Thanks for originally sharing; hoping to do something with this info and will share.

If you want some funding or impact, an accelerator for a proven deep-learning architecture could be useful. Prior work in neural computation got plenty results. Just havent seen many independents try it for deep learning. One wacky idea I had was using analog accelerators for complex, evaluation functions in genetic algorithms. The digital chip would just feed each solution to a bank of them with fitness result popping out other side. Another idea was attempting any of that with Field-Programmable, Analog Arrays to see what results they got. Lowers development costs vs making masks for custom circuits.

Those are awesome ideas, ones I never would have thought of! I was thinking more along the lines of using analog filtering for convolutional layers of deep neural networks, dropout layers and filter templates for 2-d features are just bandpass filters, which analog filters can approximate easily. That analog fpga idea is a good one also; there needs to be a way to perform rapid experimentation on these types of ideas. Are there any products/companies you know of in this area?

Your ideas are possibly better just because you already have a chance of building them if they're that close to existing analog schemes. I'd try them first to get a feel for the whole thing.

Far as reprogrammable, type in what I told you into Google: Field-Programmable, Analog Arrays (FPAA's). Or reconfigurable analog. You'll get companies and CompSci. Also remember that analog works best/easiest on oldest process nodes. That combined with multi-project wafers can get your ASIC prototypes down to thousands of dollars rather than tens to hundreds. Middle ground company is a structured ASIC that's basically like a FPGA programmed with 1-2 masks at fab time. Triad Sdmiconductor does that for digital+analog with claimed $400k per design with few weeks turn around. Idk if their analog cells are suitable, though.

Also, for good measure you might enjoy at least looking at the last general-purpose (programmable), analog computer:

http://www.comdyna.com/gp6intro.htm

Here is a neural network using analog weights and neurons: https://arxiv.org/abs/1610.02091