Interesting to see my ideas for using ternary data (zero crossings) being implemented. It turns matrix multiplications into simple additions and subtractions! I have been working with this for nearly 40 years when my investigations into doing signal processing on severely truncated data led me to examine what happens to the FFT when the orthogonal basis sets are taken to the extremes of ternary values - zero crossings. There is still an astonishing amount of information to be had from ternary weights and data! You can recover the Walsh-Hadamard Transform from the Fourier Transform this way!
What is not so great is the (by now expected) behaviour of Microsoft in not acknowledging my inputs and communications …