I worked for a while as the software/FPGA guy on a high-power RF amplifier for MRI machines. I worked with some very talented RF engineers. There were obviously some parts of the chain that were analog (like the preamps and high power amp itself). But lots of it was being replaced with digital (DSP techniques) in the next generation and the RF engineers loved it. Filters no longer needed a 15 minute tuning session by a technician with a network analyzer. A Hilbert transform is far easier and better in digital than analog. Design changes could be experimented with in a day instead of PCB spins that took weeks.
Given all these constraints I am actually quite excited to hear more of the dissenting view. Can you describe some of the research you mentioned in the last sentence? I have not heard of it and if true it would be at least interesting engineering, even if right now I doubt it would actually work.
Most machine learning algorithms are by design resilient to random noise. They can even be learned to be resilient to systematic noise (due to say.. variable hardware performance in mass production).
In those cases; a low power analog device with lower reliability guarantees is 'ok'.
I suspect the "decision engines" (the computation enforcing complex logic based on sensor readings) will probably continue to be on CPU's for some time.
It would surprise me if even unfiltered analog noise in the sensors significantly changed recall. I suppose I don't know off-hand know the magnitude of noise in typical analog systems.
As a proof of concept they did a GPS chip , with 66 times higher energy efficiency.
There are others working on analog neural compute , some released good research results .
Also, there's a huge application domain in image preprocessing: the pixel value is analog and noisy anyway, so building an analog primary visual cortex with a bank of common filters (line segments, stable features etc.) makes a lot of sense.
ADD: I'm also thinking about accelerating RANSAC-type algorithms, where you perform lots of preliminary evaluations to filter out the outliers. I bet one can cheaply implement an "100-argmins of 100 x 100 fixed10 distance matrix" computation in a handful of cycles with analog circuits. And high-dimensional RANSAC-type algorithms are what needed for robust and precise 3D visual navigation.
And that same future applies not only for neural computing, but for a field called approximate-computing, i.e. computing where results aren't accurate.Some/many signal and image processing work well with that. I've also seen some research about doing scientific computing on approximate hardware and correcting errors.