I really cannot figure out what you are saying Steve. And I don't know how what you are saying could actually apply to me. "...there could be NO way to understand and/or debug such a thing," sounds like it might have some relevant meaning but then to interpret it I have to go back to, "without this model," which refers to "constraining learning to ONLY learn things that fit this model." So my best guess is that you are saying a 'mathematical method cannot be learned by a computer program because there is no way that it could be debugged, but that does not really make much sense. A mathematical method is a way to compute something. An abstraction - even a dynamic abstraction - is typically going to refer or work with a class of particulars or data objects (that are subject to the dynamic abstraction as operands are subject to a program step.) If my guess about what you are saying is in the ball park, then I can put it another way. Can a computer program learn a sub-program (from the IO data environment)? Yes. Whenever a computer program adapts to the IO data environment it is rearranging its program. This is true even for a word processor. My gmail program is different than yours. What I mean is that it is -effectively- different than yours. It does not matter that the 'program' is 'separate' from the 'data'. From this technical view, one can make a small shift and think about some available sub-programs being rearranged in response to the user input. To make the program robust this rearrangement should not be able to introduce a bug into the program. However, it is possible that such a system could be refined via learning so that it could adapt to a greater variety of IO events of the 'kind' that it had learned to respond to. It could go through a process of refinement according to some goal for the different events of the 'kind' it was reacting to. There is the possibility that such a program could just go out on it's own, rearranging the sub-programming to fit the data and then fitting the data to the rearrangement of the sub programming - but that is just a complication, not an insurmountable impossibility. I am not sure I know what you were saying, but it is funny, as I tried to write this out I started thinking about feasible programs that could have some of the characteristics that I was thinking about. I mean that I was thinking of the possibility that different programs could be written that would demonstrate what I was trying to talk about with simple examples and as the examples got more sophisticated variations of the programming could be introduced to demonstrate the slightly higher level of sophistication. What I am getting at is that these ideas could be tested in highly controlled tests. If I was working for a software company I might propose writing some of these controlled tests in order to better study the kinds of things that I am talking about. Jim Bromer
On Wed, Jun 12, 2019 at 9:55 PM Steve Richfield <[email protected]> wrote: > Jim, > > It is (nearly?) impossible to "learn" in a way that preserves value (e.g. > 50%), dimensionality (e.g. probability), and probably significance (e.g. > +/-10%) without constraining learning to ONLY learn things that fit this > model. Without this model, it is just numerology that can NEVER EVER be > made to work - because there could be NO way to understand and/or debug > such a thing, either through automation (deep learning) or manually (as I > have tried). > > Steve > > > On Wed, Jun 12, 2019, 6:42 PM Jim Bromer <[email protected]> wrote: > >> The 'formal' part of the system can be acquired through learning. >> Jim Bromer >> >> >> On Wed, Jun 12, 2019 at 9:14 PM <[email protected]> wrote: >> >>> Yeh Steve - maybe that helps it try novel situations better??? >>> newsflash from me -> i think that formalizing the system manually ends >>> up a shallower system than what needs to be there for a developing system. >>> because its cheating it to do things, is it where the term "deep >>> learning" comes from? >>> >> *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/T395236743964cb4b-M935ffba0b4b0bfcf42422818> > ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T395236743964cb4b-M1474d9a9c8d132b6f9ac78ff Delivery options: https://agi.topicbox.com/groups/agi/subscription
