A computer that’s only as good as its software?
That’s the idea behind a new paper published in Science, but is it true?
The article, by researchers from the University of Texas, describes a system called the Neural Network Architecture, or NPAPI.
The system was created by a team at MIT and DARPA in the 1980s, but its name is a play on the word “neural net”.NBNA is a framework for designing computer systems that can represent any mathematical object in real time, allowing for parallel processing.
That’s where the word neural net comes from, and the NPAPI is its name.
The paper was co-authored by two of its co-authors, Stephen Yudkin and David A. Strominger, who were both graduate students at the time.
The authors of the paper say the NPASA system is designed to represent any computer system, and to do so, it has to be scalable, easy to implement, and provide a scalable and robust computing architecture.
They call the NPAPIS “a general purpose general-purpose network” that can be used for many different purposes.
“It’s not just a neural network; it’s a network for the world of physics,” Yudkins told the BBC.
But how scalable is it?
And can we build something like this?
According to the paper, the NPAs are capable of representing mathematical objects in “super-computing” time, that is, at the speed of light.
That means they can represent numbers that are larger than one billionth of a meter (one trillionth of an inch) in some instances.
And they can also represent numbers larger than a millionth of the Earth’s mass, but that doesn’t mean they can do anything in the real world.
“We think the NP API is a general-use general-Purpose general- purpose network,” Stromingers told the New York Times.
That’s a bit of a stretch, and there’s a big problem with the paper.
The researchers did not give an estimate of how much the NP api can represent.
“It’s a really big thing that we have not even gotten to the level of describing,” Stomings said.
And the authors of this paper have a vested interest in not revealing the details of the implementation.
“We’re not trying to prove anything here, but we’re also not trying not to have this as a topic for future research,” Yudskin told the Times.
The authors are cautious about their conclusions.
“In order to be sure that our results are generalizable to any system, we need to verify that they are indeed generalizable,” Yudiks said.
But it’s worth noting that it is possible to simulate a neural net that can do any number of things, and in that way the paper does not contradict its authors.