Circuits that work as neural networks

The goal was circuits that behave like neurons and synapses: they compute weighted sums, apply activation functions and pass signals on, so the network runs in the hardware itself. They can be trained either on the hardware or in simulation.

I split the design into small modules, wrote tests for each one's behavior, and refined them in SPICE simulation (driven from Python with PySpice and NgSpice) before anything was built. For training, I adapted established neural-network training methods to work inside the circuit, and used TensorFlow and PyTorch for the software side.