Ask HN: How can AI companies have proprietary technologies?

#1

Post by chienomi » Mon, Feb 27, 2017, 4:14 PM UTC

Machine learning technologies have doubtlessly been improved and AI companies are emerging these days.

Almost all of these technologies are open to the public, including the source code, in some cases. How can a company own a unique AI technology then?

Re: Ask HN: How can AI companies have proprietary technologies?

#2

Post by dwringer » Mon, Feb 27, 2017, 4:42 PM UTC

I'm not sure I'm able to provide a solid answer, but I would venture to say that to the extent a neuromorphic or neural "deep learning" network relies on an expensive training process, its inherent value becomes more-or-less unique to each specific implementation. Having the source code to build one's own implementation would not replace what could be hundreds of thousands of processor hours spent analyzing petabytes of data. In that case it would depend on whether the trained models were released under the open-source model, or only the underlying architecture.

Re: Ask HN: How can AI companies have proprietary technologies?

#3

Post by chienomi » Mon, Feb 27, 2017, 5:30 PM UTC

dwringer wrote: post #2

I'm not sure I'm able to provide a solid answer, but I would venture to say that to the extent a neuromorphic or neural "deep learning" network relies on an expensive training process, its inherent value becomes more-or-less unique to each specific implementation. Having the source code to build one's own implementation would not replace what could be hundreds of thousands of processor hours spent analyzing petabytes…

Thanks. This can explain the situation where most people are struggling to replicate a neat result in a paper.

Re: Ask HN: How can AI companies have proprietary technologies?

#4

Post by dwringer » Mon, Feb 27, 2017, 7:48 PM UTC

chienomi wrote: post #3

dwringer wrote: post #2

I'm not sure I'm able to provide a solid answer, but I would venture to say that to the extent a neuromorphic or neural "deep learning" network relies on an expensive training process, its inherent value becomes more-or-less unique to each specific implementation. Having the source code to build one's own implementation would not replace what could be hundreds of thousands of processor hours spent analyzing petabytes…

Thanks. This can explain the situation where most people are struggling to replicate a neat result in a paper.

Well, I'm afraid of that I am not so sure. A well-designed paper should demonstrate what the training process was and use a publicly available dataset for training. There are many repositories of training data that exist specifically to provide a standard like this in various domains, so anyone following a paper should be able to use the same training process on the same data and get the same result. A paper presenting a result that cannot be obtained in this way can be informative but does not substitute for peer-reviewed research.

Re: Ask HN: How can AI companies have proprietary technologies?

#5

Post by dwringer » Tue, Feb 28, 2017, 4:54 PM UTC

chienomi wrote: post #3

dwringer wrote: post #2

I'm not sure I'm able to provide a solid answer, but I would venture to say that to the extent a neuromorphic or neural "deep learning" network relies on an expensive training process, its inherent value becomes more-or-less unique to each specific implementation. Having the source code to build one's own implementation would not replace what could be hundreds of thousands of processor hours spent analyzing petabytes…

Thanks. This can explain the situation where most people are struggling to replicate a neat result in a paper.

Here is a discussion with some information on such data repositories as I mention in my other reply: http://academia.stackexchange.com/questions/85722/what-is-th...