When someone talks about deep learning they are talking about some kind of method that 'connects' weights and other functions derived from the weights all of which are calculated using standard arithmetic. Addition, multiplication.. I am not dismissing those ideas, in fact, I want to think about creatine something somewhat similar. However, I believe that it is worth exploring ideas that work a little differently. Instead of using arithmetic, which relies on a few simple rules, how about an arithmetic that relies on hundreds or even thousands of rules (abstractions) that have to be learned or acquired along with the data objects that the mathematical objects refer to. What is wrong with this idea? Those kinds of mathematical methods are not as flexible as traditional arithmetic. As an amateur mathematician whenever I try something different it never seems to work as smoothly as traditional arithmetic. Other people have thought about things like this. It is not original. All that makes me a crackpot. (Thank god!) But my ideas are unique in this group and it is pretty safe to say that they are unique in much larger groups as well. There may be some people who are trying something similar but it is very unlikely that we are going about it in just way. But maybe it is a little like Occam's Razor. Make it no more unique than it has to be. I just realized that your Occam criticism was not completely off. In one way I am saying let the AI (or semi-AI) program produce as many abstractions as seems useful, and then following Occam's Razor, pare it down closer to using only what is necessary. (Occam's Razor looks like an idealism which means it is therefore imperfect.) Jim Bromer
On Wed, Jun 5, 2019 at 10:44 AM Jim Bromer <[email protected]> wrote: > I did not mean to be overly critical in my last email. I was just trying > to be objective, to the best of my ability. > A major underlying goal in discrete AI has been to look for a simple set > of rules and discrete object types which could be used as a basis for all > knowledge. I believe these goals were driven by a few historical events. > The importance of arithmetic and logic, both of which could be derived from > a relatively simple set of rules; the value of arithmetic to humankind > which could be developed with a minimization of rules; and the severe > constraints on computer technology in the old days. I think that the > learned development of many rules may be achieved with modern computers but > that presupposes that there is more than one level of abstraction. So my > ideas about AI are based largely on discrete methods, but the abstractions > of the system have to be acquired and learned alongside other kinds of > knowledge (like facts and how-to stuff), and made to fit the knowledge that > is learned. > No one in this group talks this way about this stuff. Of course, a critic > can say that is implicit in someone else's ideas - except for the stuff I > am getting wrong. To put it a little more accurately, the necessity of > acquiring 'abstractions' along with 'knowledge' 'endpoints' (so to speak) > is implicit in many if not most AI theories. So how is mine original the > critic might ask? This would be an example of a dismissive argument. My > theory is my own because other people do not bother to talk about stuff > like this. So it may not be startlingly 'original' but it is my own theory > (it is not based on some white-paper or something that someone else has > talked about in these groups. You can find theories about 'abstraction' but > there is less around about the necessity of developing abstractions as > knowledge is developed and even less that suggests that these relationships > might be expressed mathematically - at least to some extent. > While I am interested in developing a fast mathematical index into > knowledge (based on discrete-based knowledge) I now believe that a great > deal of meaning can be -learned- and baked into the mathematical indexing > system. For example, mathematics does not have to be restricted to a > narrowing process (that produces a single numerical result) but it can also > be an expanding process (that produces a set of results that might be > expressed numerically.) > Again, a critic can take a dismissive attitude and might say, for example, > that fuzzy logic or any number of weighted methods can do that! Ok, but I > could defensively reply that I am interested in mathematical references to > sets that could (in theory) (typically) be derived from the values. > Radical originality is not one of my goals. However, individual uniqueness > is, because I believe that individuality is needed to develop useful > computational methods that have yet to be developed. There are many unique > aspects of my theories and the challenge is to find the theories which can > be made to function as a whole and to eventually develop computer programs > that will show their usefulness in future AI and AGI programs. > Jim Bromer > > > On Tue, Jun 4, 2019 at 6:07 PM <[email protected]> wrote: > >> Ok, sorry about that misunderstanding. >> >> You need to form the model of the of the working area of the a.i. , its >> true for us all, but how you do it, is what makes what your doing original. >> *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-M0c0dd3364303694d41da54b7> >> ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T395236743964cb4b-M52d67b75d1d7a53f63d7ccfe Delivery options: https://agi.topicbox.com/groups/agi/subscription
