SoylentNews is people
SoylentNews
IBM researchers use analog memory to train deep neural networks faster and more efficiently
Deep neural networks normally require fast, powerful graphical processing unit (GPU) hardware accelerators to support the needed high speed and computational accuracy — such as the GPU devices used in the just-announced Summit supercomputer. But GPUs are highly energy-intensive, making their use expensive and limiting their future growth, the researchers explain in a recent paper published in Nature.
Instead, the IBM researchers used large arrays of non-volatile analog memory devices (which use continuously variable signals rather than binary 0s and 1s) to perform computations. Those arrays allowed the researchers to create, in hardware, the same scale and precision of AI calculations that are achieved by more energy-intensive systems in software, but running hundreds of times faster and at hundreds of times lower power — without sacrificing the ability to create deep learning systems.
The trick was to replace conventional von Neumann architecture, which is "constrained by the time and energy spent moving data back and forth between the memory and the processor (the 'von Neumann bottleneck')," the researchers explain in the paper. "By contrast, in a non-von Neumann scheme, computing is done at the location of the data [in memory], with the strengths of the synaptic connections (the 'weights') stored and adjusted directly in memory.
Equivalent-accuracy accelerated neural-network training using analogue memory (DOI: 10.1038/s41586-018-0180-5) (DX)
(Score: 3, Informative) by Uncle_Al on Thursday June 21 2018, @01:23AM (14 children)
woot!
(Score: 2) by takyon on Thursday June 21 2018, @01:24AM (11 children)
The Bot uprising is going to be so steampunk.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by RS3 on Thursday June 21 2018, @01:56AM (10 children)
Everyone has been led to believe that digital is the end-all. We will be owned by analog bots.
(Score: 2) by bob_super on Thursday June 21 2018, @02:12AM
We're currently owned by organic bots, after all.
(Score: 0) by Anonymous Coward on Thursday June 21 2018, @02:53AM (8 children)
Unless you have a way of digitizing molecular structures then duh.
(Score: 2) by c0lo on Thursday June 21 2018, @04:03AM (6 children)
Given the discrete QM nature of the molecular structures, we are already there.
A pity the same QM nature doesn't (yet) allow us to control the so many discrete states those structures can exhibit.
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by RS3 on Thursday June 21 2018, @04:07AM (1 child)
D'oh! You beat me by 2 minutes.
(Score: 2) by c0lo on Thursday June 21 2018, @04:30AM
Sorry, I just didn't realize there was a competition going.
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by RS3 on Thursday June 21 2018, @04:22AM (3 children)
I'm curious what you mean. I believe many things function by stimulating energy states, such as pretty much anything that gives off photons / light including LASERs, magnetrons, klystrons, TWAT, vacuum tubes (electronic "valves"), X-ray, etc. So I'm guessing you mean some more advanced thing?
(Score: 2) by c0lo on Thursday June 21 2018, @04:36AM (2 children)
In all examples, you take advantages of a large set of molecular structures so that you apply sorta "macroscopic" filter over the output and discard the rest of the output. This doesn't mean control, it just mean "selection".
We aren't in the technological position to control the states of a molecular structures at individual levels - and Heisenberg warned us again thinking that we'll ever be able.
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by RS3 on Thursday June 21 2018, @06:14AM (1 child)
Okay, thanks. But aren't narrow spectrum light sources doing this?
But generally your answer is what I envisioned you were alluding to, at the most advanced concept. So one electric field exciting wire per atom. Or maybe a very focused electron beam. Hey, don't scanning tunneling electron microscopes approach that?
So what would be the purpose, output, whatever, of doing it? New molecules and compounds? Superconductivity? Something relating to nuclear fusion? Gene editing? Medicines? Tricorders? All of the above and more?
(Score: 2) by c0lo on Thursday June 21 2018, @06:45AM
Measuring != controlling
CPU at Ångström-unit scale? :)
https://www.youtube.com/watch?v=aoFiw2jMy-0 https://soylentnews.org/~MichaelDavidCrawford
(Score: 2) by RS3 on Thursday June 21 2018, @04:05AM
Pretty sure molecular structures are held together by bonds that obey laws described by quantum mechanics. But I'm not a chemist nor physicist. Well, a little.
(Score: 1, Informative) by Anonymous Coward on Thursday June 21 2018, @02:37AM (1 child)
Weren't Perceptrons analog?
[googles]
Here's one paper from 2013 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3574673/ [nih.gov]
> This study examines an analog circuit comprising a multilayer perceptron neural network (MLPNN). This study proposes a low-power and small-area analog MLP circuit to implement in an E-nose as a classifier, such that the E-nose would be relatively small, power-efficient, and portable. The analog MLP circuit had only four input neurons, four hidden neurons, and one output neuron. The circuit was designed and fabricated using a 0.18 μm standard CMOS process with a 1.8 V supply. The power consumption was 0.553 mW, and the area was approximately 1.36 × 1.36 mm2. The chip measurements showed that this MLPNN successfully identified the fruit odors of bananas, lemons, and lychees with 91.7% accuracy.
Marvin Minsky debunked Perceptrons with his famous book of the same name, but now it's looking more like the only problem with the early attempts was that they weren't big enough and/or deep enough.
(Score: 2) by fritsd on Thursday June 21 2018, @04:56PM
Minsky and Papert spotted that, if you use a linear activation function, then whatever depth of Perceptron can be re-written as a simple multivariate linear equation.
It's a pity that many people then gave up on Perceptrons until Rumelhart and McClelland (IIRC) rekindled the interest a long time after.