Hardware neural nets and the AI

This is a little question about the AI that makes some of the work in eyewire.

Don’t you use hardware neural nets? Or plan to? They should be considerably more powerful then a normal computer for this task.

It doesn’t seam that these things exist on the market. Or maybe they do? If you can buy them, then why don’t you use them? Or when you think you’ll be able to use/buy them?

*Hum… this might confuse some. I do mean “real” artificial neural nets on a silicon chip. Not a simulated software neural net on a normal computer. …or the biological neural nets that are studied.

Hi quantum_immortal,

While cool in concept, hardware ASICs are expensive to develop and rapidly outmoded as technology improves, it locks you into a certain capability level. Our lab has produced software that makes use of latest generation CPU boards ( e.g. 60 cores) efficiently so we can keep changing our architectures as new ideas are discovered or features of the software are engineered (not features of the data, the neural net implicitly discovers those).

We’ve open sourced it, you can check it out here! https://github.com/seung-lab/znn-release/

Will

1 Like

I should mention FPGAs, but those require using VHDL or something similar to program which is not the most fun thing to work with. Maybe that’s a place to find some competitive advantage. :wink:

1 Like

so… you are saying, not used because too expensive.

And too expensive because not mass produced?

You think that they might be mass produced in the near future? Maybe as a gimmick for smartphones or commercial robots or self driving cars or whatever?