Do image classification like bees do it. Find a (simple) electric curcuit that can classify a event in an analog input video signal. Train on simulated harware that gives us the gradients.
Gains
- dirt cheap to produce harware for
- 1000x electricity saving for classification
- unhackable
(could be used for classification in privacy-sensitive environments. eg. shower)
- Use NetworksDynamics.jl / DiffEq solver to simulate electric circuit
- figure out building blocks that might work (resistors, capacitors, transitors, etc etc)
- figure out encoding for input data, figure out loss function
- come up with way of generating training/test data
- generate circuit typologies
- Run training to tweak parameters (eg. resistivity, capacitor size etc.)
- Order resulting circuit as implemented boards
- Run trail whether it does the job
- Celebrate 🎉
- maybe one could come up with a deep learning thingy that predicst good candidates for curcuits (the gradient could be taken through the deep learning model AND the simulation) similar to the auto-updating heuristics in RoboGrammar.