Zeroing in on demonstrating an efficient fixed-point implementation of a neural network that fits in 64 kB data memory. Include ability to speed up and slow down CPU when processing is needed.
Can we scale the processing to accommodate less memory (cuts down on leakage) and low peak power (enables continuous, but slow, processing when connected to energy scavenging sources)?
What NN task should serve as a demonstrator?
@Titan Yuan may have a fixed point implementation somewhere
tinyML is a good project to look at too: https://tinyml.mit.edu/