1. A nueral network is in fact a turing machine. It is made up of several very simple components: a floating point multiplier and adder, an accumulation buffer, and whatever you'd like to plug in for a transfer function, all multiplied several times over. These are all easily designed "chunks" of a Turing machine (Mmmm Fundamental Theorems of Computing class...). Just because a program is non-linear in nature does not turn it into something that is not a Turing machine.
2. The halting problem is one people can solve. You can supply the problem as a GAC. The mindpixel approach can not solve it -- They claim that it can handle problems that are similar, but different from existing ones in its database, but neural networks would not be able to solve them. Rendering one of their claims in need of clarification.
4. You can indeed pull apart neural networks to see how they operate. Do you think the weights inside of one are somehow hidden to us? How are they updated during training cycles? There are tools to transform any feedforward network into a rules-based system. It is difficult to apply symbols to rules generated so that we can *easily* understand how it makes decisions, but the rules never lie...
And "regarding"... : Can it reason? Can it understand? Can it learn? A database does not do these things. The neural network structure can, but they've provided no proper explanation as to how it will operate.
So currently they have a nice project to accumulate common sense information. Good for them. But there's no solid plan to do anything with that data. I wish them luck in developing that plan, but I think there are problems that are going to be extremely difficult to overcome.
MindPixel is an interesting project. The collection of "common-sense" information has been done several times before (CYC etc.) and this does seem like an undertaking that could gain quite a bit of speed with all the media attention it's gotten.
I take issue with what it intends to (ever so vaguely) do. From the Website:
Eventually, it is hoped a GAC trained neural network will become indistinguishable from any human being when presented with any yes/no question/statement independent of whether or not GAC has seen that particular question/statement before
A Neural network is a Turing machine (a very large, hard to draw Turing machine ). This neural network will not solve the halting problem. Not too big deal, since I assume he meant "any reasonable statement", and exclude any problems that can be transformed into the halting problem. Still, it is an interesting point to bring up (I think).
Another issue: this neural network, can it reason about it's reasoning? Not terribly interesting if you can't get it to do that. Oh, it's still usefull if it can answery yes/no questions. You can always rip the network apart to figure out how it came to that conclusion. It's just very painful. And you end up with all sorts of numeric rules that are hard to give symbolic names to.
These are all just small sticking points. It would be interesting to see them addressed. I do have one large sticking point:
it's a database. There is no intelligence. There is not intelligence in the facts. It's how the facts are used. And that's my problem with the entire project. It's quite clear that the intend to use a neural network trained on these facts. What kind of network? What sort of training? What sort of validation? Justify the use of a neural network over any other form of AI, suggest a new hybrid form. How will facts be encoded? Will a binary form of a grammar tree be presented? Input sizes to the tree will vary, how will missing data be handled? Will it only give yes/no values as output? Will it be a floating point number that we can assume to be a confidence value? These are very important questions that remain unanswered.
There seems to be no available information on how the facts will be used. Seems like a bit of a scam to create a database for resale to someone who might actually be doing some research.
Slashdot Mirror
2. The halting problem is one people can solve. You can supply the problem as a GAC. The mindpixel approach can not solve it -- They claim that it can handle problems that are similar, but different from existing ones in its database, but neural networks would not be able to solve them. Rendering one of their claims in need of clarification.
4. You can indeed pull apart neural networks to see how they operate. Do you think the weights inside of one are somehow hidden to us? How are they updated during training cycles? There are tools to transform any feedforward network into a rules-based system. It is difficult to apply symbols to rules generated so that we can *easily* understand how it makes decisions, but the rules never lie...
And "regarding"... : Can it reason? Can it understand? Can it learn? A database does not do these things. The neural network structure can, but they've provided no proper explanation as to how it will operate.
So currently they have a nice project to accumulate common sense information. Good for them. But there's no solid plan to do anything with that data. I wish them luck in developing that plan, but I think there are problems that are going to be extremely difficult to overcome.
I take issue with what it intends to (ever so vaguely) do. From the Website:
Eventually, it is hoped a GAC trained neural network will become indistinguishable from any human being when presented with any yes/no question/statement independent of whether or not GAC has seen that particular question/statement before
A Neural network is a Turing machine (a very large, hard to draw Turing machine ). This neural network will not solve the halting problem. Not too big deal, since I assume he meant "any reasonable statement", and exclude any problems that can be transformed into the halting problem. Still, it is an interesting point to bring up (I think).
Another issue: this neural network, can it reason about it's reasoning? Not terribly interesting if you can't get it to do that. Oh, it's still usefull if it can answery yes/no questions. You can always rip the network apart to figure out how it came to that conclusion. It's just very painful. And you end up with all sorts of numeric rules that are hard to give symbolic names to.
These are all just small sticking points. It would be interesting to see them addressed. I do have one large sticking point: it's a database. There is no intelligence. There is not intelligence in the facts. It's how the facts are used. And that's my problem with the entire project. It's quite clear that the intend to use a neural network trained on these facts. What kind of network? What sort of training? What sort of validation? Justify the use of a neural network over any other form of AI, suggest a new hybrid form. How will facts be encoded? Will a binary form of a grammar tree be presented? Input sizes to the tree will vary, how will missing data be handled? Will it only give yes/no values as output? Will it be a floating point number that we can assume to be a confidence value? These are very important questions that remain unanswered.
There seems to be no available information on how the facts will be used. Seems like a bit of a scam to create a database for resale to someone who might actually be doing some research.