Actually I went and read a bit about the history here, and although I left a glib comment, there was major practical research into neural network approaches, with theory in 1943 and hardware by 1953 https://en.m.wikipedia.org/wiki/Perceptron
There were two major factors though that set back the neural net approach:
- a 1969 Minsky & Papert book on perceptrons lead to a belief that neural nets even of >1 layers had fundamental limits, although the book only showed such limits in 1-layer nets; this lead to a reduction in funding during various AI winters
- the “deep” in “deep learning” is all about how much larger & deeper neural systems produce substantially better results. Even if you can speculate theoretically about this, it was completely impractical to approach the scale/speed to see fruitful results until the late 90s when vector/matrix accelerators (SIMD, GPU-type things) start showing up en masse. I vaguely remember reading about advances in ML in the mid 2000s which sort of had an attitude of “huh, this neural net thing we thought was a dead end turns out to just need MOAR CORES (graph up and to the right)”
Even in the 90s through the late 2000s, when I started working in ML, people poo-poo'd it: not enough data, not good enough algorithms, and computers too slow. And I worked with supercomputers/HPC- you'd think they would have been the first groups to exploit machine learning.
The perceptron was actually a remarkable cool device, way ahead of its time.
Shouldn't need a lick of electricity! If by valves you mean fluidics [1]... at which point, harnessing the Niagra Falls and building out a fluidized supercomputer covering the great lakes would probably suffice. No worries about the waste heat, though, it's water cooled!
And how much electricity would it consume? Although the heat it generated would probably boil the Great Lakes!