Thanks for the link to Ahmed and Hawkins' paper on sparse distributed representation as it relates to HTM. It is interesting but it is taking me some time to read it. I also have started the white paper on HTM. I have seen that before but I never read it very carefully. I always thought that any kind of conceptual vector system (like a semantic vector system) would have to be n-dimensional and that it really would not map well to three dimensions. And if you wanted to use a simple vector to represent every basic concept (or every basic concept zone) then it would have to be a really big vector which would be, for the most part, blank. So the idea of a SDR does appeal to me. I am skeptical of the idea that, 'since the HTM is based on neuro-science then getting it to work is just a matter of finding the right application details' intuition. If HTM is a good model then it would be, at best, a crude model of a detail of the mind. However, it is an interesting concept and it does seem to be based on neural networks which have shown some effectiveness and it does seem feasible that it could be effective in certain areas like reconstructing an appropriate reaction to a kind of input based on partial inputs. This, for example, would be more feasible for a temporal stream of input where the interference is scattered across each 'frame' of the input. (I think research into aspects of the theory, like this one, might be worthwhile even for a skeptic.)
I haven't seen how it uses the characteristics of combinatorics that you mentioned. I just lost access to the pdf of the paper for some reason. (Cornell is either on to me or they are doing something.) Jim Bromer On Mon, Jun 8, 2015 at 9:35 PM, Matt Lind <[email protected]> wrote: > Jim: > > Agree. An interesting concept in this regard is SDR (sparse distributed > representation) which heavily relies on those characteristics of > combinatorics. Numenta's HTM (hierarchical temporary memory) framework uses > SDRs to model the brain's internal communication language. Interesting read: > http://arxiv.org/abs/1503.07469 > > Matt > > > > 发自我的iPhone > > > ------------------ 原始邮件 ------------------ > 发件人: Jim Bromer <[email protected]> > 发送时间: 2015年06月09日 02:50 > 收件人: AGI <[email protected]> > 主题: [agi] Combinatorics Is Important to Use With Discrete-Based AI > > I have not been that interested in combinatorics, graph theory, > weighted reasoning, Bayesian methods and so on because I felt that > these were all logic based and that the simplest way to develop an AGI > application would be to stick to basics and only use simple forms of > these methods if I needed them. I believed that there are certain > discreet obstacles to advancing AGI at this time and solutions to > these problems really needed to be found before any substantial > advances could be made. However, now that I am starting to appreciate > how combinatorics can help with discreet based systems and may even be > essential to using logic effectively, I am, unsurprisingly, changing > my view about this. > > So I recommend that anyone in this group who feels as I felt should > start brushing up on combinatorics, graph theory and so on. > > Many of us know that computers are good at working with narrow results > and this has produced major advances in ‘narrow AI’. Developers who > worked with probability reasoning thought that their methods should be > sufficient to overcome the narrow AI dilemma, but I haven’t seen that. > (I do appreciate the many advances that AI has made and I do not > totally accept the narrow AI dismissal, but on the other hand I am > still amazed that AI programs do not seem to exhibit the ability to > learn even in the simple ways that we would expect from toddlers.) > > So why are combinatorics so important for discrete symbolic AGI? Using > logic as an example of a tightly defined discrete system, you can see > that multiple solutions can be held in a tightly compressed formula > string. What is more, subclasses of these different solutions can > often be derived from a portion of the formula or from a simple > analysis and rearrangement using portions of the formula. > Combinatorics is one way we can get to this logical multiplicity as it > may relate to logical sub-relations. Right now these methods are > limited but there is no reason to expect that this method will be > stuck with its current limitations. > Jim Bromer > > > ------------------------------------------- > AGI > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/27149679-a7635131 > Modify Your Subscription: > https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com ------------------------------------------- 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
