On Mon, Sep 24, 2012 at 12:01 PM, Ben Goertzel <[email protected]> wrote: > A description of Ruiting Lian's current NLP development work, in the > OpenCog HK team, is here: > > http://wiki.opencog.org/w/Link2Atom
Thanks. I was not aware of this. However, it would be nice if the rules for mapping link parses to atom sets didn't have to learned from a manually created set of pairs. If you have to create this data, you'll probably end up coding the rules directly because it's faster. But there really needs to be an algorithm for learning language just from lots of raw text. I know it's possible because we all do it. >> There are several people working on development, but like any large >> software project, a lot of the work is bogged down on fixing bugs, >> porting issues, and updating the documentation. > > That's not really true at all, I'm sorry you have that false > impression ;p .... OpenCog work is difficult and can be slow, but a > small minority of the developers' time is spent on the things you > mention. I would prefer if you would stop making statements like > that; you're not part of the OpenCog development team and you really > don't know what's going on in that context !! I guess my impression was biased by the number of emails along the lines of "Help! I tried to build OpenCog and got all these errors...". I guess when things go right, they don't say anything. >>There has been some discussion of making a distributed >> version of AtomSpace but IMHO there are going to be severe performance >> problems that make scaling to large numbers of processors impractical. > > Can you provide detailed technical reasons for this opinion, based on a > specific > critique of the proposed design for a distributed Atomspace, which is > described > in the PDF attached to the following wiki page? : > > http://wiki.opencog.org/w/DistributedAtomspace > > Concrete criticisms on your part might help us improve the design. > > General criticisms as you've made, are not very useful. Again, my impression was from discussion of performance issues on the OpenCog list. Can you really tolerate 1/2 second delays or even 1 microsecond delays for what would just be a memory-read on a single processor? Also, how would you measure performance? What is your test application? As you know, I would divide up the work differently. I have a different view of AGI. It is not single human level intelligence that you have to compete with, but large organizations of people. I would divide the work by giving each processor a highly specialized task and having them communicate over the internet by routing messages in natural language text. Each peer would have a small vocabulary and only understand messages that were relevant to it. It is not just the processing that is distributed, but also the software development, administration, policy, and meta data or indexing (X knows that Y knows about Z). Because the task is huge, you need to bring in lots of people and give them an economic incentive to join the project, so that you mutually benefit from their work. Because you may not trust everyone on the internet, you need a protocol with cryptographic key exchange and signing so that intruders can't forge messages from trusted peers. (I am sure you saw this before: http://mattmahoney.net/agi2.html ). So I would say with OpenCog, keep it on a single processor and divide the work at a high level. Depending on the task, AtomSpace might not be the most efficient implementation. I realize that AtomSpace is designed to be very general form of knowledge representation, and that producing lots of specialized algorithms for specialized tasks is far more complex. Yes it is. Also, I realize it is more fun to work on AGI as a whole than to be part of a large, loosely coupled, organization that is building AGI while you work on some tiny narrow-AI part of it. Yes, that's how it is to work for a big company too. And it would be nice to be CEO of this company, but you can't because this organization has nobody in charge. -- Matt Mahoney, [email protected] ------------------------------------------- AGI Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393 Modify Your Subscription: https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968 Powered by Listbox: http://www.listbox.com
