[YKY]

Hello,

I'm new to this forum but I've been working on AGI (artificial general intelligence) for 8 years.  My project is called Genifer and it's opensource, hosted on Google Code.

This is the project web site:
http://code.google.com/p/genifer/

I and some partners have written a free online book on AGI theory, and many sets of introductory slides that can be downloaded here:
http://code.google.com/p/genifer/downloads/list

This is a 4-minute introductory video to Genifer on YouTube:



I hope to share ideas with this forum, and find more partners.  I'm very open-minded and our project welcomes everyone to join, contribute, commercialize, modify, etc.

I'm from Hong Kong, China, where we have a lot of funding and not enough expertise in AGI.  Currently we're looking for someone to work on logic-based (ie relational) inductive learning.  Contact me if you're interested!

 
YKY

[Bragi]

Hi YKY,
interesting project. Is there perhaps a demo/beta or the like that can be tested? You say you have funding? lucky you.
I'm working with neural nets myself, no funding, solo project.

[Freddy]

Welcome aboard YKY 

[claude2]

Good job! Perhaps a new Hal!

[YKY]

Hi YKY,
interesting project. Is there perhaps a demo/beta or the like that can be tested? You say you have funding? lucky you.
I'm working with neural nets myself, no funding, solo project.

Yes, I've written a partially completed prototype in Lisp, and it can be run on any Common Lisp implementation.
The prototype currently has a small test suite of ~15 logic examples.  It's not very impressive.
You may have to check out the Lisp source, branch "fuzzy", directory /hg/lisp/fuzzy.

Or, if you just want to satisfy your curiosity, here's a screenshot of the test output:


 
YKY

[YKY]

PS:  I am not having any funding now, just sustaining this on my own.  What I mean is I think I can find funding relatively easily, if I can find 1 or more partners to work on the project.  Right now I'm nearly solo with some volunteers helping me informally, so I don't need external funding yet.

[Bragi]

so, this is mostly based on a 'prolog' type of declaring logic rules? Could you perhaps give an example of a simple rule in your system?
how do you handle the natural language processing?

PS: I read you also did some experimenting with neural nets, but found some 'problems' with them. I can relate. I solved this by using a different type of network, based on resonance (my own flavor).

[YKY]

so, this is mostly based on a 'prolog' type of declaring logic rules? Could you perhaps give an example of a simple rule in your system?
how do you handle the natural language processing?

PS: I read you also did some experimenting with neural nets, but found some 'problems' with them. I can relate. I solved this by using a different type of network, based on resonance (my own flavor).

An example rule would be:
    has_beard(X) -> male(X)
but this is not a rule in a realistic KB due to various complications:  eg, the rule would be fuzzy-probabilistic, and it would be further broken down into atomic concepts such as "have", "beard", etc.  It is hard to state a rule in such details, but we don't need to care because all the rules would be learned by machine (and we can provide initial rules as seeds).

Natural language:  Genifer does not need a separate NL module.  All of natural language can be learned via inductive learning.  Actually, when the core logic engine is ready, I plan to launch a campaign to "teach Genifer English in 3 days" as a publicity stunt =)

Neural networks:  yes, it is possible to modify traditional ANNs to make it more powerful and expressive, but a problem is that the approach seems ad hoc and it may be hard to convince others to use it.  When I found out that traditional ANNs are inadequate for AGI, I switched to logic because it's more expressive and is the other most mainstream approach beside ANN.

It might be possible to find out the formal relation between your approach and the logic-based approach.

I also discovered that one can perform machine vision using logic, but it'd require a feature-extraction layer before logic takes over the processing.  And that layer may be implemented using ANN.

 
KY

[Bragi]

It is hard to state a rule in such details, but we don't need to care because all the rules would be learned by machine (and we can provide initial rules as seeds).

Do you already have initial rule seeds?

Natural language:  Genifer does not need a separate NL module.  All of natural language can be learned via inductive learning.

Can it already do something with text input?

Actually, when the core logic engine is ready, I plan to launch a campaign to "teach Genifer English in 3 days" as a publicity stunt =)

That would be pretty impressive. How would one go about teaching it?

Neural networks:  yes, it is possible to modify traditional ANNs to make it more powerful and expressive, but a problem is that the approach seems ad hoc and it may be hard to convince others to use it.  When I found out that traditional ANNs are inadequate for AGI, I switched to logic because it's more expressive and is the other most mainstream approach beside ANN.

When I say resonating neural nets, it's best to first forget everything you know about ANN. I don't work with input - output and hidden layers. It's more akin to a database system. The basic idea is that, when an impulse needs to be processed, other parts of a neural network determine the path to take, not the neuron itself (that's the resonance bit). In the end, this comes down to being able to 'program' neural nets.

[YKY]

Do you already have initial rule seeds?

No, (but see below)....

Can it already do something with text input?

No again 

That would be pretty impressive. How would one go about teaching it?

Through a combination of techniques, but mainly, by encoding rules of syntactic and semantic parsing as [higher-order] logic formulas.  It's easier than it sounds, because we can seed the system with simplistic rules at first.  Also, we can rely on inductive learning to learn from English examples, without the need to enter logic directly (so anyone who knows English can help).  Such initial rules will gradually evolve to become more and more accurate.

When I say resonating neural nets, it's best to first forget everything you know about ANN. I don't work with input - output and hidden layers. It's more akin to a database system. The basic idea is that, when an impulse needs to be processed, other parts of a neural network determine the path to take, not the neuron itself (that's the resonance bit). In the end, this comes down to being able to 'program' neural nets.

You may take a look at my idea of distributive inference, where a network of computers cooperate to perform logical inference (and learning etc).  It's not a big stretch to view a single logic formula as similar to a neuron.  In fact, a neuron when represented in coordinate space is just a half-space cut out by its equation as a hyper-plane.  And such a region is similar to a region in a Venn Diagram (ie Boolean algebra).  Thus, a neuron is like a proposition in logic.  But predicate logic makes it more powerful than a neuron because it allows variables.

I'm still making the slides about my distributive inference idea...  will upload it in a week or so...
 
KY

[Art]

Welcome aboard YKY. Sorry I missed your initial posting but work kept me busy.

It's always nice to see new members with fresh ideas. Good luck!

[YKY]

The time is just right

I am starting to explore neural-symbolic integration ideas and looking for collaborators

[Freddy]

Hi YKY,

Welcome back