Hacker News new | ask | show | jobs
by medius 2672 days ago
Just a few thoughts I've been mulling for a while about this topic:

Machine learning is something that I believe can take advantage of analog computing. A machine learning algorithm does not need highly precise or accurate representations. Most current implementations of such processing units use fewer bits (usually 8).

However, even if we use fewer bits, the engineering effort (design, layout, lithography, etc.) that goes into making the processing unit still assumes that those few bits are error free. The manufacturing process treats it like any other digital circuit. It assumes data processing part should be fault free (e.g. treat MSB and LSB the same). Digital circuits also demand higher power compared to analog versions.

If an analog circuit can be designed for such algorithms, not only could it be much faster, it will probably consume far less power. With a super high bandwidth consuming little power, an analog processing chip may give us a much better playground to try advanced algorithms. The materials can then be optimized and we might end up with something like a brain.

Brains (all animals) process far more information for the power they consume.

Digital circuits give us low level reliability and so they are really good for simple control. Analog/biology don't give us that. But they can give us a high level reliability while delegating the low level reliability to digital counterparts.

2 comments

I think you’re wrong about the ML precision. You need highly precise for most recursive machine learning tasks because you’re compounding errors otherwise.

Typically you can’t even use floating point representation: not accurate enough.

Disagree. https://arxiv.org/abs/1805.08691 demonstrates 8-bit architecture for a pre-trained CNN provides more than acceptable results with lower latency and higher throughput than a higher precision version.
Sure but we were talking about precision rather than throughput ; tbh the result is hardly surprising .
Analog can actually be much more precise than 8-bit since there's no quantization noise (yes, this is not A2D, but any intermediate results can take middle values that will only make the final value more precise than digital signal).
The D-Wave quantum annealers are analog computers, and they're being used for machine learning research as we type.