number of unique functional details in every environment that, if not learned, will consume an astronomical amount of effort to program. So, as I see it, everything boils down to the finding the secret sauce. I can't imagine amount of twiddling with code ever filling in for the absence of the secret sauce. Once you have the secret sauce, there will be no further arguments over how much is programmed and how much is learned, as functionality ends all such arguments. > > >> A human mind is not a tabula rasa, and nor need an AGI mind be... >>> >> I would like to wring out your thoughts on this. It seems obvious to me that many people think via RADICALLY different processes. Some are purely intuitive with hardly any "rational" process. Some have multiple personalities, etc. Myself, I solve complex problems by a completely unknown process, wherein I first understand a problem, then minutes, hours, days, weeks, or sometimes even years later, the solution pops into my mind fully formed. I just don't see how something could be "programmed" into systems that develop SO very differently. How do you see that this could happen? > >> OK, so here is the statement modifies as per our last go-round: >> >> "any prospective AGI platform absolutely **MUST** be capable of *learning >> at biologically comparable speeds to* perform substantially all of the >> high-level cognitive information processing functions that have been >> observed in human mind/brains *without carefully ignoring functions that >> appear incompatible with the platform, like spontaneously discovering the >> nature of the world in which they 'live'.*" >> >> Is this OK yet? >> >> > While I agree that learning is critical to AGI, I don't agree that > cognitive functions must necessarily be learned without benefit of > pre-programmed capabilities or predispositions.... I don't think the brain > carries out this sort of tabula rasa learning either... > > My view would be better summarized as something like > > "any prospective AGI platform absolutely **MUST** be capable of acquiring > (via learning, combined potentially with in-built capabilities or biases)* at > biologically comparable speeds the capability to* perform substantially > all of the high-level cognitive information processing functions that have > been observed in human mind/brains *"* > > > I have deleted "*without carefully ignoring functions that appear > incompatible with the platform, like spontaneously discovering the nature > of the world in which they 'live'.*"" as that is redundant with the prior > portion of the summary... > With exceptions noted above, we are getting pretty close. Maybe if I better understand what you see is involved in "learning, combined potentially with in-built capabilities or biases" I can polish off the paragraph. I suspect that some examples might help. Thanks for sticking with me on this. Steve ------------------------------------------- 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
